Personalized Daily Arxiv Papers 07/15/2025

arXiv:2507.10302v1 Announce Type: new Abstract: In video Multimodal Large Language Models (video MLLMs), the visual encapsulation process plays a pivotal role in converting video contents into representative tokens for LLM input. While linear projectors are widely employed for encapsulation, they introduce semantic indistinctness and temporal incoherence when applied to videos. Conversely, the structure of resamplers shows promise in tackling these challenges, but an effective solution remains unexplored. Drawing inspiration from resampler structures, we introduce DisCo, a novel visual encapsulation method designed to yield semantically distinct and temporally coherent visual tokens for video MLLMs. DisCo integrates two key components: (1) A Visual Concept Discriminator (VCD) module, assigning unique semantics for visual tokens by associating them in pair with discriminative concepts in the video. (2) A Temporal Focus Calibrator (TFC) module, ensuring consistent temporal focus of visual tokens to video elements across every video frame. Through extensive experiments on multiple video MLLM frameworks, we demonstrate that DisCo remarkably outperforms previous state-of-the-art methods across a variety of video understanding benchmarks, while also achieving higher token efficiency thanks to the reduction of semantic indistinctness. The code: https://github.com/ZJHTerry18/DisCo.

arXiv:2507.10202v1 Announce Type: new Abstract: Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities in vision-language understanding, reasoning, and generation. However, they struggle with tasks requiring fine-grained localization and reasoning in high-resolution images. This constraint stems from the fact that MLLMs are fine-tuned with fixed image resolution to align with the pre-trained image encoder used in MLLM. Consequently, feeding high-resolution images directly into MLLMs leads to poor generalization due to a train-test resolution discrepancy, while downsampling these images-although ensuring consistency-compromises fine-grained visual details and ultimately degrades performance. To address this challenge, we propose Extract Candidate then Predict (ECP), a novel training-free, task-agnostic two-stage framework designed to enhance MLLM performance on high-resolution images. The key intuition behind ECP is that while MLLMs struggle with high-resolution images, their predictions on downsampled images still contain implicit localization cues. By first identifying candidate region using the coarse prediction and then predicting the final output based on candidate region, ECP effectively preserves fine-grained details while mitigating the challenges posed by high-resolution data. We validate our framework on 4K GUI grounding and 4K, 8K MLLM perception, achieving +21.3%, +5.8%, +5.2% absolute improvement compared to baseline respectively, demonstrating its effectiveness. Code is available at https://github.com/yenncye/ECP.

arXiv:2507.09374v1 Announce Type: new Abstract: Multimodal large language models (MLLMs) still perform poorly on scientific tasks, particularly those requiring multi-step and interpretable reasoning. Their limitations include insufficient scientific reasoning patterns, lack of global coherence in multi-step inference, and the absence of reflective self-correction, making them unreliable in structured scientific contexts. We introduce EduFlow, the first end-to-end framework that covers the full pipeline of educational scientific reasoning, including data selection, MCTS-based trajectory construction, model training, and output optimization. At its core is EduPRM, a process-aware reward model that critiques reasoning steps with tags and justifications. EduPRM is trained via curriculum learning on three complementary supervision sources: MCTS-guided trajectories, error-injected critiques, and teacher-student dialogues, enabling dynamic adaptation to multi-stage problem solving and iterative refinement during inference. We further propose EduMCTS, a domain-adapted search framework that introduces bootstrapping actions specifically designed for educational reasoning, such as a self-reflection mechanism that promotes reflective error correction. It further leverages EduPRM's fine-grained feedback to guide the search toward higher-quality reasoning trajectories. By applying self-consistency and rejection sampling, we constructed EduMCTS-160K, a large-scale dataset of educational reasoning trajectories. Extensive experiments demonstrate that EduFlow enhances reasoning consistency and coherence. Code, data, and models will be released.

arXiv:2507.09404v1 Announce Type: new Abstract: Large foundation models are typically trained on data from multiple domains, with the data mixture--the proportion of each domain used--playing a critical role in model performance. The standard approach to selecting this mixture relies on trial and error, which becomes impractical for large-scale pretraining. We propose a systematic method to determine the optimal data mixture for any target domain using scaling laws. Our approach accurately predicts the loss of a model of size $N$ trained with $D$ tokens and a specific domain weight vector $h$. We validate the universality of these scaling laws by demonstrating their predictive power in three distinct and large-scale settings: large language model (LLM), native multimodal model (NMM), and large vision models (LVM) pretraining. We further show that these scaling laws can extrapolate to new data mixtures and across scales: their parameters can be accurately estimated using a few small-scale training runs, and used to estimate the performance at larger scales and unseen domain weights. The scaling laws allow to derive the optimal domain weights for any target domain under a given training budget ($N$,$D$), providing a principled alternative to costly trial-and-error methods.

arXiv:2507.09693v1 Announce Type: new Abstract: Experiment commentary is crucial in describing the experimental procedures, delving into underlying scientific principles, and incorporating content-related safety guidelines. In practice, human teachers rely heavily on subject-specific expertise and invest significant time preparing such commentary. To address this challenge, we introduce the task of automatic commentary generation across multi-discipline scientific experiments. While recent progress in large multimodal models (LMMs) has demonstrated promising capabilities in video understanding and reasoning, their ability to generate fine-grained and insightful experiment commentary remains largely underexplored. In this paper, we make the following contributions: (i) We construct \textit{ExpInstruct}, the first dataset tailored for experiment commentary generation, featuring over 7\textit{K} step-level commentaries across 21 scientific subjects from 3 core disciplines (\ie, science, healthcare and engineering). Each sample includes procedural descriptions along with potential scientific principles (\eg, chemical equations and physical laws) and safety guidelines. (ii) We propose ExpStar, an automatic experiment commentary generation model that leverages a retrieval-augmented mechanism to adaptively access, evaluate, and utilize external knowledge. (iii) Extensive experiments show that our ExpStar substantially outperforms 14 leading LMMs, which highlights the superiority of our dataset and model. We believe that ExpStar holds great potential for advancing AI-assisted scientific experiment instruction.

arXiv:2507.10007v1 Announce Type: new Abstract: Chain of Thought (CoT) reasoning has demonstrated remarkable deep reasoning capabilities in both large language models (LLMs) and multimodal large language models (MLLMs). However, its reliability is often undermined by the accumulation of errors in intermediate steps. This paper introduces an novel approach to calibrate the CoT reasoning accuracy by leveraging the model's intrinsic veracity encoding. We discover that specific attention head activations reliably reflect the truthfulness of reasoning steps in CoT. Based on this insight, we train a confidence predictor to evaluate the correctness of each reasoning step using these truthfulness-sensitive activations, dynamically selecting the most plausible reasoning path via beam search. Experimental results demonstrate that our method significantly outperforms the state-of-the-art baselines (e.g., Few-Shot CoT, Self-Consistency, and Self-Evaluation Guided Beam Search) across the mathematical, symbolic, and commonsense reasoning tasks, exhibiting superior accuracy and reliability in both unimodal and multimodal settings. We further validate the approach on large reasoning models, confirming its applicability to specialized reasoning models. Additionally, we explore the role of the model's self-correction ability in CoT reasoning. This work provides a novel reliability improvement path for CoT reasoning with broad application potential.

arXiv:2507.09184v1 Announce Type: new Abstract: Hallucinations pose a significant challenge in Large Vision Language Models (LVLMs), with misalignment between multimodal features identified as a key contributing factor. This paper reveals the negative impact of the long-term decay in Rotary Position Encoding (RoPE), used for positional modeling in LVLMs, on multimodal alignment. Concretely, under long-term decay, instruction tokens exhibit uneven perception of image tokens located at different positions within the two-dimensional space: prioritizing image tokens from the bottom-right region since in the one-dimensional sequence, these tokens are positionally closer to the instruction tokens. This biased perception leads to insufficient image-instruction interaction and suboptimal multimodal alignment. We refer to this phenomenon as image alignment bias. To enhance instruction's perception of image tokens at different spatial locations, we propose MCA-LLaVA, based on Manhattan distance, which extends the long-term decay to a two-dimensional, multi-directional spatial decay. MCA-LLaVA integrates the one-dimensional sequence order and two-dimensional spatial position of image tokens for positional modeling, mitigating hallucinations by alleviating image alignment bias. Experimental results of MCA-LLaVA across various hallucination and general benchmarks demonstrate its effectiveness and generality. The code can be accessed in https://github.com/ErikZ719/MCA-LLaVA.

arXiv:2507.08979v1 Announce Type: new Abstract: We introduce Projection-based Reduction of Implicit Spurious bias in vision-language Models (PRISM), a new data-free and task-agnostic solution for bias mitigation in VLMs like CLIP. VLMs often inherit and amplify biases in their training data, leading to skewed predictions. PRISM is designed to debias VLMs without relying on predefined bias categories or additional external data. It operates in two stages: first, an LLM is prompted with simple class prompts to generate scene descriptions that contain spurious correlations. Next, PRISM uses our novel contrastive-style debiasing loss to learn a projection that maps the embeddings onto a latent space that minimizes spurious correlations while preserving the alignment between image and text embeddings.Extensive experiments demonstrate that PRISM outperforms current debiasing methods on the commonly used Waterbirds and CelebA datasets We make our code public at: https://github.com/MahdiyarMM/PRISM.

arXiv:2507.09985v1 Announce Type: new Abstract: Touch is recognized as a vital sense for humans and an equally important modality for robots, especially for dexterous manipulation, material identification, and scenarios involving visual occlusion. Building upon very recent work in touch foundation models, this demonstration will feature Octopi-1.5, our latest visual-tactile-language model. Compared to its predecessor, Octopi-1.5 introduces the ability to process tactile signals from multiple object parts and employs a simple retrieval-augmented generation (RAG) module to improve performance on tasks and potentially learn new objects on-the-fly. The system can be experienced live through a new handheld tactile-enabled interface, the TMI, equipped with GelSight and TAC-02 tactile sensors. This convenient and accessible setup allows users to interact with Octopi-1.5 without requiring a robot. During the demonstration, we will showcase Octopi-1.5 solving tactile inference tasks by leveraging tactile inputs and commonsense knowledge. For example, in a Guessing Game, Octopi-1.5 will identify objects being grasped and respond to follow-up queries about how to handle it (e.g., recommending careful handling for soft fruits). We also plan to demonstrate Octopi-1.5's RAG capabilities by teaching it new items. With live interactions, this demonstration aims to highlight both the progress and limitations of VTLMs such as Octopi-1.5 and to foster further interest in this exciting field. Code for Octopi-1.5 and design files for the TMI gripper are available at https://github.com/clear-nus/octopi-1.5.

arXiv:2507.09279v1 Announce Type: new Abstract: Multimodal large language models (MLLMs) hold considerable promise for applications in healthcare. However, their deployment in safety-critical settings is hindered by two key limitations: (i) sensitivity to prompt design, and (ii) a tendency to generate incorrect responses with high confidence. As clinicians may rely on a model's stated confidence to gauge the reliability of its predictions, it is especially important that when a model expresses high confidence, it is also highly accurate. We introduce Prompt4Trust, the first reinforcement learning (RL) framework for prompt augmentation targeting confidence calibration in MLLMs. A lightweight LLM is trained to produce context-aware auxiliary prompts that guide a downstream task MLLM to generate responses in which the expressed confidence more accurately reflects predictive accuracy. Unlike conventional calibration techniques, Prompt4Trust specifically prioritizes aspects of calibration most critical for safe and trustworthy clinical decision-making. Beyond improvements driven by this clinically motivated calibration objective, our proposed method also improves task accuracy, achieving state-of-the-art medical visual question answering (VQA) performance on the PMC-VQA benchmark, which is composed of multiple-choice questions spanning diverse medical imaging modalities. Moreover, our framework trained with a small downstream task MLLM showed promising zero-shot generalization to larger MLLMs in our experiments, suggesting the potential for scalable calibration without the associated computational costs. This work demonstrates the potential of automated yet human-aligned prompt engineering for improving the the trustworthiness of MLLMs in safety critical settings. Our codebase can be found at https://github.com/xingbpshen/vccrl-llm.

arXiv:2507.09334v1 Announce Type: new Abstract: While 3D Multi-modal Large Language Models (MLLMs) demonstrate remarkable scene understanding capabilities, their practical deployment faces critical challenges due to computational inefficiency. The key bottleneck stems from processing excessive object-centric visual tokens required for comprehensive 3D scene representation. Although visual token pruning has shown promise in accelerating 2D MLLMs, its applicability to 3D domains remains largely unexplored due to fundamental disparities in token structures. In this paper, we reveal two critical insights: (1) Significant redundancy exists in object-level 3D token representations, analogous to patch-level redundancy in 2D systems; (2) Global attention patterns exhibit strong predictive power for identifying non-essential tokens in 3D contexts. Building on these observations, we propose Fast3D, a plug-and-play visual token pruning framework for 3D MLLMs featuring two technical innovations: (1) Global Attention Prediction (GAP), where a lightweight neural network learns to predict the global attention distributions of the target model, enabling efficient token importance estimation for precise pruning guidance; (2) Sample-Adaptive visual token Pruning (SAP), which introduces dynamic token budgets through attention-based complexity assessment, automatically adjusting layer-wise pruning ratios based on input characteristics. Both of these two techniques operate without modifying the parameters of the target model. Extensive evaluations across five benchmarks validate the effectiveness of Fast3D, particularly under high visual token pruning ratios. Code is available at https://github.com/wencan25/Fast3D

arXiv:2507.10539v1 Announce Type: new Abstract: World models (WMs) demonstrate strong capabilities in prediction, generation, and planning tasks. Existing WMs primarily focus on unstructured data and cannot leverage the ubiquitous structured data, often represented as graphs, in the digital world. While multiple graph foundation models have been proposed, they focus on graph learning tasks and cannot extend to diverse multi-modal data and interdisciplinary tasks. To address these challenges, we propose the Graph World Model (GWM), a world model that supports both unstructured and graph-structured states with multi-modal information and represents diverse tasks as actions. The core of a GWM is a generic message-passing algorithm to aggregate structured information, either over a unified multi-modal token space by converting multi-modal data into text (GWM-T) or a unified multi-modal embedding space by modality-specific encoders (GWM-E). Notably, GWM introduces action nodes to support diverse tasks, where action nodes are linked to other nodes via direct reference or similarity computation. Extensive experiments on six tasks from diverse domains, including multi-modal generation and matching, recommendation, graph prediction, multi-agent, retrieval-augmented generation, and planning and optimization, show that the same GWM outperforms or matches domain-specific baselines' performance, benefits from multi-hop structures, and demonstrates strong zero-shot/few-shot capabilities on unseen new tasks. Our code for GWM is released at https://github.com/ulab-uiuc/GWM.

arXiv:2507.09531v1 Announce Type: new Abstract: Key Information Extraction (KIE) underpins the understanding of visual documents (e.g., receipts and contracts) by extracting precise semantic content and accurately capturing spatial structure. Yet existing multimodal large language models (MLLMs) often perform poorly on dense documents and rely on vision tokenization approaches that scale with image size, leading to redundant computation and memory inefficiency. To address these challenges, we introduce VDInstruct, an MLLM that separates spatial region detection from semantic feature extraction. Central to our model is a content-aware tokenization strategy: rather than fragmenting the entire image uniformly, it generates tokens in proportion to document complexity, preserving critical structure while eliminating wasted tokens. Leveraging a three-stage training paradigm, our model achieves state-of-the-art (SOTA) results on KIE benchmarks, matching or exceeding the accuracy of leading approaches while reducing the number of image tokens by roughly 3.6x. In zero-shot evaluations, VDInstruct surpasses strong baselines-such as DocOwl 1.5-by +5.5 F1 points, highlighting its robustness to unseen documents. These findings show that content-aware tokenization combined with explicit layout modeling offers a promising direction forward for document understanding. Data, source code, and model weights will be made publicly available.

arXiv:2507.09406v1 Announce Type: new Abstract: Large language models (LLMs) aligned for safety through techniques like reinforcement learning from human feedback (RLHF) often exhibit emergent deceptive behaviors, where outputs appear compliant but subtly mislead or omit critical information. This paper introduces adversarial activation patching, a novel mechanistic interpretability framework that leverages activation patching as an adversarial tool to induce, detect, and mitigate such deception in transformer-based models. By sourcing activations from "deceptive" prompts and patching them into safe forward passes at specific layers, we simulate vulnerabilities and quantify deception rates. Through toy neural network simulations across multiple scenarios (e.g., 1000 trials per setup), we demonstrate that adversarial patching increases deceptive outputs to 23.9% from a 0% baseline, with layer-specific variations supporting our hypotheses. We propose six hypotheses, including transferability across models, exacerbation in multimodal settings, and scaling effects. An expanded literature review synthesizes over 20 key works in interpretability, deception, and adversarial attacks. Mitigation strategies, such as activation anomaly detection and robust fine-tuning, are detailed, alongside ethical considerations and future research directions. This work advances AI safety by highlighting patching's dual-use potential and provides a roadmap for empirical studies on large-scale models.

