论文检索

Guangjing Yang, Ziyuan Qin, Chaoran Zhang, Chenlin Du, Jinglin Wang, Wanran Sun, Zhenyu Zhang, Bing Ji, Qicheng Lao

Medical visual grounding serves as a crucial foundation for fine-grained multimodal reasoning and interpretable clinical decision support. Despite recent advances in reinforcement learning (RL) for grounding tasks, existing approaches such as Group Relative Policy Optimization (GRPO) suffer from severe reward sparsity when directly applied to medical images, primarily due to the inherent difficulty of localizing small or ambiguous regions of interest, which is further exacerbated by the rigid and suboptimal nature of fixed IoU-based reward schemes in RL. This leads to vanishing policy gradients and stagnated optimization, particularly during early training. To address this challenge, we propose MedLoc-R1, a performance-aware reward scheduling framework that progressively tightens the reward criterion in accordance with model readiness. MedLoc-R1 introduces a sliding-window performance tracker and a multi-condition update rule that automatically adjust the reward schedule from dense, easily obtainable signals to stricter, fine-grained localization requirements, while preserving the favorable properties of GRPO without introducing auxiliary networks or additional gradient paths. Experiments on three medical visual grounding benchmarks demonstrate that MedLoc-R1 consistently improves both localization accuracy and training stability over GRPO-based baselines. Our framework offers a general, lightweight, and effective solution for RL-based grounding in high-stakes medical applications. Code and checkpoints will be released after acceptance.

Di Zhang, Zhangpeng Gong, Xiaobo Pang, Jiashuai Liu, Junbo Lu, Hao Cui, Jiusong Ge, Zhi Zeng, Kai Yi, Yinghua Li 等

Foundation models have achieved success in computational pathology, demonstrating generalization across histopathology tasks. However, existing models overlook the heterogeneous and non-uniform organization of regions of interest (ROIs) because they rely on natural image backbones not tailored for tissue morphology. Consequently, they fail to capture the coherent tissue architecture beyond patches, limiting interpretability and clinical relevance. To address these challenges, we present Cross-modal Adaptive Region Encoder (CARE), a foundation model for pathology that partitions WSIs into several morphologically relevant regions. Specifically, CARE employs a two-stage pretraining strategy: (1) a self-supervised unimodal pretraining stage that learns morphological representations from 34,277 whole-slide images (WSIs) without segmentation annotations, and (2) a cross-modal alignment stage that leverages RNA and protein profiles to refine the construction and representation of adaptive regions. This molecular guidance enables CARE to identify biologically relevant patterns and generate irregular yet coherent tissue regions, selecting the representative area as ROI. CARE supports a broad range of pathology-related tasks, using either the ROI feature or the slide-level feature obtained by aggregating adaptive regions. Based on only one-tenth of the pretraining data typically used by mainstream foundation models, CARE achieves superior average performance across 33 downstream benchmarks, including morphological classification, molecular prediction, and survival analysis, and outperforms other foundation model baselines overall.

Ao Cheng, Xingming Li, Xuanyu Ji, Xixiang He, Qiyao Sun, Chunping Qiu, Runke Huang, Qingyong Hu

Electronic Navigational Charts (ENCs) are the safety-critical backbone of modern maritime navigation, yet it remains unclear whether multimodal large language models (MLLMs) can reliably interpret them. Unlike natural images or conventional charts, ENCs encode regulations, bathymetry, and route constraints via standardized vector symbols, scale-dependent rendering, and precise geometric structure---requiring specialized maritime expertise for interpretation. We introduce ENC-Bench, the first benchmark dedicated to professional ENC understanding. ENC-Bench contains 20,490 expert-validated samples from 840 authentic National Oceanic and Atmospheric Administration (NOAA) ENCs, organized into a three-level hierarchy: Perception (symbol and feature recognition), Spatial Reasoning (coordinate localization, bearing, distance), and Maritime Decision-Making (route legality, safety assessment, emergency planning under multiple constraints). All samples are generated from raw S-57 data through a calibrated vector-to-image pipeline with automated consistency checks and expert review. We evaluate 10 state-of-the-art MLLMs such as GPT-4o, Gemini 2.5, Qwen3-VL, InternVL-3, and GLM-4.5V, under a unified zero-shot protocol. The best model achieves only 47.88% accuracy, with systematic challenges in symbolic grounding, spatial computation, multi-constraint reasoning, and robustness to lighting and scale variations. By establishing the first rigorous ENC benchmark, we open a new research frontier at the intersection of specialized symbolic reasoning and safety-critical AI, providing essential infrastructure for advancing MLLMs toward professional maritime applications.