arXiv:2507.10300v1 Announce Type: new Abstract: Multimodal large language models (MLLMs) have shown remarkable performance in vision-language tasks. However, existing MLLMs are primarily trained on generic datasets, limiting their ability to reason on domain-specific visual cues such as those in facial images. In particular, tasks that require detailed understanding of facial structure, expression, emotion, and demographic features remain underexplored by MLLMs due to the lack of large-scale annotated face image-text datasets. In this work, we introduce FaceLLM, a multimodal large language model trained specifically for facial image understanding. To construct the training data, we propose a novel weakly supervised pipeline that uses ChatGPT with attribute-aware prompts to generate high-quality question-answer pairs based on images from the FairFace dataset. The resulting corpus, called FairFaceGPT, covers a diverse set of attributes including expression, pose, skin texture, and forensic information. Our experiments demonstrate that FaceLLM improves the performance of MLLMs on various face-centric tasks and achieves state-of-the-art performance. This work highlights the potential of synthetic supervision via language models for building domain-specialized MLLMs, and sets a precedent for trustworthy, human-centric multimodal AI systems. FairFaceGPT dataset and pretrained FaceLLM models are publicly available in the project page.

arXiv:2507.09068v1 Announce Type: new Abstract: The rapid advancements in Large Language Models (LLMs) and their multimodal extensions (MLLMs) have ushered in remarkable progress in video understanding. However, a fundamental challenge persists: effectively processing and comprehending video content that extends beyond minutes or hours. While recent efforts like Video-XL-2 have demonstrated novel architectural solutions for extreme efficiency, and advancements in positional encoding such as HoPE and VideoRoPE++ aim to improve spatio-temporal understanding over extensive contexts, current state-of-the-art models still encounter significant computational and memory constraints when faced with the sheer volume of visual tokens from lengthy sequences. Furthermore, maintaining temporal coherence, tracking complex events, and preserving fine-grained details over extended periods remain formidable hurdles, despite progress in agentic reasoning systems like Deep Video Discovery. This position paper posits that a logical, albeit ambitious, next frontier for multimedia research is Infinite Video Understanding -- the capability for models to continuously process, understand, and reason about video data of arbitrary, potentially never-ending duration. We argue that framing Infinite Video Understanding as a blue-sky research objective provides a vital north star for the multimedia, and the wider AI, research communities, driving innovation in areas such as streaming architectures, persistent memory mechanisms, hierarchical and adaptive representations, event-centric reasoning, and novel evaluation paradigms. Drawing inspiration from recent work on long/ultra-long video understanding and several closely related fields, we outline the core challenges and key research directions towards achieving this transformative capability.

arXiv:2507.09071v1 Announce Type: new Abstract: Large vision-language models (VLMs) enable the joint processing of text and images. However, the inclusion of vision data significantly expands the prompt length. Along with the quadratic complexity of the attention computation, this results in a longer prefill duration. An approach to mitigate this bottleneck is to leverage the inherent sparsity in the attention computation. In our analysis of attention patterns in VLMs, we observe that a substantial portion of layers exhibit minimal cross-image attention, except through attention-sink tokens per image. These sparse attention patterns fall into distinct categories: sink-only, document mask and a hybrid document-sink mask. Based on this, we propose BlindSight: a training-free approach to optimize VLM inference using a input template-aware attention sparsity mask. We utilize samples from a dataset to derive a prompt-agnostic sparsity categorization for every attention head. We evaluate the proposed technique using VLMs such as Qwen2-VL, Qwen2.5-VL and Gemma-3. BlindSight results in a 32%-41% reduction in FLOPs on average with -2%-+2% accuracy compared to the original model in most evaluated multi-image understanding benchmarks.

arXiv:2507.10095v1 Announce Type: new Abstract: CLIP has shown promising performance across many short-text tasks in a zero-shot manner. However, limited by the input length of the text encoder, CLIP struggles on under-stream tasks with long-text inputs (>77 tokens). To remedy this issue, we propose FIX-CLIP which includes three novel modules: (1) A dual-branch training pipeline that aligns short and long texts with masked and raw images respectively, which boosts the long-text representation while preserving the short-text ability. (2) Multiple learnable regional prompts with unidirectional masks in Transformer layers for regional information extraction. (3) A hierarchical feature alignment module in the intermediate encoder layers to promote the consistency of multi-scale features. Furthermore, we collect 30M images and utilize existing MLLMs to synthesize long-text captions for training. Extensive experiments show that FIX-CLIP achieves state-of-the-art performance on both long-text and short-text retrieval benchmarks. For downstream applications, we reveal that FIX-CLIP's text encoder delivers promising performance in a plug-and-play manner for diffusion models with long-text input.

arXiv:2507.09139v1 Announce Type: new Abstract: Human pose estimation traditionally relies on architectures that encode keypoint priors, limiting their generalization to novel poses or unseen keypoints. Recent language-guided approaches like LocLLM reformulate keypoint localization as a vision-language task, enabling zero-shot generalization through textual descriptions. However, LocLLM's linear projector fails to capture complex spatial-textual interactions critical for high-precision localization. To address this, we propose PoseLLM, the first Large Language Model (LLM)-based pose estimation framework that replaces the linear projector with a nonlinear MLP vision-language connector. This lightweight two-layer MLP with GELU activation enables hierarchical cross-modal feature transformation, enhancing the fusion of visual patches and textual keypoint descriptions. Trained exclusively on COCO data, PoseLLM achieves 77.8 AP on the COCO validation set, outperforming LocLLM by +0.4 AP, while maintaining strong zero-shot generalization on Human-Art and MPII. Our work demonstrates that a simple yet powerful nonlinear connector significantly boosts localization accuracy without sacrificing generalization, advancing the state-of-the-art in language-guided pose estimation. Code is available at https://github.com/Ody-trek/PoseLLM.

arXiv:2507.10442v1 Announce Type: new Abstract: Vision-language Models (VLMs) have emerged as general-purpose tools for addressing a variety of complex computer vision problems. Such models have been shown to be highly capable, but, at the same time, lacking some basic visual understanding skills. In this paper, we set out to understand the limitations of SoTA VLMs on fundamental visual tasks by constructing a series of tests that probe which components of design, specifically, may be lacking. Importantly, we go significantly beyond the current benchmarks, which simply measure the final performance of VLM response, by also comparing and contrasting it to the performance of probes trained directly on features obtained from the visual encoder, intermediate vision-language projection and LLM-decoder output. In doing so, we uncover shortcomings in VLMs and make a number of important observations about their capabilities, robustness and how they process visual information. We hope our insights will guide progress in further improving VLMs.

arXiv:2507.10213v1 Announce Type: new Abstract: Multimodal learning often encounters the under-optimized problem and may have worse performance than unimodal learning. Existing methods attribute this problem to the imbalanced learning between modalities and rebalance them through gradient modulation. However, they fail to explain why the dominant modality in multimodal models also underperforms that in unimodal learning. In this work, we reveal the optimization conflict between the modality encoder and modality fusion module in multimodal models. Specifically, we prove that the cross-modal fusion in multimodal models decreases the gradient passed back to each modality encoder compared with unimodal models. Consequently, the performance of each modality in the multimodal model is inferior to that in the unimodal model. To this end, we propose a disentangled gradient learning (DGL) framework to decouple the optimization of the modality encoder and modality fusion module in the multimodal model. DGL truncates the gradient back-propagated from the multimodal loss to the modality encoder and replaces it with the gradient from unimodal loss. Besides, DGL removes the gradient back-propagated from the unimodal loss to the modality fusion module. This helps eliminate the gradient interference between the modality encoder and modality fusion module while ensuring their respective optimization processes. Finally, extensive experiments on multiple types of modalities, tasks, and frameworks with dense cross-modal interaction demonstrate the effectiveness and versatility of the proposed DGL. Code is available at \href{https://github.com/shicaiwei123/ICCV2025-GDL}{https://github.com/shicaiwei123/ICCV2025-GDL}

arXiv:2507.10015v1 Announce Type: new Abstract: Foundation multi-modal models are often designed by stitching of multiple existing pretrained uni-modal models: for example, an image classifier with an autoregressive text model. This stitching process is performed by training a connector module that aims to align the representation-representation or representation-input spaces of these uni-modal models. However, given the complexity of training such connectors on large scale web-based datasets coupled with the ever-increasing number of available pretrained uni-modal models, the task of uni-modal models selection and subsequent connector module training becomes computationally demanding. To address this under-studied critical problem, we propose Hypernetwork Model Alignment (Hyma), a novel all-in-one solution for optimal uni-modal model selection and connector training by leveraging hypernetworks. Specifically, our framework utilizes the parameter prediction capability of a hypernetwork to obtain jointly trained connector modules for $N \times M$ combinations of uni-modal models. In our experiments, Hyma reduces the optimal uni-modal model pair search cost by $10\times$ (averaged across all experiments), while matching the ranking and trained connector performance obtained via grid search across a suite of diverse multi-modal benchmarks.

arXiv:2507.09500v1 Announce Type: new Abstract: Vision-language models (VLMs) exhibit remarkable zero-shot capabilities but struggle with distribution shifts in downstream tasks when labeled data is unavailable, which has motivated the development of Test-Time Adaptation (TTA) to improve VLMs' performance during inference without annotations. Among various TTA approaches, cache-based methods show promise by preserving historical knowledge from low-entropy samples in a dynamic cache and fostering efficient adaptation. However, these methods face two critical reliability challenges: (1) entropy often becomes unreliable under distribution shifts, causing error accumulation in the cache and degradation in adaptation performance; (2) the final predictions may be unreliable due to inflexible decision boundaries that fail to accommodate large downstream shifts. To address these challenges, we propose a Reliable Test-time Adaptation (ReTA) method that integrates two complementary strategies to enhance reliability from two perspectives. First, to mitigate the unreliability of entropy as a sample selection criterion for cache construction, we introduce Consistency-aware Entropy Reweighting (CER), which incorporates consistency constraints to weight entropy during cache updating. While conventional approaches rely solely on low entropy for cache prioritization and risk introducing noise, our method leverages predictive consistency to maintain a high-quality cache and facilitate more robust adaptation. Second, we present Diversity-driven Distribution Calibration (DDC), which models class-wise text embeddings as multivariate Gaussian distributions, enabling adaptive decision boundaries for more accurate predictions across visually diverse content. Extensive experiments demonstrate that ReTA consistently outperforms state-of-the-art methods, particularly under challenging real-world distribution shifts.

arXiv:2507.09910v1 Announce Type: new Abstract: Graphic design visually conveys information and data by creating and combining text, images and graphics. Two-stage methods that rely primarily on layout generation lack creativity and intelligence, making graphic design still labor-intensive. Existing diffusion-based methods generate non-editable graphic design files at image level with poor legibility in visual text rendering, which prevents them from achieving satisfactory and practical automated graphic design. In this paper, we propose Instructional Graphic Designer (IGD) to swiftly generate multimodal layers with editable flexibility with only natural language instructions. IGD adopts a new paradigm that leverages parametric rendering and image asset generation. First, we develop a design platform and establish a standardized format for multi-scenario design files, thus laying the foundation for scaling up data. Second, IGD utilizes the multimodal understanding and reasoning capabilities of MLLM to accomplish attribute prediction, sequencing and layout of layers. It also employs a diffusion model to generate image content for assets. By enabling end-to-end training, IGD architecturally supports scalability and extensibility in complex graphic design tasks. The superior experimental results demonstrate that IGD offers a new solution for graphic design.

arXiv:2507.09404v1 Announce Type: new Abstract: Large foundation models are typically trained on data from multiple domains, with the data mixture--the proportion of each domain used--playing a critical role in model performance. The standard approach to selecting this mixture relies on trial and error, which becomes impractical for large-scale pretraining. We propose a systematic method to determine the optimal data mixture for any target domain using scaling laws. Our approach accurately predicts the loss of a model of size $N$ trained with $D$ tokens and a specific domain weight vector $h$. We validate the universality of these scaling laws by demonstrating their predictive power in three distinct and large-scale settings: large language model (LLM), native multimodal model (NMM), and large vision models (LVM) pretraining. We further show that these scaling laws can extrapolate to new data mixtures and across scales: their parameters can be accurately estimated using a few small-scale training runs, and used to estimate the performance at larger scales and unseen domain weights. The scaling laws allow to derive the optimal domain weights for any target domain under a given training budget ($N$,$D$), providing a principled alternative to costly trial-and-error methods.

arXiv:2507.09937v1 Announce Type: new Abstract: Large language models are susceptible to memorizing repeated sequences, posing privacy and copyright concerns. A popular mitigation strategy is to remove memorized information from specific neurons post-hoc. However, such approaches have shown limited success so far. In a controlled setting, we show that the memorization of natural sequences (those that resemble linguistically plausible text) become mechanistically entangled with general language abilities, thereby becoming challenging to remove post-hoc. In this work, we put forward a new paradigm of MemSinks that promotes isolation of memorization by design. We leverage a sequence identifier that activates a unique set of memorization neurons for each sequence across repetitions. By analyzing the dynamics of learning and forgetting, we argue that MemSinks facilitates isolation of memorized content, making it easier to remove without compromising general language capabilities. We implement MemSinks at the billion-parameter and billion-token scale, and observe both effective isolation and strong generalization. To our knowledge, this is the first proof-of-concept on real data demonstrating that simultaneous generalization and isolation is achievable. We open-source our code at http://github.com/grghosal/MemSinks.

arXiv:2507.09025v1 Announce Type: new Abstract: We propose Lizard, a linearization framework that transforms pretrained Transformer-based Large Language Models (LLMs) into flexible, subquadratic architectures for infinite-context generation. Transformer-based LLMs face significant memory and computational bottlenecks as context lengths increase, due to the quadratic complexity of softmax attention and the growing key-value (KV) cache. Lizard addresses these limitations by introducing a subquadratic attention mechanism that closely approximates softmax attention while preserving the output quality. Unlike previous linearization methods, which are often limited by fixed model structures and therefore exclude gating mechanisms, Lizard incorporates a gating module inspired by recent state-of-the-art linear models. This enables adaptive memory control, supports constant-memory inference, offers strong length generalization, and allows more flexible model design. Lizard combines gated linear attention for global context compression with sliding window attention enhanced by meta memory, forming a hybrid mechanism that captures both long-range dependencies and fine-grained local interactions. Moreover, we introduce a hardware-aware algorithm that accelerates the training speed of our models. Extensive experiments show that Lizard achieves near-lossless recovery of the teacher model's performance across standard language modeling tasks, while significantly outperforming previous linearization methods. On the 5-shot MMLU benchmark, Lizard improves over prior models by 18 points and shows significant improvements on associative recall tasks.

arXiv:2507.09185v1 Announce Type: new Abstract: Large language models (LLMs) often develop learned mechanisms specialized to specific datasets, such as reliance on domain-specific correlations, which yield high-confidence predictions without generalizable reasoning. While beneficial in one setting, these dataset-specific mechanisms typically degrade performance when models encounter novel tasks or distributions. In this work, we introduce a fine-tuning approach designed to enhance generalization by identifying and pruning neurons associated with dataset-specific mechanisms in transformer-based LLMs. Our method employs Integrated Gradients to quantify each neuron's influence on high-confidence predictions, pinpointing those that disproportionately contribute to dataset-specific performance without supporting robust, transferable reasoning. Selectively pruning these neurons compels the model to depend on generalizable representations. Evaluated across multiple-choice benchmarks, our pruning-based fine-tuning significantly enhances performance, surpassing prior (non-pruning) adaptation methods.

arXiv:2507.09955v1 Announce Type: new Abstract: DeepSeek, a Chinese Artificial Intelligence (AI) startup, has released their V3 and R1 series models, which attracted global attention due to their low cost, high performance, and open-source advantages. This paper begins by reviewing the evolution of large AI models focusing on paradigm shifts, the mainstream Large Language Model (LLM) paradigm, and the DeepSeek paradigm. Subsequently, the paper highlights novel algorithms introduced by DeepSeek, including Multi-head Latent Attention (MLA), Mixture-of-Experts (MoE), Multi-Token Prediction (MTP), and Group Relative Policy Optimization (GRPO). The paper then explores DeepSeek engineering breakthroughs in LLM scaling, training, inference, and system-level optimization architecture. Moreover, the impact of DeepSeek models on the competitive AI landscape is analyzed, comparing them to mainstream LLMs across various fields. Finally, the paper reflects on the insights gained from DeepSeek innovations and discusses future trends in the technical and engineering development of large AI models, particularly in data, training, and reasoning.

arXiv:2507.09839v1 Announce Type: new Abstract: An increasing number of NLP applications interact with large language models (LLMs) through black-box APIs, making prompt engineering critical for controlling model outputs. While recent Automatic Prompt Optimization (APO) methods iteratively refine prompts using model-generated feedback, textual gradients, they primarily focus on error correction and neglect valuable insights from correct predictions. This limits both their effectiveness and efficiency. In this paper, we propose a novel APO framework centered on enhancing the feedback mechanism. We reinterpret the textual gradient as a form of negative reinforcement and introduce the complementary positive reinforcement to explicitly preserve beneficial prompt components identified through successful predictions. To mitigate the noise inherent in LLM-generated feedback, we introduce a technique called feedback diversification, which aggregates multiple feedback signals, emphasizing consistent, actionable advice while filtering out outliers. Motivated by the rapid evolution and diversity of available LLMs, we also formalize Continual Prompt Optimization (CPO), addressing the practical challenge of efficiently migrating optimized prompts between different model versions or API providers. Our experiments reveal that naive prompt migration often degrades performance due to loss of critical instructions. In contrast, our approach consistently outperforms strong baselines, achieving significant accuracy improvements, faster convergence, and lower computational costs in both standard and migration scenarios.

arXiv:2507.10524v1 Announce Type: new Abstract: Scaling language models unlocks impressive capabilities, but the accompanying computational and memory demands make both training and deployment expensive. Existing efficiency efforts typically target either parameter sharing or adaptive computation, leaving open the question of how to attain both simultaneously. We introduce Mixture-of-Recursions (MoR), a unified framework that combines the two axes of efficiency inside a single Recursive Transformer. MoR reuses a shared stack of layers across recursion steps to achieve parameter efficiency, while lightweight routers enable adaptive token-level thinking by dynamically assigning different recursion depths to individual tokens. This allows MoR to focus quadratic attention computation only among tokens still active at a given recursion depth, further improving memory access efficiency by selectively caching only their key-value pairs. Beyond these core mechanisms, we also propose a KV sharing variant that reuses KV pairs from the first recursion, specifically designed to decrease prefill latency and memory footprint. Across model scales ranging from 135M to 1.7B parameters, MoR forms a new Pareto frontier: at equal training FLOPs and smaller model sizes, it significantly lowers validation perplexity and improves few-shot accuracy, while delivering higher throughput compared with vanilla and existing recursive baselines. These gains demonstrate that MoR is an effective path towards large-model quality without incurring large-model cost.

arXiv:2507.08967v1 Announce Type: new Abstract: Model steering represents a powerful technique that dynamically aligns large language models (LLMs) with human preferences during inference. However, conventional model-steering methods rely heavily on externally annotated data, not only limiting their adaptability to varying contexts but also tethering their effectiveness to annotation quality. In this paper, we present SIMS, the first self-improving model-steering framework that operates without relying on external supervision. At its core, SIMS autonomously generates and refines contrastive samples through iterative self-improvement cycles, enabling adaptive, context-specific steering. Additionally, SIMS employs novel strategies, including prompt ranking and contrast sampling, to further enhance steering efficacy. Extensive evaluation across diverse LLMs and benchmarks demonstrates that SIMS substantially outperforms existing methods in steering effectiveness and adaptability, highlighting self-improving model steering as a promising direction for future research on inference-time LLM alignment.

arXiv:2507.09850v1 Announce Type: new Abstract: Reasoning-capable language models achieve state-of-the-art performance in diverse complex tasks by generating long, explicit Chain-of-Thought (CoT) traces. While recent works show that base models can acquire such reasoning traces via reinforcement learning or distillation from stronger models like DeepSeek-R1, previous works demonstrate that even short CoT prompting without fine-tuning is able to improve reasoning. We ask whether long CoT can be induced in a base model using only prompting or minimal tuning. Using just 20 long CoT examples from the reasoning model \texttt{QwQ-32B-Preview}, we lightly fine-tune the base model \texttt{Qwen2.5-32B}. The resulting model outperforms the much larger \texttt{Qwen2.5-Math-72B-Instruct}, showing that a handful of high-quality examples can unlock strong reasoning capabilities. We further explore using CoT data from non-reasoning models and human annotators, enhanced with prompt engineering, multi-pass editing, and structural guidance. However, neither matches the performance of reasoning model traces, suggesting that certain latent qualities of expert CoT are difficult to replicate. We analyze key properties of reasoning data, such as problem difficulty, diversity, and answer length, that influence reasoning distillation. While challenges remain, we are optimistic that carefully curated human-written CoT, even in small quantities, can activate reasoning behaviors in base models. We release our human-authored dataset across refinement stages and invite further investigation into what makes small-scale reasoning supervision so effective.

arXiv:2507.10532v1 Announce Type: new Abstract: The reasoning capabilities of large language models (LLMs) have been a longstanding focus of research. Recent works have further enhanced these capabilities using reinforcement learning (RL), with many new methods claiming significant improvements with minimal or no external supervision. Surprisingly, some studies even suggest that random or incorrect reward signals can enhance reasoning performance. However, these breakthroughs are mostly reported on the Qwen2.5 model family and evaluated on well-known benchmarks such as MATH-500, AMC, and AIME, while failing to achieve similar gains on other models like Llama, which warrants further investigation. Our analysis shows that although Qwen2.5 achieves strong mathematical reasoning performance, its pretraining on large-scale web corpora makes it vulnerable to data contamination in popular benchmarks. As a result, results derived from these benchmarks may be unreliable. To address this, we introduce a generator that produces fully synthetic arithmetic problems of arbitrary length and difficulty, yielding a clean dataset we call RandomCalculation. Using these leakage-free datasets, we show that only accurate reward signals consistently improve performance, while noisy or incorrect signals do not. We advocate for evaluating RL methods on uncontaminated benchmarks and across diverse model families to ensure trustworthy conclusions.

arXiv:2507.09709v1 Announce Type: new Abstract: Understanding the latent space geometry of large language models (LLMs) is key to interpreting their behavior and improving alignment. \baturay{However, it remains unclear to what extent LLMs internally organize representations related to semantic understanding. To investigate this, we conduct a large-scale empirical study of hidden states in transformer-based LLMs, analyzing 11 decoder-only models across 6 scientific topics and 12 layers each. We find that high-level semantic information consistently lies in low-dimensional subspaces that form linearly separable representations across distinct domains. This separability becomes more pronounced in deeper layers and under prompts that trigger structured reasoning or alignment behaviors$\unicode{x2013}$even when surface content is unchanged. This geometry enables simple yet effective causal interventions in hidden space; for example, reasoning patterns like chain-of-thought can be captured by a single vector direction. Together, these findings support the development of geometry-aware tools that operate directly on latent representations to detect and mitigate harmful or adversarial content, using methods such as transport-based defenses that leverage this separability. As a proof of concept, we demonstrate this potential by training a simple MLP classifier as a lightweight latent-space guardrail, which detects adversarial and malicious prompts with high precision.

arXiv:2507.09406v1 Announce Type: new Abstract: Large language models (LLMs) aligned for safety through techniques like reinforcement learning from human feedback (RLHF) often exhibit emergent deceptive behaviors, where outputs appear compliant but subtly mislead or omit critical information. This paper introduces adversarial activation patching, a novel mechanistic interpretability framework that leverages activation patching as an adversarial tool to induce, detect, and mitigate such deception in transformer-based models. By sourcing activations from "deceptive" prompts and patching them into safe forward passes at specific layers, we simulate vulnerabilities and quantify deception rates. Through toy neural network simulations across multiple scenarios (e.g., 1000 trials per setup), we demonstrate that adversarial patching increases deceptive outputs to 23.9% from a 0% baseline, with layer-specific variations supporting our hypotheses. We propose six hypotheses, including transferability across models, exacerbation in multimodal settings, and scaling effects. An expanded literature review synthesizes over 20 key works in interpretability, deception, and adversarial attacks. Mitigation strategies, such as activation anomaly detection and robust fine-tuning, are detailed, alongside ethical considerations and future research directions. This work advances AI safety by highlighting patching's dual-use potential and provides a roadmap for empirical studies on large-scale models.

arXiv:2507.09482v1 Announce Type: new Abstract: Human emotions are complex, with sarcasm being a subtle and distinctive form. Despite progress in sarcasm research, sarcasm generation remains underexplored, primarily due to the overreliance on textual modalities and the neglect of visual cues, as well as the mismatch between image content and sarcastic intent in existing datasets. In this paper, we introduce M2SaG, a multimodal sarcasm generation dataset with 4,970 samples, each containing an image, a sarcastic text, and a sarcasm target. To benchmark M2SaG, we propose ViSP, a generation framework that integrates Proximal Policy Optimization (PPO) and contrastive learning. PPO utilizes reward scores from DIP to steer the generation of sarcastic texts, while contrastive learning encourages the model to favor outputs with higher reward scores. These strategies improve overall generation quality and produce texts with more pronounced sarcastic intent. We evaluate ViSP across five metric sets and find it surpasses all baselines, including large language models, underscoring their limitations in sarcasm generation. Furthermore, we analyze the distributions of Sarcasm Scores and Factual Incongruity for both M2SaG and the texts generated by ViSP. The generated texts exhibit higher mean Sarcasm Scores (0.898 vs. 0.770) and Factual Incongruity (0.768 vs. 0.739), demonstrating that ViSP produces higher-quality sarcastic content than the original dataset. % The dataset and code will be publicly available. Our dataset and code will be released at \textit{https://github.com/wclapply/ViSP}.

arXiv:2507.09751v1 Announce Type: new Abstract: Large language models (LLMs) have demonstrated impressive capabilities in natural language understanding and generation, but they exhibit problems with logical consistency in the output they generate. How can we harness LLMs' broad-coverage parametric knowledge in formal reasoning despite their inconsistency? We present a method for directly integrating an LLM into the interpretation function of the formal semantics for a paraconsistent logic. We provide experimental evidence for the feasibility of the method by evaluating the function using datasets created from several short-form factuality benchmarks. Unlike prior work, our method offers a theoretical framework for neuro-symbolic reasoning that leverages an LLM's knowledge while preserving the underlying logic's soundness and completeness properties.

arXiv:2507.10540v1 Announce Type: new Abstract: The rapid advancement of large language models (LLMs) has created a vibrant ecosystem of diverse architectures, each with unique strengths due to differences in design, training data, and objectives. However, most applications still rely on a single backend model, limiting coverage of capabilities and leading to inefficiencies in performance and token cost when tackling complex tasks. We highlight an underexploited opportunity: LLM routing data, produced when hosting platforms route diverse queries to different models, which can reveal comparative strengths across tasks. To address this, we propose FusionBench, a comprehensive routing benchmark covering 14 tasks across five domains with 20 open-source LLMs (8B to 671B parameters), capturing 103M tokens and summarizing reusable thought templates from top models. Building on this, we introduce FusionFactory, a systematic fusion framework with three levels: (1) query-level fusion, tailoring routers for each query using both direct responses and reasoning-augmented outputs; (2) thought-level fusion, leveraging abstract templates derived from top-performing LLMs' answers to similar queries; and (3) model-level fusion, transferring capabilities between models via distillation, using top responses or highest judge scores as training data. Experiments show FusionFactory consistently outperforms the best individual LLM across all 14 benchmarks, with optimal fusion configurations varying by benchmark, demonstrating the value of systematic LLM fusion in harnessing complementary strengths and improving overall performance.

arXiv:2507.09616v1 Announce Type: new Abstract: Deploying transformer-based neural networks on resource-constrained edge devices presents a significant challenge. This challenge is often addressed through various techniques, such as low-rank approximation and mixed-precision quantization. In this work, we introduce Mixed Low-Rank and Quantization (MLoRQ), a novel method that integrates both techniques. MLoRQ employs a two-stage optimization process to determine optimal bit-width and rank assignments for each layer, adhering to predefined memory constraints. This process includes: (i) an intra-layer optimization that identifies potentially optimal compression solutions out of all low-rank and quantization combinations; (ii) an inter-layer optimization that assigns bit-width precision and rank to each layer while ensuring the memory constraint is met. An optional final step applies a sequential optimization process using a modified adaptive rounding technique to mitigate compression-induced errors in joint low-rank approximation and quantization. The method is compatible and can be seamlessly integrated with most existing quantization algorithms. MLoRQ shows state-of-the-art results with up to 15% performance improvement, evaluated on Vision Transformers for image classification, object detection, and instance segmentation tasks.

arXiv:2507.08833v1 Announce Type: new Abstract: Low-Rank Adaptation (LoRA) is one of the most widely used techniques for fine-tuning large language models (LLMs). By introducing a small number of trainable low-rank weight matrices, LoRA substantially reduces the number of parameters that need to be updated, offering significant advantages in memory consumption and computational efficiency compared to full fine-tuning. However, we observed that LoRA does not consistently provide speed improvements across all model architectures and training setups. Motivated by this inconsistency, we conduct a comprehensive analysis of LoRA's performance and investigate the underlying factors limiting its speedup. Based on our findings, we propose several methods for more efficient fine-tuning of LLMs. We empirically evaluate these methods and compare them to LoRA, demonstrating that our approach achieves comparable or superior performance while delivering more consistent training speed improvements. Our work offers valuable insights and practical guidelines for practitioners seeking to optimize LLM fine-tuning under resource constraints.

arXiv:2507.09931v1 Announce Type: new Abstract: The integration of Large Language Models (LLMs) into safety-critical domains, such as nuclear engineering, necessitates a deep understanding of their internal reasoning processes. This paper presents a novel methodology for interpreting how an LLM encodes and utilizes domain-specific knowledge, using a Boiling Water Reactor system as a case study. We adapted a general-purpose LLM (Gemma-3-1b-it) to the nuclear domain using a parameter-efficient fine-tuning technique known as Low-Rank Adaptation. By comparing the neuron activation patterns of the base model to those of the fine-tuned model, we identified a sparse set of neurons whose behavior was significantly altered during the adaptation process. To probe the causal role of these specialized neurons, we employed a neuron silencing technique. Our results demonstrate that while silencing most of these specialized neurons individually did not produce a statistically significant effect, deactivating the entire group collectively led to a statistically significant degradation in task performance. Qualitative analysis further revealed that silencing these neurons impaired the model's ability to generate detailed, contextually accurate technical information. This paper provides a concrete methodology for enhancing the transparency of an opaque black-box model, allowing domain expertise to be traced to verifiable neural circuits. This offers a pathway towards achieving nuclear-grade artificial intelligence (AI) assurance, addressing the verification and validation challenges mandated by nuclear regulatory frameworks (e.g., 10 CFR 50 Appendix B), which have limited AI deployment in safety-critical nuclear operations.

arXiv:2507.09846v1 Announce Type: new Abstract: As both model and dataset sizes continue to scale rapidly, conventional pretraining strategies with fixed compute budgets-such as cosine learning rate schedules-are increasingly inadequate for large-scale training. Recent alternatives, including warmup-stable-decay (WSD) schedules and weight averaging, offer greater flexibility. However, WSD relies on explicit decay phases to track progress, while weight averaging addresses this limitation at the cost of additional memory. In search of a more principled and scalable alternative, we revisit the Schedule-Free (SF) method [Defazio et al., 2024], which has shown strong empirical performance across diverse settings. We show that SF-AdamW effectively navigates the "river" structure of the loss landscape without decay phases or auxiliary averaging, making it particularly suitable for continuously scaling training workloads. To understand this behavior, we conduct a theoretical and empirical analysis of SF dynamics, revealing that it implicitly performs weight averaging without memory overhead. Guided by this analysis, we propose a refined variant of SF that improves robustness to momentum and performs better under large batch sizes, addressing key limitations of the original method. Together, these results establish SF as a practical, scalable, and theoretically grounded approach for language model training.

arXiv:2507.10445v1 Announce Type: new Abstract: In this work we examine LLMs' ability to ask clarification questions in task-oriented dialogues that follow the asynchronous instruction-giver/instruction-follower format. We present a new corpus that combines two existing annotations of the Minecraft Dialogue Corpus -- one for reference and ambiguity in reference, and one for SDRT including clarifications -- into a single common format providing the necessary information to experiment with clarifications and their relation to ambiguity. With this corpus we compare LLM actions with original human-generated clarification questions, examining how both humans and LLMs act in the case of ambiguity. We find that there is only a weak link between ambiguity and humans producing clarification questions in these dialogues, and low correlation between humans and LLMs. Humans hardly ever produce clarification questions for referential ambiguity, but often do so for task-based uncertainty. Conversely, LLMs produce more clarification questions for referential ambiguity, but less so for task uncertainty. We question if LLMs' ability to ask clarification questions is predicated on their recent ability to simulate reasoning, and test this with different reasoning approaches, finding that reasoning does appear to increase question frequency and relevancy.