Renbo Tu, Ali SaraerToosi, Nicholas S. Conroy, Gennady Pekhimenko, Aviad Levis

The Event Horizon Telescope (EHT) delivered the first image of a black hole by capturing the light from its surrounding accretion flow, revealing structure but not dynamics. Simulations of black hole accretion dynamics are essential for interpreting EHT images but costly to generate and impractical for inference. Motivated by this bottleneck, BHCast presents a framework for forecasting black hole plasma dynamics from a single, blurry snapshot, such as those captured by the EHT. At its core, BHCast is a neural model that transforms a static image into forecasted future frames, revealing the underlying dynamics hidden within one snapshot. With a multi-scale pyramid loss, we demonstrate how autoregressive forecasting can simultaneously super-resolve and evolve a blurry frame into a coherent, high-resolution movie that remains stable over long time horizons. From forecasted dynamics, we can then extract interpretable spatio-temporal features, such as pattern speed (rotation rate) and pitch angle. Finally, BHCast uses gradient-boosting trees to recover black hole properties from these plasma features, including the spin and viewing inclination angle. The separation between forecasting and inference provides modular flexibility, interpretability, and robust uncertainty quantification. We demonstrate the effectiveness of BHCast on simulations of two distinct black hole accretion systems, Sagittarius A* and M87*, by testing on simulated frames blurred to EHT resolution and real EHT images of M87*. Ultimately, our methodology establishes a scalable paradigm for solving inverse problems, demonstrating the potential of learned dynamics to unlock insights from resolution-limited scientific data.

Ruiying Liu, Yuanzhi Liang, Haibin Huang, Tianshu Yu, Chi Zhang

Group Relative Policy Optimization (GRPO) has emerged as an effective and lightweight framework for post-training visual generative models. However, its performance is fundamentally limited by the ambiguity of textual-visual correspondence: a single prompt may validly describe diverse visual outputs, and a single image or video may support multiple equally correct interpretations. This many-to-many relationship leads reward models to generate uncertain and weakly discriminative signals, causing GRPO to underutilize reliable feedback and overfit noisy ones. We introduce Bayesian Prior-Guided Optimization (BPGO), a novel extension of GRPO that explicitly models reward uncertainty through a semantic prior anchor. BPGO adaptively modulates optimization trust at two levels: inter-group Bayesian trust allocation emphasizes updates from groups consistent with the prior while down-weighting ambiguous ones, and intra-group prior-anchored renormalization sharpens sample distinctions by expanding confident deviations and compressing uncertain scores.Across both image and video generation tasks, BPGO delivers consistently stronger semantic alignment, enhanced perceptual fidelity, and faster convergence than standard GRPO and recent variants.

Lingxiao Li, Dongwon Kim, Lingyan Ruan, Bin Chen, Taesoo Kwon, Taehyun Rhee

Text-guided motion generation in 3D scenes has advanced the synthesis of human-scene interactions, contributing to embodied AI, scene understanding, and virtual agent simulation. While recent studies have begun exploring multi-agent scenarios, achieving temporally synchronised interactions among multiple agents remains an open challenge. Existing methods are often limited in flexibility and scalability when handling diverse interaction contexts.We present a method that enables synchronised multi-agent interaction using a single-agent motion synthesis model through two key components: a text-guided dependency-aware story planner and a temporal synchronisation module. The story planner interprets natural language instructions into structured event sequences with temporal dependencies. Our synchronisation module, built upon time-warping control and diffusion posterior sampling, aligns interaction timing across agents without retraining.Experimental results demonstrate that the proposed framework effectively models temporal dependencies and causal order between events. Evaluations across diverse interaction types show improved temporal alignment and coherent multi-agent motion generation consistent with textual instructions.