arXiv:2507.10326v1 Announce Type: new Abstract: Prompt engineering has proven to be a crucial step in leveraging pretrained large language models (LLMs) in solving various real-world tasks. Numerous solutions have been proposed that seek to automate prompt engineering by using the model itself to edit prompts. However, the majority of state-of-the-art approaches are evaluated on tasks that require minimal prompt templates and on very large and highly capable LLMs. In contrast, solving complex tasks that require detailed information to be included in the prompt increases the amount of text that needs to be optimised. Furthermore, smaller models have been shown to be more sensitive to prompt design. To address these challenges, we propose an evolutionary search approach to automated discrete prompt optimisation consisting of two phases. In the first phase, grammar-guided genetic programming is invoked to synthesise prompt-creating programmes by searching the space of programmes populated by function compositions of syntactic, dictionary-based and LLM-based prompt-editing functions. In the second phase, local search is applied to explore the neighbourhoods of best-performing programmes in an attempt to further fine-tune their performance. Our approach outperforms three state-of-the-art prompt optimisation approaches, PromptWizard, OPRO, and RL-Prompt, on three relatively small general-purpose LLMs in four domain-specific challenging tasks. We also illustrate several examples where these benchmark methods suffer relatively severe performance degradation, while our approach improves performance in almost all task-model combinations, only incurring minimal degradation when it does not.

arXiv:2507.10435v1 Announce Type: new Abstract: Recent studies suggest that large language models (LLMs) possess the capability to solve graph reasoning tasks. Notably, even when graph structures are embedded within textual descriptions, LLMs can still effectively answer related questions. This raises a fundamental question: How can a decoder-only Transformer architecture understand underlying graph structures? To address this, we start with the substructure extraction task, interpreting the inner mechanisms inside the transformers and analyzing the impact of the input queries. Specifically, through both empirical results and theoretical analysis, we present Induced Substructure Filtration (ISF), a perspective that captures the substructure identification in the multi-layer transformers. We further validate the ISF process in LLMs, revealing consistent internal dynamics across layers. Building on these insights, we explore the broader capabilities of Transformers in handling diverse graph types. Specifically, we introduce the concept of thinking in substructures to efficiently extract complex composite patterns, and demonstrate that decoder-only Transformers can successfully extract substructures from attributed graphs, such as molecular graphs. Together, our findings offer a new insight on how sequence-based Transformers perform the substructure extraction task over graph data.

arXiv:2507.09394v1 Announce Type: new Abstract: In this work, we study how multi-head latent attention (MLA), a popular strategy for compressing key/value memory, affects a transformer's internal capacity during pretraining. Using a lightweight suite of Marchenko-Pastur (MP) diagnostics, we analyze the spectrum of the $W_{Q}W_{K}^\top$ gram matrix throughout training, comparing three variants: the standard multi-head attention (MHA) baseline, MLA-PreRoPE with rotary applied before compression, and MLA-Decoupled, which shares a single rotary sub-vector across all heads. Our random matrix analysis reveals \textbf{three key findings:} \textbf{ i)} capacity bottlenecks emerge locally: both MHA and MLA-PreRoPE exhibit sharp, early spikes in specific layers that persist and propagate, disrupting the balance between bulk and outlier directions; \textbf{ ii)} these spikes coincide with rank collapse, concentrating the model's expressivity into narrow subspaces; \textbf{ iii)} only the decoupled variant prevents this cascade, maintaining broad spectral support and suppressing outlier formation across layers. These results underscore that \emph{how} rotary embeddings are applied is just as critical as \emph{where} compression occurs. Sharing rotary components across heads mitigates spectral fragmentation and preserves representational capacity.

arXiv:2507.10124v1 Announce Type: new Abstract: Identifying bias in LLMs is ongoing. Because they are still in development, what is true today may be false tomorrow. We therefore need general strategies for debiasing that will outlive current models. Strategies developed for debiasing human decision making offer one promising approach as they incorporate an LLM-style prompt intervention designed to bring latent knowledge into awareness during decision making. LLMs trained on vast amounts of information contain information about potential biases, counter-arguments, and contradictory evidence, but that information may only be brought to bear if prompted. Metacognitive prompts developed in the human decision making literature are designed to achieve this, and as I demonstrate here, they show promise with LLMs. The prompt I focus on here is "could you be wrong?" Following an LLM response, this prompt leads LLMs to produce additional information, including why they answered as they did, errors, biases, contradictory evidence, and alternatives, none of which were apparent in their initial response. Indeed, this metaknowledge often reveals that how LLMs and users interpret prompts are not aligned. Here I demonstrate this prompt using a set of questions taken from recent articles about LLM biases, including implicit discriminatory biases and failures of metacognition. "Could you be wrong" prompts the LLM to identify its own biases and produce cogent metacognitive reflection. I also present another example involving convincing but incomplete information, which is readily corrected by the metacognitive prompt. In sum, this work argues that human psychology offers a new avenue for prompt engineering, leveraging a long history of effective prompt-based improvements to human decision making.

arXiv:2507.10403v1 Announce Type: new Abstract: Retrieving relevant imagery from vast satellite archives is crucial for applications like disaster response and long-term climate monitoring. However, most text-to-image retrieval systems are limited to RGB data, failing to exploit the unique physical information captured by other sensors, such as the all-weather structural sensitivity of Synthetic Aperture Radar (SAR) or the spectral signatures in optical multispectral data. To bridge this gap, we introduce CrisisLandMark, a new large-scale corpus of over 647,000 Sentinel-1 SAR and Sentinel-2 multispectral images paired with structured textual annotations for land cover, land use, and crisis events harmonized from authoritative land cover systems (CORINE and Dynamic World) and crisis-specific sources. We then present CLOSP (Contrastive Language Optical SAR Pretraining), a novel framework that uses text as a bridge to align unpaired optical and SAR images into a unified embedding space. Our experiments show that CLOSP achieves a new state-of-the-art, improving retrieval nDGC by 54% over existing models. Additionally, we find that the unified training strategy overcomes the inherent difficulty of interpreting SAR imagery by transferring rich semantic knowledge from the optical domain with indirect interaction. Furthermore, GeoCLOSP, which integrates geographic coordinates into our framework, creates a powerful trade-off between generality and specificity: while the CLOSP excels at general semantic tasks, the GeoCLOSP becomes a specialized expert for retrieving location-dependent crisis events and rare geographic features. This work highlights that the integration of diverse sensor data and geographic context is essential for unlocking the full potential of remote sensing archives.

arXiv:2507.09574v1 Announce Type: new Abstract: Recent text-to-image models produce high-quality results but still struggle with precise visual control, balancing multimodal inputs, and requiring extensive training for complex multimodal image generation. To address these limitations, we propose MENTOR, a novel autoregressive (AR) framework for efficient Multimodal-conditioned Tuning for Autoregressive multimodal image generation. MENTOR combines an AR image generator with a two-stage training paradigm, enabling fine-grained, token-level alignment between multimodal inputs and image outputs without relying on auxiliary adapters or cross-attention modules. The two-stage training consists of: (1) a multimodal alignment stage that establishes robust pixel- and semantic-level alignment, followed by (2) a multimodal instruction tuning stage that balances the integration of multimodal inputs and enhances generation controllability. Despite modest model size, suboptimal base components, and limited training resources, MENTOR achieves strong performance on the DreamBench++ benchmark, outperforming competitive baselines in concept preservation and prompt following. Additionally, our method delivers superior image reconstruction fidelity, broad task adaptability, and improved training efficiency compared to diffusion-based methods. Dataset, code, and models are available at: https://github.com/HaozheZhao/MENTOR

arXiv:2507.09487v1 Announce Type: new Abstract: Visual and semantic concepts are often structured in a hierarchical manner. For instance, textual concept `cat' entails all images of cats. A recent study, MERU, successfully adapts multimodal learning techniques from Euclidean space to hyperbolic space, effectively capturing the visual-semantic hierarchy. However, a critical question remains: how can we more efficiently train a model to capture and leverage this hierarchy? In this paper, we propose the \textit{Hyperbolic Masked Image and Distillation Network} (HMID-Net), a novel and efficient method that integrates Masked Image Modeling (MIM) and knowledge distillation techniques within hyperbolic space. To the best of our knowledge, this is the first approach to leverage MIM and knowledge distillation in hyperbolic space to train highly efficient models. In addition, we introduce a distillation loss function specifically designed to facilitate effective knowledge transfer in hyperbolic space. Our experiments demonstrate that MIM and knowledge distillation techniques in hyperbolic space can achieve the same remarkable success as in Euclidean space. Extensive evaluations show that our method excels across a wide range of downstream tasks, significantly outperforming existing models like MERU and CLIP in both image classification and retrieval.

arXiv:2507.09537v1 Announce Type: new Abstract: Predicting the future of surrounding agents and accordingly planning a safe, goal-directed trajectory are crucial for automated vehicles. Current methods typically rely on imitation learning to optimize metrics against the ground truth, often overlooking how scene understanding could enable more holistic trajectories. In this paper, we propose Plan-MAE, a unified pretraining framework for prediction and planning that capitalizes on masked autoencoders. Plan-MAE fuses critical contextual understanding via three dedicated tasks: reconstructing masked road networks to learn spatial correlations, agent trajectories to model social interactions, and navigation routes to capture destination intents. To further align vehicle dynamics and safety constraints, we incorporate a local sub-planning task predicting the ego-vehicle's near-term trajectory segment conditioned on earlier segment. This pretrained model is subsequently fine-tuned on downstream tasks to jointly generate the prediction and planning trajectories. Experiments on large-scale datasets demonstrate that Plan-MAE outperforms current methods on the planning metrics by a large margin and can serve as an important pre-training step for learning-based motion planner.

arXiv:2507.09308v1 Announce Type: new Abstract: Recent advances in latent diffusion models have achieved remarkable results in high-fidelity RGB image synthesis by leveraging pretrained VAEs to compress and reconstruct pixel data at low computational cost. However, the generation of transparent or layered content (RGBA image) remains largely unexplored, due to the lack of large-scale benchmarks. In this work, we propose ALPHA, the first comprehensive RGBA benchmark that adapts standard RGB metrics to four-channel images via alpha blending over canonical backgrounds. We further introduce ALPHAVAE, a unified end-to-end RGBA VAE that extends a pretrained RGB VAE by incorporating a dedicated alpha channel. The model is trained with a composite objective that combines alpha-blended pixel reconstruction, patch-level fidelity, perceptual consistency, and dual KL divergence constraints to ensure latent fidelity across both RGB and alpha representations. Our RGBA VAE, trained on only 8K images in contrast to 1M used by prior methods, achieves a +4.9 dB improvement in PSNR and a +3.2% increase in SSIM over LayerDiffuse in reconstruction. It also enables superior transparent image generation when fine-tuned within a latent diffusion framework. Our code, data, and models are released on https://github.com/o0o0o00o0/AlphaVAE for reproducibility.

arXiv:2507.10072v1 Announce Type: new Abstract: Diffusion models exhibit impressive generative capabilities but are significantly impacted by exposure bias. In this paper, we make a key observation: the energy of the predicted noisy images decreases during the diffusion process. Building on this, we identify two important findings: 1) The reduction in energy follows distinct patterns in the low-frequency and high-frequency subbands; 2) This energy reduction results in amplitude variations between the network-reconstructed clean data and the real clean data. Based on the first finding, we introduce a frequency-domain regulation mechanism utilizing wavelet transforms, which separately adjusts the low- and high-frequency subbands. Leveraging the second insight, we provide a more accurate analysis of exposure bias in the two subbands. Our method is training-free and plug-and-play, significantly improving the generative quality of various diffusion models and providing a robust solution to exposure bias across different model architectures. The source code is available at https://github.com/kunzhan/wpp.

arXiv:2507.09984v1 Announce Type: new Abstract: In spite of remarkable potential of the Latent Diffusion Models (LDMs) in image generation, the desired properties and optimal design of the autoencoders have been underexplored. In this work, we analyze the role of autoencoders in LDMs and identify three key properties: latent smoothness, perceptual compression quality, and reconstruction quality. We demonstrate that existing autoencoders fail to simultaneously satisfy all three properties, and propose Variational Masked AutoEncoders (VMAEs), taking advantage of the hierarchical features maintained by Masked AutoEncoder. We integrate VMAEs into the LDM framework, introducing Latent Diffusion Models with Masked AutoEncoders (LDMAEs). Through comprehensive experiments, we demonstrate significantly enhanced image generation quality and computational efficiency.

arXiv:2507.10217v1 Announce Type: new Abstract: Recent diffusion models achieve personalization by learning specific subjects, allowing learned attributes to be integrated into generated images. However, personalized human image generation remains challenging due to the need for precise and consistent attribute preservation (e.g., identity, clothing details). Existing subject-driven image generation methods often require either (1) inference-time fine-tuning with few images for each new subject or (2) large-scale dataset training for generalization. Both approaches are computationally expensive and impractical for real-time applications. To address these limitations, we present Wardrobe Polyptych LoRA, a novel part-level controllable model for personalized human image generation. By training only LoRA layers, our method removes the computational burden at inference while ensuring high-fidelity synthesis of unseen subjects. Our key idea is to condition the generation on the subject's wardrobe and leverage spatial references to reduce information loss, thereby improving fidelity and consistency. Additionally, we introduce a selective subject region loss, which encourages the model to disregard some of reference images during training. Our loss ensures that generated images better align with text prompts while maintaining subject integrity. Notably, our Wardrobe Polyptych LoRA requires no additional parameters at the inference stage and performs generation using a single model trained on a few training samples. We construct a new dataset and benchmark tailored for personalized human image generation. Extensive experiments show that our approach significantly outperforms existing techniques in fidelity and consistency, enabling realistic and identity-preserving full-body synthesis.

arXiv:2507.08980v1 Announce Type: new Abstract: Diffusion models can be improved with additional guidance towards more effective representations of input. Indeed, prior empirical work has already shown that aligning internal representations of the diffusion model with those of pre-trained models improves generation quality. In this paper, we present a systematic framework for incorporating representation guidance into diffusion models. We provide alternative decompositions of denoising models along with their associated training criteria, where the decompositions determine when and how the auxiliary representations are incorporated. Guided by our theoretical insights, we introduce two new strategies for enhancing representation alignment in diffusion models. First, we pair examples with target representations either derived from themselves or arisen from different synthetic modalities, and subsequently learn a joint model over the multimodal pairs. Second, we design an optimal training curriculum that balances representation learning and data generation. Our experiments across image, protein sequence, and molecule generation tasks demonstrate superior performance as well as accelerated training. In particular, on the class-conditional ImageNet $256\times 256$ benchmark, our guidance results in $23.3$ times faster training than the original SiT-XL as well as four times speedup over the state-of-the-art method REPA. The code is available at https://github.com/ChenyuWang-Monica/REED.

arXiv:2507.09733v1 Announce Type: new Abstract: We present the first foundational AI model for universal physics simulation that learns physical laws directly from boundary-condition data without requiring a priori equation encoding. Traditional physics-informed neural networks (PINNs) and finite-difference methods necessitate explicit mathematical formulation of governing equations, fundamentally limiting their generalizability and discovery potential. Our sketch-guided diffusion transformer approach reimagines computational physics by treating simulation as a conditional generation problem, where spatial boundary conditions guide the synthesis of physically accurate steady-state solutions. By leveraging enhanced diffusion transformer architectures with novel spatial relationship encoding, our model achieves direct boundary-to-equilibrium mapping and is generalizable to diverse physics domains. Unlike sequential time-stepping methods that accumulate errors over iterations, our approach bypasses temporal integration entirely, directly generating steady-state solutions with SSIM > 0.8 while maintaining sub-pixel boundary accuracy. Our data-informed approach enables physics discovery through learned representations analyzable via Layer-wise Relevance Propagation (LRP), revealing emergent physical relationships without predetermined mathematical constraints. This work represents a paradigm shift from AI-accelerated physics to AI-discovered physics, establishing the first truly universal physics simulation framework.

arXiv:2507.09212v1 Announce Type: new Abstract: Iterative generative models, like diffusion and flow-matching, create high-fidelity samples by progressively refining a noise vector into data. However, this process is notoriously slow, often requiring hundreds of function evaluations. We introduce the warm-start model, a simple, deterministic model that dramatically accelerates conditional generation by providing a better starting point. Instead of starting generation from an uninformed N(0, I) prior, our warm-start model predicts an informed prior N(mu, sigma), whose moments are conditioned on the input context. This "warm start" substantially reduces the distance the generative process must traverse, particularly when the conditioning information is strongly informative. On tasks like image inpainting, our method achieves results competitive with a 1000-step DDPM baseline using only 11 total function evaluations (1 for the warm start, 10 for generation). A simple conditional normalization trick makes our method compatible with any standard generative model and sampler without modification, allowing it to be combined with other efficient sampling techniques for further acceleration. Our implementation is available at https://github.com/jonas-scholz123/warm-start-model.

arXiv:2507.09574v1 Announce Type: new Abstract: Recent text-to-image models produce high-quality results but still struggle with precise visual control, balancing multimodal inputs, and requiring extensive training for complex multimodal image generation. To address these limitations, we propose MENTOR, a novel autoregressive (AR) framework for efficient Multimodal-conditioned Tuning for Autoregressive multimodal image generation. MENTOR combines an AR image generator with a two-stage training paradigm, enabling fine-grained, token-level alignment between multimodal inputs and image outputs without relying on auxiliary adapters or cross-attention modules. The two-stage training consists of: (1) a multimodal alignment stage that establishes robust pixel- and semantic-level alignment, followed by (2) a multimodal instruction tuning stage that balances the integration of multimodal inputs and enhances generation controllability. Despite modest model size, suboptimal base components, and limited training resources, MENTOR achieves strong performance on the DreamBench++ benchmark, outperforming competitive baselines in concept preservation and prompt following. Additionally, our method delivers superior image reconstruction fidelity, broad task adaptability, and improved training efficiency compared to diffusion-based methods. Dataset, code, and models are available at: https://github.com/HaozheZhao/MENTOR

arXiv:2507.10547v1 Announce Type: new Abstract: Visual tokenizers are pivotal in multimodal large models, acting as bridges between continuous inputs and discrete tokens. Nevertheless, training high-compression-rate VQ-VAEs remains computationally demanding, often necessitating thousands of GPU hours. This work demonstrates that a pre-trained VAE can be efficiently transformed into a VQ-VAE by controlling quantization noise within the VAE's tolerance threshold. We present \textbf{Quantize-then-Rectify (ReVQ)}, a framework leveraging pre-trained VAEs to enable rapid VQ-VAE training with minimal computational overhead. By integrating \textbf{channel multi-group quantization} to enlarge codebook capacity and a \textbf{post rectifier} to mitigate quantization errors, ReVQ compresses ImageNet images into at most 512 tokens while sustaining competitive reconstruction quality (rFID = 1.06). Significantly, ReVQ reduces training costs by over two orders of magnitude relative to state-of-the-art approaches: ReVQ finishes full training on a single NVIDIA 4090 in approximately 22 hours, whereas comparable methods require 4.5 days on 32 A100 GPUs. Experimental results show that ReVQ achieves superior efficiency-reconstruction trade-offs.

arXiv:2507.10029v1 Announce Type: new Abstract: Memory-efficient personalization is critical for adapting text-to-image diffusion models while preserving user privacy and operating within the limited computational resources of edge devices. To this end, we propose a selective optimization framework that adaptively chooses between backpropagation on low-resolution images (BP-low) and zeroth-order optimization on high-resolution images (ZO-high), guided by the characteristics of the diffusion process. As observed in our experiments, BP-low efficiently adapts the model to target-specific features, but suffers from structural distortions due to resolution mismatch. Conversely, ZO-high refines high-resolution details with minimal memory overhead but faces slow convergence when applied without prior adaptation. By complementing both methods, our framework leverages BP-low for effective personalization while using ZO-high to maintain structural consistency, achieving memory-efficient and high-quality fine-tuning. To maximize the efficacy of both BP-low and ZO-high, we introduce a timestep-aware probabilistic function that dynamically selects the appropriate optimization strategy based on diffusion timesteps. This function mitigates the overfitting from BP-low at high timesteps, where structural information is critical, while ensuring ZO-high is applied more effectively as training progresses. Experimental results demonstrate that our method achieves competitive performance while significantly reducing memory consumption, enabling scalable, high-quality on-device personalization without increasing inference latency.

arXiv:2507.09168v1 Announce Type: new Abstract: Text-guided image and 3D editing have advanced with diffusion-based models, yet methods like Delta Denoising Score often struggle with stability, spatial control, and editing strength. These limitations stem from reliance on complex auxiliary structures, which introduce conflicting optimization signals and restrict precise, localized edits. We introduce Stable Score Distillation (SSD), a streamlined framework that enhances stability and alignment in the editing process by anchoring a single classifier to the source prompt. Specifically, SSD utilizes Classifier-Free Guidance (CFG) equation to achieves cross-prompt alignment, and introduces a constant term null-text branch to stabilize the optimization process. This approach preserves the original content's structure and ensures that editing trajectories are closely aligned with the source prompt, enabling smooth, prompt-specific modifications while maintaining coherence in surrounding regions. Additionally, SSD incorporates a prompt enhancement branch to boost editing strength, particularly for style transformations. Our method achieves state-of-the-art results in 2D and 3D editing tasks, including NeRF and text-driven style edits, with faster convergence and reduced complexity, providing a robust and efficient solution for text-guided editing.

arXiv:2507.10340v1 Announce Type: new Abstract: Despite the success of diffusion models in image generation tasks such as text-to-image, the enormous computational complexity of diffusion models limits their use in resource-constrained environments. To address this, network quantization has emerged as a promising solution for designing efficient diffusion models. However, existing diffusion model quantization methods do not consider input conditions, such as text prompts, as an essential source of information for quantization. In this paper, we propose a novel quantization method dubbed Quantization of Language-to-Image diffusion models using text Prompts (QLIP). QLIP leverages text prompts to guide the selection of bit precision for every layer at each time step. In addition, QLIP can be seamlessly integrated into existing quantization methods to enhance quantization efficiency. Our extensive experiments demonstrate the effectiveness of QLIP in reducing computational complexity and improving the quality of the generated images across various datasets.

arXiv:2507.09595v1 Announce Type: new Abstract: FLUX.1 is a diffusion-based text-to-image generation model developed by Black Forest Labs, designed to achieve faithful text-image alignment while maintaining high image quality and diversity. FLUX is considered state-of-the-art in text-to-image generation, outperforming popular models such as Midjourney, DALL-E 3, Stable Diffusion 3 (SD3), and SDXL. Although publicly available as open source, the authors have not released official technical documentation detailing the model's architecture or training setup. This report summarizes an extensive reverse-engineering effort aimed at demystifying FLUX's architecture directly from its source code, to support its adoption as a backbone for future research and development. This document is an unofficial technical report and is not published or endorsed by the original developers or their affiliated institutions.

arXiv:2507.09915v1 Announce Type: new Abstract: The scarcity of data in various scenarios, such as medical, industry and autonomous driving, leads to model overfitting and dataset imbalance, thus hindering effective detection and segmentation performance. Existing studies employ the generative models to synthesize more training samples to mitigate data scarcity. However, these synthetic samples are repetitive or simplistic and fail to provide "crucial information" that targets the downstream model's weaknesses. Additionally, these methods typically require separate training for different objects, leading to computational inefficiencies. To address these issues, we propose Crucial-Diff, a domain-agnostic framework designed to synthesize crucial samples. Our method integrates two key modules. The Scene Agnostic Feature Extractor (SAFE) utilizes a unified feature extractor to capture target information. The Weakness Aware Sample Miner (WASM) generates hard-to-detect samples using feedback from the detection results of downstream model, which is then fused with the output of SAFE module. Together, our Crucial-Diff framework generates diverse, high-quality training data, achieving a pixel-level AP of 83.63% and an F1-MAX of 78.12% on MVTec. On polyp dataset, Crucial-Diff reaches an mIoU of 81.64% and an mDice of 87.69%. Code will be released after acceptance.

arXiv:2507.10065v1 Announce Type: new Abstract: We present MoVieS, a novel feed-forward model that synthesizes 4D dynamic novel views from monocular videos in one second. MoVieS represents dynamic 3D scenes using pixel-aligned grids of Gaussian primitives, explicitly supervising their time-varying motion. This allows, for the first time, the unified modeling of appearance, geometry and motion, and enables view synthesis, reconstruction and 3D point tracking within a single learning-based framework. By bridging novel view synthesis with dynamic geometry reconstruction, MoVieS enables large-scale training on diverse datasets with minimal dependence on task-specific supervision. As a result, it also naturally supports a wide range of zero-shot applications, such as scene flow estimation and moving object segmentation. Extensive experiments validate the effectiveness and efficiency of MoVieS across multiple tasks, achieving competitive performance while offering several orders of magnitude speedups.

arXiv:2507.09052v1 Announce Type: new Abstract: Training data for class-conditional image synthesis often exhibit a long-tailed distribution with limited images for tail classes. Such an imbalance causes mode collapse and reduces the diversity of synthesized images for tail classes. For class-conditional diffusion models trained on imbalanced data, we aim to improve the diversity of tail class images without compromising the fidelity and diversity of head class images. We achieve this by introducing two deceptively simple but highly effective contrastive loss functions. Firstly, we employ an unsupervised InfoNCE loss utilizing negative samples to increase the distance/dissimilarity among synthetic images, particularly for tail classes. To further enhance the diversity of tail classes, our second loss is an MSE loss that contrasts class-conditional generation with unconditional generation at large timesteps. This second loss makes the denoising process insensitive to class conditions for the initial steps, which enriches tail classes through knowledge sharing from head classes. Conditional-unconditional alignment has been shown to enhance the performance of long-tailed GAN. We are the first to adapt such alignment to diffusion models. We successfully leveraged contrastive learning for class-imbalanced diffusion models. Our contrastive learning framework is easy to implement and outperforms standard DDPM and alternative methods for class-imbalanced diffusion models across various datasets, including CIFAR10/100-LT, PlacesLT, TinyImageNetLT, and ImageNetLT.

arXiv:2507.08956v1 Announce Type: new Abstract: Diffusion models have quickly become some of the most popular and powerful generative models for high-dimensional data. The key insight that enabled their development was the realization that access to the score -- the gradient of the log-density at different noise levels -- allows for sampling from data distributions by solving a reverse-time stochastic differential equation (SDE) via forward discretization, and that popular denoisers allow for unbiased estimators of this score. In this paper, we demonstrate that an alternative, backward discretization of these SDEs, using proximal maps in place of the score, leads to theoretical and practical benefits. We leverage recent results in proximal matching to learn proximal operators of the log-density and, with them, develop Proximal Diffusion Models (ProxDM). Theoretically, we prove that $\widetilde{O}(d/\sqrt{\varepsilon})$ steps suffice for the resulting discretization to generate an $\varepsilon$-accurate distribution w.r.t. the KL divergence. Empirically, we show that two variants of ProxDM achieve significantly faster convergence within just a few sampling steps compared to conventional score-matching methods.

arXiv:2507.09573v1 Announce Type: new Abstract: Artistic typography aims to stylize input characters with visual effects that are both creative and legible. Traditional approaches rely heavily on manual design, while recent generative models, particularly diffusion-based methods, have enabled automated character stylization. However, existing solutions remain limited in interactivity, lacking support for localized edits, iterative refinement, multi-character composition, and open-ended prompt interpretation. We introduce WordCraft, an interactive artistic typography system that integrates diffusion models to address these limitations. WordCraft features a training-free regional attention mechanism for precise, multi-region generation and a noise blending that supports continuous refinement without compromising visual quality. To support flexible, intent-driven generation, we incorporate a large language model to parse and structure both concrete and abstract user prompts. These components allow our framework to synthesize high-quality, stylized typography across single- and multi-character inputs across multiple languages, supporting diverse user-centered workflows. Our system significantly enhances interactivity in artistic typography synthesis, opening up creative possibilities for artists and designers.

arXiv:2507.09524v1 Announce Type: new Abstract: Recent advancements in unpaired dehazing, particularly those using GANs, show promising performance in processing real-world hazy images. However, these methods tend to face limitations due to the generator's limited transport mapping capability, which hinders the full exploitation of their effectiveness in unpaired training paradigms. To address these challenges, we propose DehazeSB, a novel unpaired dehazing framework based on the Schr"odinger Bridge. By leveraging optimal transport (OT) theory, DehazeSB directly bridges the distributions between hazy and clear images. This enables optimal transport mappings from hazy to clear images in fewer steps, thereby generating high-quality results. To ensure the consistency of structural information and details in the restored images, we introduce detail-preserving regularization, which enforces pixel-level alignment between hazy inputs and dehazed outputs. Furthermore, we propose a novel prompt learning to leverage pre-trained CLIP models in distinguishing hazy images and clear ones, by learning a haze-aware vision-language alignment. Extensive experiments on multiple real-world datasets demonstrate our method's superiority. Code: https://github.com/ywxjm/DehazeSB.

arXiv:2507.10347v1 Announce Type: new Abstract: In this paper, we design an algorithm to accelerate the diffusion process on the $SO(3)$ manifold. The inherently sequential nature of diffusion models necessitates substantial time for denoising perturbed data. To overcome this limitation, we proposed to adapt the numerical Picard iteration for the $SO(3)$ space. We demonstrate our algorithm on an existing method that employs diffusion models to address the pose ambiguity problem. Moreover, we show that this acceleration advantage occurs without any measurable degradation in task reward. The experiments reveal that our algorithm achieves a speed-up of up to 4.9$\times$, significantly reducing the latency for generating a single sample.

arXiv:2507.10085v1 Announce Type: new Abstract: Representation Fine-tuning (ReFT), a recently proposed Parameter-Efficient Fine-Tuning (PEFT) method, has attracted widespread attention for significantly improving parameter efficiency by editing representation space alone. In this work, we investigate applying ReFT to complex reasoning tasks. However, directly using the native ReFT method, which modifies fixed representations at the beginning and end of each layer, yields suboptimal performance, as these fixed-position representations have uncertain impact on the outputs. We observe that, in complex reasoning tasks, there often exist certain critical representations. These representations either integrate significant information from preceding layers or regulate subsequent layer representations. Through layer-by-layer propagation, they exert a substantial influence on the final output. Naturally, fine-tuning these critical representations has the potential to greatly enhance reasoning performance. Building upon these insights, we propose Critical Representation Fine-Tuning (CRFT), a novel method that identifies and optimizes these critical representations through information flow analysis. CRFT operates within a supervised learning framework, dynamically optimizing critical representations in a low-rank linear subspace while freezing the base model. The effectiveness and efficiency of our method are validated across eight benchmarks for arithmetic and commonsense reasoning, using LLaMA and Mistral model families. Furthermore, our method also adapts effectively to few-shot settings, boosting one-shot accuracy by 16.4%. Our work highlights the untapped potential of representation-level optimization for CoT reasoning, offering a lightweight yet powerful alternative to traditional PEFT methods.

arXiv:2507.09407v1 Announce Type: new Abstract: We introduce the framework of LLM-Stackelberg games, a class of sequential decision-making models that integrate large language models (LLMs) into strategic interactions between a leader and a follower. Departing from classical Stackelberg assumptions of complete information and rational agents, our formulation allows each agent to reason through structured prompts, generate probabilistic behaviors via LLMs, and adapt their strategies through internal cognition and belief updates. We define two equilibrium concepts: reasoning and behavioral equilibrium, which aligns an agent's internal prompt-based reasoning with observable behavior, and conjectural reasoning equilibrium, which accounts for epistemic uncertainty through parameterized models over an opponent's response. These layered constructs capture bounded rationality, asymmetric information, and meta-cognitive adaptation. We illustrate the framework through a spearphishing case study, where a sender and a recipient engage in a deception game using structured reasoning prompts. This example highlights the cognitive richness and adversarial potential of LLM-mediated interactions. Our results show that LLM-Stackelberg games provide a powerful paradigm for modeling decision-making in domains such as cybersecurity, misinformation, and recommendation systems.

arXiv:2507.10007v1 Announce Type: new Abstract: Chain of Thought (CoT) reasoning has demonstrated remarkable deep reasoning capabilities in both large language models (LLMs) and multimodal large language models (MLLMs). However, its reliability is often undermined by the accumulation of errors in intermediate steps. This paper introduces an novel approach to calibrate the CoT reasoning accuracy by leveraging the model's intrinsic veracity encoding. We discover that specific attention head activations reliably reflect the truthfulness of reasoning steps in CoT. Based on this insight, we train a confidence predictor to evaluate the correctness of each reasoning step using these truthfulness-sensitive activations, dynamically selecting the most plausible reasoning path via beam search. Experimental results demonstrate that our method significantly outperforms the state-of-the-art baselines (e.g., Few-Shot CoT, Self-Consistency, and Self-Evaluation Guided Beam Search) across the mathematical, symbolic, and commonsense reasoning tasks, exhibiting superior accuracy and reliability in both unimodal and multimodal settings. We further validate the approach on large reasoning models, confirming its applicability to specialized reasoning models. Additionally, we explore the role of the model's self-correction ability in CoT reasoning. This work provides a novel reliability improvement path for CoT reasoning with broad application potential.