Yufei Zhan, Ziheng Wu, Yousong Zhu, Rongkun Xue, Guanghao Zhou, Ruipu Luo, Zhenghao Chen, Can Zhang, Yifan Li, Zhentao He 等

Despite recent advances in multimodal reasoning, Multimodal Large Language Models (MLLMs) still struggle on complex tasks where initial visual perceptions can be misleading. This performance gap stems from a critical reasoning flaw we term Visual Inertia: while MLLMs excel at iterative reflection in textual contexts, they tend to uncritically commit to their initial visual interpretations and rarely revise them. To overcome this limitation, we introduce GThinker, an MLLM equipped with a novel adaptive visual rethinking capability. GThinker leverages Cue-Rethinking, a flexible reasoning pattern that not only grounds reasoning in visual cues but also strategically triggers a re-examination of these cues to resolve inconsistencies. To instill this capability, we introduce a novel two-stage training framework. It begins with a pattern-guided cold start, enhanced by a judge-guided selective mechanism to learn from failure cases, followed by incentive reinforcement learning. We further curate the GThinker-11K dataset to power the training with an iterative multimodal annotation pipeline. Extensive experiments demonstrate that GThinker significantly mitigates visual inertia during reasoning, achieving a leading 81.5% on the M3CoT benchmark, which is rich in such challenges, surpassing the powerful o4-mini model. Furthermore, GThinker shows consistent improvements across a range of multimodal reasoning benchmarks with an average gain of 2.1%, showcasing the broad benefits of equipping MLLMs with the ability to rethink both what they see and how they think. Data, model, and code are available at https://github.com/jefferyZhan/GThinker.

Zhuofan Xie, Zishan Lin, Jinliang Lin, Jie Qi, Shaohua Hong, Shuo Li

Active Learning (AL) reduces annotation costs in medical imaging by selecting only the most informative samples for labeling, but suffers from cold-start when labeled data are scarce. Vision-Language Models (VLMs) address the cold-start problem via zero-shot predictions, yet their temperature-scaled softmax outputs treat text-image similarities as deterministic scores while ignoring inherent uncertainty, leading to overconfidence. This overconfidence misleads sample selection, wasting annotation budgets on uninformative cases. To overcome these limitations, the Similarity-as-Evidence (SaE) framework calibrates text-image similarities by introducing a Similarity Evidence Head (SEH), which reinterprets the similarity vector as evidence and parameterizes a Dirichlet distribution over labels. In contrast to a standard softmax that enforces confident predictions even under weak signals, the Dirichlet formulation explicitly quantifies lack of evidence (vacuity) and conflicting evidence (dissonance), thereby mitigating overconfidence caused by rigid softmax normalization. Building on this, SaE employs a dual-factor acquisition strategy: high-vacuity samples (e.g., rare diseases) are prioritized in early rounds to ensure coverage, while high-dissonance samples (e.g., ambiguous diagnoses) are prioritized later to refine boundaries, providing clinically interpretable selection rationales. Experiments on ten public medical imaging datasets with a 20% label budget show that SaE attains state-of-the-art macro-averaged accuracy of 82.57%. On the representative BTMRI dataset, SaE also achieves superior calibration, with a negative log-likelihood (NLL) of 0.425.

Zhuoli Zhuang, Yu-Cheng Chang, Yu-Kai Wang, Thomas Do, Chin-Teng Lin

Recent advancements in computer vision have accelerated the development of autonomous driving. Despite these advancements, training machines to drive in a way that aligns with human expectations remains a significant challenge. Human factors are still essential, as humans possess a sophisticated cognitive system capable of rapidly interpreting scene information and making accurate decisions. Aligning machine with human intent has been explored with Reinforcement Learning with Human Feedback (RLHF). Conventional RLHF methods rely on collecting human preference data by manually ranking AI-generated outputs, which is time-consuming and indirect. In this work, we propose an electroencephalography (EEG)-guided decision-making framework to incorporate human cognitive insights into reinforcement learning (RL) for autonomous driving. We collected EEG signals from 20 participants in a realistic driving simulator and analyzed event-related potentials (ERP) in response to sudden environmental changes. Our proposed framework employs a neural network to predict the strength of ERP based on the cognitive information from visual scene information. Moreover, we explore the integration of such cognitive information into the reward signal of the RL algorithm. Experimental results show that our framework can improve the collision avoidance ability of the RL algorithm, highlighting the potential of neuro-cognitive feedback in enhancing autonomous driving systems.