arXiv:2507.09495v1 Announce Type: new Abstract: Multi-agent reinforcement learning faces fundamental challenges that conventional approaches have failed to overcome: exponentially growing joint action spaces, non-stationary environments where simultaneous learning creates moving targets, and partial observability that constrains coordination. Current methods remain reactive, employing stimulus-response mechanisms that fail when facing novel scenarios. We argue for a transformative paradigm shift from reactive to proactive multi-agent intelligence through generative AI-based reinforcement learning. This position advocates reconceptualizing agents not as isolated policy optimizers, but as sophisticated generative models capable of synthesizing complex multi-agent dynamics and making anticipatory decisions based on predictive understanding of future interactions. Rather than responding to immediate observations, generative-RL agents can model environment evolution, predict other agents' behaviors, generate coordinated action sequences, and engage in strategic reasoning accounting for long-term dynamics. This approach leverages pattern recognition and generation capabilities of generative AI to enable proactive decision-making, seamless coordination through enhanced communication, and dynamic adaptation to evolving scenarios. We envision this paradigm shift will unlock unprecedented possibilities for distributed intelligence, moving beyond individual optimization toward emergent collective behaviors representing genuine collaborative intelligence. The implications extend across autonomous systems, robotics, and human-AI collaboration, promising solutions to coordination challenges intractable under traditional reactive frameworks.

arXiv:2507.09839v1 Announce Type: new Abstract: An increasing number of NLP applications interact with large language models (LLMs) through black-box APIs, making prompt engineering critical for controlling model outputs. While recent Automatic Prompt Optimization (APO) methods iteratively refine prompts using model-generated feedback, textual gradients, they primarily focus on error correction and neglect valuable insights from correct predictions. This limits both their effectiveness and efficiency. In this paper, we propose a novel APO framework centered on enhancing the feedback mechanism. We reinterpret the textual gradient as a form of negative reinforcement and introduce the complementary positive reinforcement to explicitly preserve beneficial prompt components identified through successful predictions. To mitigate the noise inherent in LLM-generated feedback, we introduce a technique called feedback diversification, which aggregates multiple feedback signals, emphasizing consistent, actionable advice while filtering out outliers. Motivated by the rapid evolution and diversity of available LLMs, we also formalize Continual Prompt Optimization (CPO), addressing the practical challenge of efficiently migrating optimized prompts between different model versions or API providers. Our experiments reveal that naive prompt migration often degrades performance due to loss of critical instructions. In contrast, our approach consistently outperforms strong baselines, achieving significant accuracy improvements, faster convergence, and lower computational costs in both standard and migration scenarios.

arXiv:2507.09279v1 Announce Type: new Abstract: Multimodal large language models (MLLMs) hold considerable promise for applications in healthcare. However, their deployment in safety-critical settings is hindered by two key limitations: (i) sensitivity to prompt design, and (ii) a tendency to generate incorrect responses with high confidence. As clinicians may rely on a model's stated confidence to gauge the reliability of its predictions, it is especially important that when a model expresses high confidence, it is also highly accurate. We introduce Prompt4Trust, the first reinforcement learning (RL) framework for prompt augmentation targeting confidence calibration in MLLMs. A lightweight LLM is trained to produce context-aware auxiliary prompts that guide a downstream task MLLM to generate responses in which the expressed confidence more accurately reflects predictive accuracy. Unlike conventional calibration techniques, Prompt4Trust specifically prioritizes aspects of calibration most critical for safe and trustworthy clinical decision-making. Beyond improvements driven by this clinically motivated calibration objective, our proposed method also improves task accuracy, achieving state-of-the-art medical visual question answering (VQA) performance on the PMC-VQA benchmark, which is composed of multiple-choice questions spanning diverse medical imaging modalities. Moreover, our framework trained with a small downstream task MLLM showed promising zero-shot generalization to larger MLLMs in our experiments, suggesting the potential for scalable calibration without the associated computational costs. This work demonstrates the potential of automated yet human-aligned prompt engineering for improving the the trustworthiness of MLLMs in safety critical settings. Our codebase can be found at https://github.com/xingbpshen/vccrl-llm.

arXiv:2507.08838v1 Announce Type: new Abstract: Improving the reasoning capabilities of diffusion-based large language models (dLLMs) through reinforcement learning (RL) remains an open problem. The intractability of dLLMs likelihood function necessitates approximating the current, old, and reference policy likelihoods at each policy optimization step. This reliance introduces additional computational overhead and lead to potentially large bias -- particularly when approximation errors occur in the denominator of policy ratios used for importance sampling. To mitigate these issues, we introduce $\mathtt{wd1}$, a novel policy optimization approach that reformulates the objective as a weighted likelihood, requiring only a single approximation for the current parametrized policy likelihood. Experiments on widely used reasoning benchmarks demonstrate that $\mathtt{wd1}$, without supervised fine-tuning (SFT) or any supervised data, outperforms existing RL methods for dLLMs, achieving up to 16% higher accuracy. $\mathtt{wd1}$ delivers additional computational gains, including reduced training time and fewer function evaluations (NFEs) per gradient step. These findings, combined with the simplicity of method's implementation and R1-Zero-like training (no SFT), position $\mathtt{wd1}$ as a more effective and efficient method for applying RL to dLLMs reasoning.

arXiv:2507.09638v1 Announce Type: new Abstract: The Retrieval-Augmented Generation (RAG) systems' performance on Thai legal question answering is still limited, especially for questions requiring extensive, complex legal reasoning. To address these limitations, we introduce an approach aligning LLMs toward improved law citation accuracy and better response quality using Group-Relative Policy Optimization (GRPO). Our approach leverages BGE-M3 embeddings as a cost-efficient semantic-similarity reward, significantly reducing computational expenses up to 2.5x compared to large language model judges. Experiments on the NitiBench benchmark demonstrate substantial improvements: GRPO achieves up to 90% citation-F1 gains from the base model and a 31% increase in joint quality metrics over instruction tuning. Crucially, our method shows enhanced robustness on complex legal reasoning tasks compared to instruction tuning, providing an effective and resource-efficient solution for enhancing Thai legal LLMs.

arXiv:2507.10532v1 Announce Type: new Abstract: The reasoning capabilities of large language models (LLMs) have been a longstanding focus of research. Recent works have further enhanced these capabilities using reinforcement learning (RL), with many new methods claiming significant improvements with minimal or no external supervision. Surprisingly, some studies even suggest that random or incorrect reward signals can enhance reasoning performance. However, these breakthroughs are mostly reported on the Qwen2.5 model family and evaluated on well-known benchmarks such as MATH-500, AMC, and AIME, while failing to achieve similar gains on other models like Llama, which warrants further investigation. Our analysis shows that although Qwen2.5 achieves strong mathematical reasoning performance, its pretraining on large-scale web corpora makes it vulnerable to data contamination in popular benchmarks. As a result, results derived from these benchmarks may be unreliable. To address this, we introduce a generator that produces fully synthetic arithmetic problems of arbitrary length and difficulty, yielding a clean dataset we call RandomCalculation. Using these leakage-free datasets, we show that only accurate reward signals consistently improve performance, while noisy or incorrect signals do not. We advocate for evaluating RL methods on uncontaminated benchmarks and across diverse model families to ensure trustworthy conclusions.

arXiv:2507.09482v1 Announce Type: new Abstract: Human emotions are complex, with sarcasm being a subtle and distinctive form. Despite progress in sarcasm research, sarcasm generation remains underexplored, primarily due to the overreliance on textual modalities and the neglect of visual cues, as well as the mismatch between image content and sarcastic intent in existing datasets. In this paper, we introduce M2SaG, a multimodal sarcasm generation dataset with 4,970 samples, each containing an image, a sarcastic text, and a sarcasm target. To benchmark M2SaG, we propose ViSP, a generation framework that integrates Proximal Policy Optimization (PPO) and contrastive learning. PPO utilizes reward scores from DIP to steer the generation of sarcastic texts, while contrastive learning encourages the model to favor outputs with higher reward scores. These strategies improve overall generation quality and produce texts with more pronounced sarcastic intent. We evaluate ViSP across five metric sets and find it surpasses all baselines, including large language models, underscoring their limitations in sarcasm generation. Furthermore, we analyze the distributions of Sarcasm Scores and Factual Incongruity for both M2SaG and the texts generated by ViSP. The generated texts exhibit higher mean Sarcasm Scores (0.898 vs. 0.770) and Factual Incongruity (0.768 vs. 0.739), demonstrating that ViSP produces higher-quality sarcastic content than the original dataset. % The dataset and code will be publicly available. Our dataset and code will be released at \textit{https://github.com/wclapply/ViSP}.

arXiv:2507.09662v1 Announce Type: new Abstract: Large reasoning models (LRMs) like OpenAI o1 and DeepSeek R1 have demonstrated impressive performance on complex reasoning tasks like mathematics and programming with long Chain-of-Thought (CoT) reasoning sequences (slow-thinking), compared with traditional large language models (fast-thinking). However, these reasoning models also face a huge challenge that generating unnecessarily lengthy and redundant reasoning chains even for trivial questions. This phenomenon leads to a significant waste of inference resources, increases the response time for simple queries, and hinders the practical application of LRMs in real-world products. To this end, it is crucial to shorten lengthy reasoning chains and learn adaptive reasoning between fast and slow thinking based on input difficulty. In this survey, we provide a comprehensive overview of recent progress in concise and adaptive thinking for efficient reasoning of LRMs, including methodologies, benchmarks, and challenges for future exploration. We hope this survey can help researchers quickly understand the landscape of this field and inspire novel adaptive thinking ideas to facilitate better usage of LRMs.

arXiv:2507.09104v1 Announce Type: new Abstract: Recently, the role of LLM-as-judge in evaluating large language models has gained prominence. However, current judge models suffer from narrow specialization and limited robustness, undermining their capacity for comprehensive evaluations. In this work, we present CompassJudger-2, a novel generalist judge model that overcomes these limitations via a task-driven, multi-domain data curation strategy. Central to our approach is supervising judgment tasks with verifiable rewards, guiding intrinsic critical reasoning through rejection sampling to foster robust, generalizable judgment capabilities. We introduce a refined learning objective with margin policy gradient loss to enhance performance. Empirically, CompassJudger-2 achieves superior results across multiple judge and reward benchmarks, and our 7B model demonstrates competitive judgment accuracy with significantly larger models like DeepSeek-V3 and Qwen3-235B-A22B. Additionally, we propose JudgerBenchV2, a comprehensive benchmark evaluating cross-domain judgment accuracy and rank consistency to standardize judge model evaluation. These contributions advance robust, scalable LLM judgment and establish new performance and evaluation standards.

arXiv:2507.09016v1 Announce Type: new Abstract: Reinforcement Learning from Human Feedback (RLHF) aligns language models with human preferences but is computationally expensive. We explore two approaches that leverage human gaze modeling to enhance RLHF: (1) gaze-aware reward models and (2) gaze-based distribution of sparse rewards at token level. Our experiments demonstate that gaze-informed RLHF achieves faster convergence while maintaining or slightly improving performance, thus, reducing computational costs during policy optimization. These results show that human gaze provides a valuable and underused signal for policy optimization, pointing to a promising direction for improving RLHF efficiency.

arXiv:2507.09884v1 Announce Type: new Abstract: Large language models (LLMs) increasingly rely on reinforcement learning (RL) to enhance their reasoning capabilities through feedback. A critical challenge is verifying the consistency of model-generated responses and reference answers, since these responses are often lengthy, diverse, and nuanced. Rule-based verifiers struggle with complexity, prompting the use of model-based verifiers. However, specialized verifiers lack flexibility, while general LLM judges can be inconsistent. Existing research primarily focuses on building better verifiers, yet a systematic evaluation of different types of verifiers' performance across domains remains lacking, severely constraining the reliable development of Reinforcement Learning with Verifiable Reward (RLVR). To address this, we propose VerifyBench--a cross-domain comprehensive benchmark for systematically evaluating verifiers. We construct 4,000 expert-level questions covering mathematics, physics, chemistry, and biology. Each question is equipped with reference answers and diverse responses. The reliability of the evaluation is ensured through a rigorous annotation process conducted by a multidisciplinary expert team. We design a four-dimensional experimental framework to comprehensively compare the performance boundaries of specialized verifiers and general LLMs under combined conditions of extracted answers vs. complete responses, and short vs. long outputs. Our evaluation uncovers fundamental trade-offs in verifiers: while specialized verifiers achieve leading accuracy, they exhibit deficiencies in recall; general models show stronger inclusivity but unstable precision. More importantly, we discover verifiers' high sensitivity to input structure and inherent limitations in cross-domain generalization, providing critical insights into the bottlenecks of current verifier technology.

arXiv:2507.09477v1 Announce Type: new Abstract: Retrieval-Augmented Generation (RAG) lifts the factuality of Large Language Models (LLMs) by injecting external knowledge, yet it falls short on problems that demand multi-step inference; conversely, purely reasoning-oriented approaches often hallucinate or mis-ground facts. This survey synthesizes both strands under a unified reasoning-retrieval perspective. We first map how advanced reasoning optimizes each stage of RAG (Reasoning-Enhanced RAG). Then, we show how retrieved knowledge of different type supply missing premises and expand context for complex inference (RAG-Enhanced Reasoning). Finally, we spotlight emerging Synergized RAG-Reasoning frameworks, where (agentic) LLMs iteratively interleave search and reasoning to achieve state-of-the-art performance across knowledge-intensive benchmarks. We categorize methods, datasets, and open challenges, and outline research avenues toward deeper RAG-Reasoning systems that are more effective, multimodally-adaptive, trustworthy, and human-centric. The collection is available at https://github.com/DavidZWZ/Awesome-RAG-Reasoning.

arXiv:2507.09854v1 Announce Type: new Abstract: Neurosymbolic artificial intelligence (AI) systems combine neural network and classical symbolic AI mechanisms to exploit the complementary strengths of large scale, generalizable learning and robust, verifiable reasoning. Numerous classifications of neurosymbolic AI illustrate how these two components can be integrated in distinctly different ways. In this work, we propose reinterpreting instruction tuned large language models as model grounded symbolic AI systems where natural language serves as the symbolic layer and grounding is achieved through the models internal representation space. Within this framework, we investigate and develop novel learning and reasoning approaches that preserve structural similarities to traditional learning and reasoning paradigms. Preliminary evaluations across axiomatic deductive reasoning procedures of varying complexity provide insights into the effectiveness of our approach in improving learning efficiency and reasoning reliability.

arXiv:2507.10124v1 Announce Type: new Abstract: Identifying bias in LLMs is ongoing. Because they are still in development, what is true today may be false tomorrow. We therefore need general strategies for debiasing that will outlive current models. Strategies developed for debiasing human decision making offer one promising approach as they incorporate an LLM-style prompt intervention designed to bring latent knowledge into awareness during decision making. LLMs trained on vast amounts of information contain information about potential biases, counter-arguments, and contradictory evidence, but that information may only be brought to bear if prompted. Metacognitive prompts developed in the human decision making literature are designed to achieve this, and as I demonstrate here, they show promise with LLMs. The prompt I focus on here is "could you be wrong?" Following an LLM response, this prompt leads LLMs to produce additional information, including why they answered as they did, errors, biases, contradictory evidence, and alternatives, none of which were apparent in their initial response. Indeed, this metaknowledge often reveals that how LLMs and users interpret prompts are not aligned. Here I demonstrate this prompt using a set of questions taken from recent articles about LLM biases, including implicit discriminatory biases and failures of metacognition. "Could you be wrong" prompts the LLM to identify its own biases and produce cogent metacognitive reflection. I also present another example involving convincing but incomplete information, which is readily corrected by the metacognitive prompt. In sum, this work argues that human psychology offers a new avenue for prompt engineering, leveraging a long history of effective prompt-based improvements to human decision making.

arXiv:2507.09650v1 Announce Type: new Abstract: How can large language models (LLMs) serve users with varying preferences that may conflict across cultural, political, or other dimensions? To advance this challenge, this paper establishes four key results. First, we demonstrate, through a large-scale multilingual human study with representative samples from five countries (N=15,000), that humans exhibit significantly more variation in preferences than the responses of 21 state-of-the-art LLMs. Second, we show that existing methods for preference dataset collection are insufficient for learning the diversity of human preferences even along two of the most salient dimensions of variability in global values, due to the underlying homogeneity of candidate responses. Third, we argue that this motivates the need for negatively-correlated sampling when generating candidate sets, and we show that simple prompt-based techniques for doing so significantly enhance the performance of alignment methods in learning heterogeneous preferences. Fourth, based on this novel candidate sampling approach, we collect and open-source Community Alignment, the largest and most representative multilingual and multi-turn preference dataset to date, featuring almost 200,000 comparisons from annotators spanning five countries. We hope that the Community Alignment dataset will be a valuable resource for improving the effectiveness of LLMs for a diverse global population.

arXiv:2507.10548v1 Announce Type: new Abstract: Recent advanced vision-language models(VLMs) have demonstrated strong performance on passive, offline image and video understanding tasks. However, their effectiveness in embodied settings, which require online interaction and active scene understanding remains limited. In such scenarios, an agent perceives the environment from a first-person perspective, with each action dynamically shaping subsequent observations. Even state-of-the-art models such as GPT-4o, Claude 3.5 Sonnet, and Gemini 2.5 Pro struggle in open-environment interactions, exhibiting clear limitations in spatial reasoning and long-horizon planning. To address this gap, we introduce EmRACE-3K, a dataset of over 3,000 language-guided tasks situated in diverse, photorealistic environments constructed using Unreal Engine and the UnrealCV-Zoo framework. The tasks encompass a wide range of embodied challenges, including navigation, object manipulation, and multi-stage goal execution. Each task unfolds as a multi-step trajectory, pairing first-person visual observations with high-level instructions, grounded actions, and natural language rationales that express the agent's intent at every step. Using EmRACE-3K, we establish a benchmark to evaluate the embodied reasoning capabilities of VLMs across three key dimensions: Exploration, Dynamic Spatial-Semantic Reasoning, and Multi-stage Goal Execution. In zero-shot settings, all models achieve success rates below 20%, underscoring the challenge posed by our benchmark and the current limitations of VLMs in interactive environments. To demonstrate the utility of EmRACE-3K, we further fine-tune Qwen2.5-VL-7B using supervised learning followed by reinforcement learning. This approach yields substantial improvements across all three challenge categories, highlighting the dataset's effectiveness in enabling the development of embodied reasoning capabilities.

arXiv:2507.09693v1 Announce Type: new Abstract: Experiment commentary is crucial in describing the experimental procedures, delving into underlying scientific principles, and incorporating content-related safety guidelines. In practice, human teachers rely heavily on subject-specific expertise and invest significant time preparing such commentary. To address this challenge, we introduce the task of automatic commentary generation across multi-discipline scientific experiments. While recent progress in large multimodal models (LMMs) has demonstrated promising capabilities in video understanding and reasoning, their ability to generate fine-grained and insightful experiment commentary remains largely underexplored. In this paper, we make the following contributions: (i) We construct \textit{ExpInstruct}, the first dataset tailored for experiment commentary generation, featuring over 7\textit{K} step-level commentaries across 21 scientific subjects from 3 core disciplines (\ie, science, healthcare and engineering). Each sample includes procedural descriptions along with potential scientific principles (\eg, chemical equations and physical laws) and safety guidelines. (ii) We propose ExpStar, an automatic experiment commentary generation model that leverages a retrieval-augmented mechanism to adaptively access, evaluate, and utilize external knowledge. (iii) Extensive experiments show that our ExpStar substantially outperforms 14 leading LMMs, which highlights the superiority of our dataset and model. We believe that ExpStar holds great potential for advancing AI-assisted scientific experiment instruction.

arXiv:2507.09313v1 Announce Type: new Abstract: With the growing research focus on multimodal dialogue systems, the capability for proactive interaction is gradually gaining recognition. As an alternative to conventional turn-by-turn dialogue, users increasingly expect multimodal systems to be more initiative, for example, by autonomously determining the timing of multi-turn responses in real time during video playback. To facilitate progress in this emerging area, we introduce ProactiveBench, the first comprehensive benchmark to evaluate a system's ability to engage in proactive interaction. Since model responses are generated at varying timestamps, we further propose PAUC, the first metric that accounts for the temporal dynamics of model responses. This enables a more accurate evaluation of systems operating in proactive settings. Through extensive benchmarking of various baseline systems on ProactiveBench and a user study of human preferences, we show that PAUC is in better agreement with human preferences than traditional evaluation metrics, which typically only consider the textual content of responses. These findings demonstrate that PAUC provides a more faithful assessment of user experience in proactive interaction scenarios. Project homepage: https://github.com/yellow-binary-tree/ProactiveBench

arXiv:2507.09491v1 Announce Type: new Abstract: Existing video benchmarks often resemble image-based benchmarks, with question types like "What actions does the person perform throughout the video?" or "What color is the woman's dress in the video?" For these, models can often answer by scanning just a few key frames, without deep temporal reasoning. This limits our ability to assess whether large vision-language models (LVLMs) can truly think with videos rather than perform superficial frame-level analysis. To address this, we introduce GLIMPSE, a benchmark specifically designed to evaluate whether LVLMs can genuinely think with videos. Unlike prior benchmarks, GLIMPSE emphasizes comprehensive video understanding beyond static image cues. It consists of 3,269 videos and over 4,342 highly visual-centric questions across 11 categories, including Trajectory Analysis, Temporal Reasoning, and Forensics Detection. All questions are carefully crafted by human annotators and require watching the entire video and reasoning over full video context-this is what we mean by thinking with video. These questions cannot be answered by scanning selected frames or relying on text alone. In human evaluations, GLIMPSE achieves 94.82% accuracy, but current LVLMs face significant challenges. Even the best-performing model, GPT-o3, reaches only 66.43%, highlighting that LVLMs still struggle to move beyond surface-level reasoning to truly think with videos.

arXiv:2507.09815v1 Announce Type: new Abstract: Ensuring the safety of vulnerable road users (VRUs), such as pedestrians and cyclists, is a critical challenge for autonomous driving systems, as crashes involving VRUs often result in severe or fatal consequences. While multimodal large language models (MLLMs) have shown promise in enhancing scene understanding and decision making in autonomous vehicles, there is currently no standardized benchmark to quantitatively evaluate their reasoning abilities in complex, safety-critical scenarios involving VRUs. To address this gap, we present VRU-Accident, a large-scale vision-language benchmark designed to evaluate MLLMs in high-risk traffic scenarios involving VRUs. VRU-Accident comprises 1K real-world dashcam accident videos, annotated with 6K multiple-choice question-answer pairs across six safety-critical categories (with 24K candidate options and 3.4K unique answer choices), as well as 1K dense scene descriptions. Unlike prior works, our benchmark focuses explicitly on VRU-vehicle accidents, providing rich, fine-grained annotations that capture both spatial-temporal dynamics and causal semantics of accidents. To assess the current landscape of MLLMs, we conduct a comprehensive evaluation of 17 state-of-the-art models on the multiple-choice VQA task and on the dense captioning task. Our findings reveal that while MLLMs perform reasonably well on visually grounded attributes, they face significant challenges in reasoning and describing accident causes, types, and preventability.

arXiv:2507.09862v1 Announce Type: new Abstract: The rapid development of large-scale models has catalyzed significant breakthroughs in the digital human domain. These advanced methodologies offer high-fidelity solutions for avatar driving and rendering, leading academia to focus on the next major challenge: audio-visual dyadic interactive virtual human. To facilitate research in this emerging area, we present SpeakerVid-5M dataset, the first large-scale, high-quality dataset designed for audio-visual dyadic interactive virtual human generation. Totaling over 8,743 hours, SpeakerVid-5M contains more than 5.2 million video clips of human portraits. It covers diverse scales and interaction types, including monadic talking, listening, and dyadic conversations. Crucially, the dataset is structured along two key dimensions: interaction type and data quality. First, it is categorized into four types (dialogue branch, single branch, listening branch and multi-turn branch) based on the interaction scenario. Second, it is stratified into a large-scale pre-training subset and a curated, high-quality subset for Supervised Fine-Tuning (SFT). This dual structure accommodates a wide array of 2D virtual human tasks. In addition, we provide an autoregressive (AR)-based video chat baseline trained on this data, accompanied by a dedicated set of metrics and test data to serve as a benchmark VidChatBench for future work. Both the dataset and the corresponding data processing code will be publicly released. Project page: https://dorniwang.github.io/SpeakerVid-5M/

arXiv:2507.10449v1 Announce Type: new Abstract: Coral reefs are vital yet vulnerable ecosystems that require continuous monitoring to support conservation. While coral reef images provide essential information in coral monitoring, interpreting such images remains challenging due to the need for domain expertise. Visual Question Answering (VQA), powered by Large Vision-Language Models (LVLMs), has great potential in user-friendly interaction with coral reef images. However, applying VQA to coral imagery demands a dedicated dataset that addresses two key challenges: domain-specific annotations and multidimensional questions. In this work, we introduce CoralVQA, the first large-scale VQA dataset for coral reef analysis. It contains 12,805 real-world coral images from 67 coral genera collected from 3 oceans, along with 277,653 question-answer pairs that comprehensively assess ecological and health-related conditions. To construct this dataset, we develop a semi-automatic data construction pipeline in collaboration with marine biologists to ensure both scalability and professional-grade data quality. CoralVQA presents novel challenges and provides a comprehensive benchmark for studying vision-language reasoning in the context of coral reef images. By evaluating several state-of-the-art LVLMs, we reveal key limitations and opportunities. These insights form a foundation for future LVLM development, with a particular emphasis on supporting coral conservation efforts.

arXiv:2507.10541v1 Announce Type: new Abstract: Recent Large Reasoning Models (LRMs) have achieved remarkable progress on task-specific benchmarks, yet their evaluation methods remain constrained by isolated problem-solving paradigms. Existing benchmarks predominantly assess single-question reasoning through sequential testing, resulting critical limitations: (1) vulnerability to data contamination and less challenging (e.g., DeepSeek-R1 achieves 97.0% on MATH500), forcing costly and perpetual creation of new questions with large human efforts, (2) failure to evaluate models under multi-context pressure, a key requirement for real-world deployment. To bridge this gap, we present REST (Reasoning Evaluation through Simultaneous Testing), a stress-testing framework that concurrently exposes LRMs to multiple problems simultaneously. Beyond basic reasoning, REST specifically evaluates several under-tested capabilities: contextual priority allocation, cross-problem interference resistance, and dynamic cognitive load management. Our evaluation reveals several striking findings: Even state-of-the-art (SOTA) models like DeepSeek-R1 exhibit substantial performance degradation under stress testing. Crucially, REST demonstrates stronger discriminative power than existing benchmarks, revealing pronounced performance differences among models that exhibit similar, near-ceiling performance under single-question evaluations. Some key mechanistic insights emerge from our analysis: (1) the "overthinking trap" is a critical factor contributing to the performance degradation; (2) the models trained with "long2short" technique preserve more accuracy of their single-problem performance under REST, outperforming standard-trained counterparts. These results establish REST as a cost-efficient, future-proof evaluation paradigm that better reflects real-world reasoning demands while reducing reliance on continuous human annotation.

arXiv:2507.10106v1 Announce Type: new Abstract: As AI systems become increasingly capable and ubiquitous, ensuring the safety of these systems is critical. However, existing safety tools often target different aspects of model safety and cannot provide full assurance in isolation, highlighting a need for integrated and composite methodologies. This paper introduces BlueGlass, a framework designed to facilitate composite AI safety workflows by providing a unified infrastructure enabling the integration and composition of diverse safety tools that operate across model internals and outputs. Furthermore, to demonstrate the utility of this framework, we present three safety-oriented analyses on vision-language models for the task of object detection: (1) distributional evaluation, revealing performance trade-offs and potential failure modes across distributions; (2) probe-based analysis of layer dynamics highlighting shared hierarchical learning via phase transition; and (3) sparse autoencoders identifying interpretable concepts. More broadly, this work contributes foundational infrastructure and findings for building more robust and reliable AI systems.

arXiv:2507.09884v1 Announce Type: new Abstract: Large language models (LLMs) increasingly rely on reinforcement learning (RL) to enhance their reasoning capabilities through feedback. A critical challenge is verifying the consistency of model-generated responses and reference answers, since these responses are often lengthy, diverse, and nuanced. Rule-based verifiers struggle with complexity, prompting the use of model-based verifiers. However, specialized verifiers lack flexibility, while general LLM judges can be inconsistent. Existing research primarily focuses on building better verifiers, yet a systematic evaluation of different types of verifiers' performance across domains remains lacking, severely constraining the reliable development of Reinforcement Learning with Verifiable Reward (RLVR). To address this, we propose VerifyBench--a cross-domain comprehensive benchmark for systematically evaluating verifiers. We construct 4,000 expert-level questions covering mathematics, physics, chemistry, and biology. Each question is equipped with reference answers and diverse responses. The reliability of the evaluation is ensured through a rigorous annotation process conducted by a multidisciplinary expert team. We design a four-dimensional experimental framework to comprehensively compare the performance boundaries of specialized verifiers and general LLMs under combined conditions of extracted answers vs. complete responses, and short vs. long outputs. Our evaluation uncovers fundamental trade-offs in verifiers: while specialized verifiers achieve leading accuracy, they exhibit deficiencies in recall; general models show stronger inclusivity but unstable precision. More importantly, we discover verifiers' high sensitivity to input structure and inherent limitations in cross-domain generalization, providing critical insights into the bottlenecks of current verifier technology.

arXiv:2507.10442v1 Announce Type: new Abstract: Vision-language Models (VLMs) have emerged as general-purpose tools for addressing a variety of complex computer vision problems. Such models have been shown to be highly capable, but, at the same time, lacking some basic visual understanding skills. In this paper, we set out to understand the limitations of SoTA VLMs on fundamental visual tasks by constructing a series of tests that probe which components of design, specifically, may be lacking. Importantly, we go significantly beyond the current benchmarks, which simply measure the final performance of VLM response, by also comparing and contrasting it to the performance of probes trained directly on features obtained from the visual encoder, intermediate vision-language projection and LLM-decoder output. In doing so, we uncover shortcomings in VLMs and make a number of important observations about their capabilities, robustness and how they process visual information. We hope our insights will guide progress in further improving VLMs.

arXiv:2507.10548v1 Announce Type: new Abstract: Recent advanced vision-language models(VLMs) have demonstrated strong performance on passive, offline image and video understanding tasks. However, their effectiveness in embodied settings, which require online interaction and active scene understanding remains limited. In such scenarios, an agent perceives the environment from a first-person perspective, with each action dynamically shaping subsequent observations. Even state-of-the-art models such as GPT-4o, Claude 3.5 Sonnet, and Gemini 2.5 Pro struggle in open-environment interactions, exhibiting clear limitations in spatial reasoning and long-horizon planning. To address this gap, we introduce EmRACE-3K, a dataset of over 3,000 language-guided tasks situated in diverse, photorealistic environments constructed using Unreal Engine and the UnrealCV-Zoo framework. The tasks encompass a wide range of embodied challenges, including navigation, object manipulation, and multi-stage goal execution. Each task unfolds as a multi-step trajectory, pairing first-person visual observations with high-level instructions, grounded actions, and natural language rationales that express the agent's intent at every step. Using EmRACE-3K, we establish a benchmark to evaluate the embodied reasoning capabilities of VLMs across three key dimensions: Exploration, Dynamic Spatial-Semantic Reasoning, and Multi-stage Goal Execution. In zero-shot settings, all models achieve success rates below 20%, underscoring the challenge posed by our benchmark and the current limitations of VLMs in interactive environments. To demonstrate the utility of EmRACE-3K, we further fine-tune Qwen2.5-VL-7B using supervised learning followed by reinforcement learning. This approach yields substantial improvements across all three challenge categories, highlighting the dataset's effectiveness in enabling the development of embodied reasoning capabilities.

arXiv:2507.09407v1 Announce Type: new Abstract: We introduce the framework of LLM-Stackelberg games, a class of sequential decision-making models that integrate large language models (LLMs) into strategic interactions between a leader and a follower. Departing from classical Stackelberg assumptions of complete information and rational agents, our formulation allows each agent to reason through structured prompts, generate probabilistic behaviors via LLMs, and adapt their strategies through internal cognition and belief updates. We define two equilibrium concepts: reasoning and behavioral equilibrium, which aligns an agent's internal prompt-based reasoning with observable behavior, and conjectural reasoning equilibrium, which accounts for epistemic uncertainty through parameterized models over an opponent's response. These layered constructs capture bounded rationality, asymmetric information, and meta-cognitive adaptation. We illustrate the framework through a spearphishing case study, where a sender and a recipient engage in a deception game using structured reasoning prompts. This example highlights the cognitive richness and adversarial potential of LLM-mediated interactions. Our results show that LLM-Stackelberg games provide a powerful paradigm for modeling decision-making in domains such as cybersecurity, misinformation, and recommendation systems.

arXiv:2507.09495v1 Announce Type: new Abstract: Multi-agent reinforcement learning faces fundamental challenges that conventional approaches have failed to overcome: exponentially growing joint action spaces, non-stationary environments where simultaneous learning creates moving targets, and partial observability that constrains coordination. Current methods remain reactive, employing stimulus-response mechanisms that fail when facing novel scenarios. We argue for a transformative paradigm shift from reactive to proactive multi-agent intelligence through generative AI-based reinforcement learning. This position advocates reconceptualizing agents not as isolated policy optimizers, but as sophisticated generative models capable of synthesizing complex multi-agent dynamics and making anticipatory decisions based on predictive understanding of future interactions. Rather than responding to immediate observations, generative-RL agents can model environment evolution, predict other agents' behaviors, generate coordinated action sequences, and engage in strategic reasoning accounting for long-term dynamics. This approach leverages pattern recognition and generation capabilities of generative AI to enable proactive decision-making, seamless coordination through enhanced communication, and dynamic adaptation to evolving scenarios. We envision this paradigm shift will unlock unprecedented possibilities for distributed intelligence, moving beyond individual optimization toward emergent collective behaviors representing genuine collaborative intelligence. The implications extend across autonomous systems, robotics, and human-AI collaboration, promising solutions to coordination challenges intractable under traditional reactive frameworks.

arXiv:2507.09041v1 Announce Type: new Abstract: Developing autonomous agents that quickly explore an environment and adapt their behavior online is a canonical challenge in robotics and machine learning. While humans are able to achieve such fast online exploration and adaptation, often acquiring new information and skills in only a handful of interactions, existing algorithmic approaches tend to rely on random exploration and slow, gradient-based behavior updates. How can we endow autonomous agents with such capabilities on par with humans? Taking inspiration from recent progress on both in-context learning and large-scale behavioral cloning, in this work we propose behavioral exploration: training agents to internalize what it means to explore and adapt in-context over the space of expert'' behaviors. To achieve this, given access to a dataset of expert demonstrations, we train a long-context generative model to predict expert actions conditioned on a context of past observations and a measure of how exploratory'' the expert's behaviors are relative to this context. This enables the model to not only mimic the behavior of an expert, but also, by feeding its past history of interactions into its context, to select different expert behaviors than what have been previously selected, thereby allowing for fast online adaptation and targeted, ``expert-like'' exploration. We demonstrate the effectiveness of our method in both simulated locomotion and manipulation settings, as well as on real-world robotic manipulation tasks, illustrating its ability to learn adaptive, exploratory behavior.

arXiv:2507.08885v1 Announce Type: new Abstract: How to enable robots to predict the outcomes of their own motion intentions in three-dimensional space has been a fundamental problem in embodied intelligence. To explore more general spatial imagination capabilities, here we present AirScape, the first world model designed for six-degree-of-freedom aerial agents. AirScape predicts future observation sequences based on current visual inputs and motion intentions. Specifically, we construct an dataset for aerial world model training and testing, which consists of 11k video-intention pairs. This dataset includes first-person-view videos capturing diverse drone actions across a wide range of scenarios, with over 1,000 hours spent annotating the corresponding motion intentions. Then we develop a two-phase training schedule to train a foundation model -- initially devoid of embodied spatial knowledge -- into a world model that is controllable by motion intentions and adheres to physical spatio-temporal constraints.

arXiv:2507.09160v1 Announce Type: new Abstract: Vision-Language-Action (VLA) models have shown remarkable achievements, driven by the rich implicit knowledge of their vision-language components. However, achieving generalist robotic agents demands precise grounding into physical interactions, especially in contact-rich scenarios where fine-grained force control is essential. We advance VLAs' implicit knowledge beyond identifying what to do, towards guiding how to physically interact with real world. This paper introduces Tactile-VLA, a novel framework that deeply fuses vision, language, action, and tactile sensing. This framework incorporates a hybrid position-force controller to translate the model's intentions into precise physical actions and a reasoning module that allows the robot to adapt its strategy based on tactile feedback. Experiments demonstrate Tactile-VLA's effectiveness and generalizability in three key aspects: (1) enabling tactile-aware instruction following, (2) utilizing tactile-relevant commonsense, and (3) facilitating adaptive tactile-involved reasoning. A key finding is that the VLM's prior knowledge already contains semantic understanding of physical interaction; by connecting it to the robot's tactile sensors with only a few demonstrations, we can activate this prior knowledge to achieve zero-shot generalization in contact-rich tasks.

arXiv:2507.09985v1 Announce Type: new Abstract: Touch is recognized as a vital sense for humans and an equally important modality for robots, especially for dexterous manipulation, material identification, and scenarios involving visual occlusion. Building upon very recent work in touch foundation models, this demonstration will feature Octopi-1.5, our latest visual-tactile-language model. Compared to its predecessor, Octopi-1.5 introduces the ability to process tactile signals from multiple object parts and employs a simple retrieval-augmented generation (RAG) module to improve performance on tasks and potentially learn new objects on-the-fly. The system can be experienced live through a new handheld tactile-enabled interface, the TMI, equipped with GelSight and TAC-02 tactile sensors. This convenient and accessible setup allows users to interact with Octopi-1.5 without requiring a robot. During the demonstration, we will showcase Octopi-1.5 solving tactile inference tasks by leveraging tactile inputs and commonsense knowledge. For example, in a Guessing Game, Octopi-1.5 will identify objects being grasped and respond to follow-up queries about how to handle it (e.g., recommending careful handling for soft fruits). We also plan to demonstrate Octopi-1.5's RAG capabilities by teaching it new items. With live interactions, this demonstration aims to highlight both the progress and limitations of VTLMs such as Octopi-1.5 and to foster further interest in this exciting field. Code for Octopi-1.5 and design files for the TMI gripper are available at https://github.com/clear-nus/octopi-1.5.

arXiv:2507.10284v1 Announce Type: new Abstract: Visual coverage path planning with unmanned aerial vehicles (UAVs) requires agents to strategically coordinate UAV motion and camera control to maximize coverage, minimize redundancy, and maintain battery efficiency. Traditional reinforcement learning (RL) methods rely on environment-specific reward formulations that lack semantic adaptability. This study proposes Prompt-Informed Reinforcement Learning (PIRL), a novel approach that integrates the zero-shot reasoning ability and in-context learning capability of large language models with curiosity-driven RL. PIRL leverages semantic feedback from an LLM, GPT-3.5, to dynamically shape the reward function of the Proximal Policy Optimization (PPO) RL policy guiding the agent in position and camera adjustments for optimal visual coverage. The PIRL agent is trained using OpenAI Gym and evaluated in various environments. Furthermore, the sim-to-real-like ability and zero-shot generalization of the agent are tested by operating the agent in Webots simulator which introduces realistic physical dynamics. Results show that PIRL outperforms multiple learning-based baselines such as PPO with static rewards, PPO with exploratory weight initialization, imitation learning, and an LLM-only controller. Across different environments, PIRL outperforms the best-performing baseline by achieving up to 14% higher visual coverage in OpenAI Gym and 27% higher in Webots, up to 25% higher battery efficiency, and up to 18% lower redundancy, depending on the environment. The results highlight the effectiveness of LLM-guided reward shaping in complex spatial exploration tasks and suggest a promising direction for integrating natural language priors into RL for robotics.

arXiv:2507.08851v1 Announce Type: new Abstract: Understanding open-world semantics is critical for robotic planning and control, particularly in unstructured outdoor environments. Current vision-language mapping approaches rely on object-centric segmentation priors, which often fail outdoors due to semantic ambiguities and indistinct semantic class boundaries. We propose OTAS - an Open-vocabulary Token Alignment method for Outdoor Segmentation. OTAS overcomes the limitations of open-vocabulary segmentation models by extracting semantic structure directly from the output tokens of pretrained vision models. By clustering semantically similar structures across single and multiple views and grounding them in language, OTAS reconstructs a geometrically consistent feature field that supports open-vocabulary segmentation queries. Our method operates zero-shot, without scene-specific fine-tuning, and runs at up to ~17 fps. OTAS provides a minor IoU improvement over fine-tuned and open-vocabulary 2D segmentation methods on the Off-Road Freespace Detection dataset. Our model achieves up to a 151% IoU improvement over open-vocabulary mapping methods in 3D segmentation on TartanAir. Real-world reconstructions demonstrate OTAS' applicability to robotic applications. The code and ROS node will be made publicly available upon paper acceptance.

arXiv:2507.10539v1 Announce Type: new Abstract: World models (WMs) demonstrate strong capabilities in prediction, generation, and planning tasks. Existing WMs primarily focus on unstructured data and cannot leverage the ubiquitous structured data, often represented as graphs, in the digital world. While multiple graph foundation models have been proposed, they focus on graph learning tasks and cannot extend to diverse multi-modal data and interdisciplinary tasks. To address these challenges, we propose the Graph World Model (GWM), a world model that supports both unstructured and graph-structured states with multi-modal information and represents diverse tasks as actions. The core of a GWM is a generic message-passing algorithm to aggregate structured information, either over a unified multi-modal token space by converting multi-modal data into text (GWM-T) or a unified multi-modal embedding space by modality-specific encoders (GWM-E). Notably, GWM introduces action nodes to support diverse tasks, where action nodes are linked to other nodes via direct reference or similarity computation. Extensive experiments on six tasks from diverse domains, including multi-modal generation and matching, recommendation, graph prediction, multi-agent, retrieval-augmented generation, and planning and optimization, show that the same GWM outperforms or matches domain-specific baselines' performance, benefits from multi-hop structures, and demonstrates strong zero-shot/few-shot capabilities on unseen new tasks. Our code for GWM is released at https://github.com/ulab-uiuc/GWM.

arXiv:2507.10500v1 Announce Type: new Abstract: While autonomous driving technologies continue to advance, current Advanced Driver Assistance Systems (ADAS) remain limited in their ability to interpret scene context or engage with drivers through natural language. These systems typically rely on predefined logic and lack support for dialogue-based interaction, making them inflexible in dynamic environments or when adapting to driver intent. This paper presents Scene-Aware Conversational ADAS (SC-ADAS), a modular framework that integrates Generative AI components including large language models, vision-to-text interpretation, and structured function calling to enable real-time, interpretable, and adaptive driver assistance. SC-ADAS supports multi-turn dialogue grounded in visual and sensor context, allowing natural language recommendations and driver-confirmed ADAS control. Implemented in the CARLA simulator with cloud-based Generative AI, the system executes confirmed user intents as structured ADAS commands without requiring model fine-tuning. We evaluate SC-ADAS across scene-aware, conversational, and revisited multi-turn interactions, highlighting trade-offs such as increased latency from vision-based context retrieval and token growth from accumulated dialogue history. These results demonstrate the feasibility of combining conversational reasoning, scene perception, and modular ADAS control to support the next generation of intelligent driver assistance.

arXiv:2507.09329v1 Announce Type: new Abstract: LLM-based coding agents are rapidly being deployed in software development, yet their security implications remain poorly understood. These agents, while capable of accelerating software development, may inadvertently introduce insecure practices. We conducted the first systematic security evaluation of autonomous coding agents, analyzing over 12,000 actions across five state-of-the-art models (GPT-4o, GPT-4.1, Claude variants) on 93 real-world software setup tasks. Our findings reveal significant security concerns: 21% of agent trajectories contained insecure actions, with models showing substantial variation in security behavior. We developed a high-precision detection system that identified four major vulnerability categories, with information exposure (CWE-200) being the most prevalent one. We also evaluated mitigation strategies including feedback mechanisms and security reminders with various effectiveness between models. GPT-4.1 demonstrated exceptional security awareness with 96.8% mitigation success. Our work provides the first comprehensive framework for evaluating coding agent security and highlights the need for security-aware design of next generation LLM-based coding agents.

arXiv:2507.09477v1 Announce Type: new Abstract: Retrieval-Augmented Generation (RAG) lifts the factuality of Large Language Models (LLMs) by injecting external knowledge, yet it falls short on problems that demand multi-step inference; conversely, purely reasoning-oriented approaches often hallucinate or mis-ground facts. This survey synthesizes both strands under a unified reasoning-retrieval perspective. We first map how advanced reasoning optimizes each stage of RAG (Reasoning-Enhanced RAG). Then, we show how retrieved knowledge of different type supply missing premises and expand context for complex inference (RAG-Enhanced Reasoning). Finally, we spotlight emerging Synergized RAG-Reasoning frameworks, where (agentic) LLMs iteratively interleave search and reasoning to achieve state-of-the-art performance across knowledge-intensive benchmarks. We categorize methods, datasets, and open challenges, and outline research avenues toward deeper RAG-Reasoning systems that are more effective, multimodally-adaptive, trustworthy, and human-centric. The collection is available at https://github.com/DavidZWZ/Awesome-RAG-Reasoning.

arXiv:2507.08831v1 Announce Type: new Abstract: Vision-Language Navigation in Continuous Environments (VLNCE), where an agent follows instructions and moves freely to reach a destination, is a key research problem in embodied AI. However, most navigation policies are sensitive to viewpoint changes, i.e., variations in camera height and viewing angle that alter the agent's observation. In this paper, we introduce a generalized scenario, V2-VLNCE (VLNCE with Varied Viewpoints), and propose VIL (View Invariant Learning), a view-invariant post-training strategy that enhances the robustness of existing navigation policies to changes in camera viewpoint. VIL employs a contrastive learning framework to learn sparse and view-invariant features. Additionally, we introduce a teacher-student framework for the Waypoint Predictor Module, a core component of most VLNCE baselines, where a view-dependent teacher model distills knowledge into a view-invariant student model. We employ an end-to-end training paradigm to jointly optimize these components, thus eliminating the cost for individual module training. Empirical results show that our method outperforms state-of-the-art approaches on V2-VLNCE by 8-15% measured on Success Rate for two standard benchmark datasets R2R-CE and RxR-CE. Furthermore, we evaluate VIL under the standard VLNCE setting and find that, despite being trained for varied viewpoints, it often still improves performance. On the more challenging RxR-CE dataset, our method also achieved state-of-the-art performance across all metrics when compared to other map-free methods. This suggests that adding VIL does not diminish the standard viewpoint performance and can serve as a plug-and-play post-training method.

arXiv:2507.10087v1 Announce Type: new Abstract: The rapid emergence of foundation models, particularly Large Language Models (LLMs) and Vision-Language Models (VLMs), has introduced a transformative paradigm in robotics. These models offer powerful capabilities in semantic understanding, high-level reasoning, and cross-modal generalization, enabling significant advances in perception, planning, control, and human-robot interaction. This critical review provides a structured synthesis of recent developments, categorizing applications across simulation-driven design, open-world execution, sim-to-real transfer, and adaptable robotics. Unlike existing surveys that emphasize isolated capabilities, this work highlights integrated, system-level strategies and evaluates their practical feasibility in real-world environments. Key enabling trends such as procedural scene generation, policy generalization, and multimodal reasoning are discussed alongside core bottlenecks, including limited embodiment, lack of multimodal data, safety risks, and computational constraints. Through this lens, this paper identifies both the architectural strengths and critical limitations of foundation model-based robotics, highlighting open challenges in real-time operation, grounding, resilience, and trust. The review concludes with a roadmap for future research aimed at bridging semantic reasoning and physical intelligence through more robust, interpretable, and embodied models.

arXiv:2507.09174v1 Announce Type: new Abstract: The rapid proliferation of multimodal misinformation presents significant challenges for automated fact-checking systems, especially when claims are ambiguous or lack sufficient context. We introduce RAMA, a novel retrieval-augmented multi-agent framework designed for verifying multimedia misinformation. RAMA incorporates three core innovations: (1) strategic query formulation that transforms multimodal claims into precise web search queries; (2) cross-verification evidence aggregation from diverse, authoritative sources; and (3) a multi-agent ensemble architecture that leverages the complementary strengths of multiple multimodal large language models and prompt variants. Extensive experiments demonstrate that RAMA achieves superior performance on benchmark datasets, particularly excelling in resolving ambiguous or improbable claims by grounding verification in retrieved factual evidence. Our findings underscore the necessity of integrating web-based evidence and multi-agent reasoning for trustworthy multimedia verification, paving the way for more reliable and scalable fact-checking solutions. RAMA will be publicly available at https://github.com/kalendsyang/RAMA.git.

arXiv:2507.10134v1 Announce Type: new Abstract: Unmanned Aerial Vehicles (UAVs) are vital for public safety, particularly in wildfire monitoring, where early detection minimizes environmental impact. In UAV-Assisted Wildfire Monitoring (UAWM) systems, joint optimization of sensor transmission scheduling and velocity is critical for minimizing Age of Information (AoI) from stale sensor data. Deep Reinforcement Learning (DRL) has been used for such optimization; however, its limitations such as low sampling efficiency, simulation-to-reality gaps, and complex training render it unsuitable for time-critical applications like wildfire monitoring. This paper introduces a new online Flight Resource Allocation scheme based on LLM-Enabled In-Context Learning (FRSICL) to jointly optimize the UAV's flight control and data collection schedule along the trajectory in real time, thereby asymptotically minimizing the average AoI across ground sensors. In contrast to DRL, FRSICL generates data collection schedules and controls velocity using natural language task descriptions and feedback from the environment, enabling dynamic decision-making without extensive retraining. Simulation results confirm the effectiveness of the proposed FRSICL compared to Proximal Policy Optimization (PPO) and Nearest-Neighbor baselines.