Van-Nguyen Pham, Duc-Tai Le, Junghyun Bum, Hyunseung Choo

Deep neural networks have achieved strong performance in fundus image classification, yet their flattened feature representations are often highly redundant. Such redundancy can lead to poor generalization across imaging devices, reduced interpretability, and inefficient use of model capacity. To address this issue, this study proposes a post-training feature pruning framework, termed greedy feature pruning (GFP) , which removes weak or redundant dimensions from the flattened features of trained backbones. GFP employs a greedy build-up process guided by performance metrics on the training set, constrained by a minimum feature keeping ratio, to identify compact yet discriminative subsets of features. Experiments are conducted on five public fundus datasets covering multiple tasks, including diabetic retinopathy detection (DDR, Messidor-2), glaucoma detection (PAPILA), multi-label classification (ODIR) and multi-class retinal disease classification (RETINA), using EfficientNetV2, ViT, and CoAtNet as backbones. Results show that GFP consistently improves AUROC and AUPRC across datasets while reducing the number of flattened features by up to 96%. Feature visualizations and quantitative analyses confirm that GFP enhances the compactness and separability of latent features. Moreover, cross-dataset evaluation demonstrates that GFP improves transferability between datasets, indicating better domain robustness. Overall, the proposed GFP framework provides a simple yet effective approach for compressing feature representations and improving both discriminability and generalization in fundus image classification.

Peirong Zhang, Yidan Zhang, Luxiao Xu, Jinliang Lin, Zonghao Guo, Fengxiang Wang, Xue Yang, Kaiwen Wei, Lei Wang

Recent advances in multimodal large language models (MLLMs) have led to remarkable progress in visual grounding, enabling fine-grained cross-modal alignment between textual queries and image regions. However, transferring such capabilities to remote sensing imagery remains challenging, as targets are often extremely small within kilometer-scale scenes, and queries typically involve intricate geospatial relations such as relative positions, spatial hierarchies, or contextual dependencies across distant objects.To address these challenges, we propose GeoViS, a Geospatially Rewarded Visual Search framework that reformulates remote sensing visual grounding as a progressive search-and-reasoning process. Rather than directly predicting the target location in a single step, GeoViS actively explores the global image through a tree-structured sequence of visual cues, integrating multimodal perception, spatial reasoning, and reward-guided exploration to refine geospatial hypotheses iteratively. This design enables the model to detect subtle small-scale targets while maintaining holistic scene awareness.Extensive experiments on five remote sensing grounding benchmarks demonstrate that GeoViS achieves precise geospatial understanding and consistently surpasses existing methods across key visual grounding metrics, highlighting its strong cross-domain generalization and interpretability.

Khushboo Mishra, Varun Trivedi, Tanima Dutta

Unsupervised domain adaptation (UDA) for videos and 1D time-series data faces significant challenges due to domain shifts in terms of both temporal dynamics and feature distributions. Existing UDA approaches for time-series data often address temporal alignment and uncertainty mitigation as separate objectives, leading to unstable training, noisy pseudo-labels, and incomplete feature transfer. This disjoint treatment fails to capture inter-channel causal dependencies and also overlooks the impact of prediction uncertainty on adaptation quality. This limits the transferability of learned representations and results in suboptimal adaptation. To address the aforementioned limitations, we propose a novel UDA framework, named Causally-Regularized Optimal Transport (in short Causal-OT), that preserves domain-invariant causal mechanisms by embedding causal graph regularization into robust OT alignment process. First we estimate inter-channel causal graphs in both source and target domains and learn a transport plan that not only aligns feature distributions but also improves interpretability and minimizes the discrepancy between causal structures of the Granger graphs. However, pseudo-labeling may still prone to error propagation allowing incorrect target predictions during self-training, degrading the model stability and transfer quality across domains. To mitigate this, we further introduce a causality-aware pseudo-labeling strategy that selects high-confidence target samples based on both entropy and structural consistency with the causal graph of the source domain. This enhances robustness against pseudo-label noise.Extensive experiments on six time-series benchmarks achieving 4.5% gain in accuracy and a 3.8% improvement in F1-score. We conduct experiments on four benchmark video datasets that achieve a 2.5% gain in accuracy.

Seongrae Noh, SeungWon Seo, Gyeong-Moon Park, HyeongYeop Kang

Editing a 3D indoor scene from natural language is conceptually straightforward but technically challenging. Existing open-vocabulary systems often regenerate large portions of a scene or rely on image-space edits that disrupt spatial structure, resulting in unintended global changes or physically inconsistent layouts. These limitations stem from treating editing primarily as a generative task.We take a different view. A user instruction defines a desired world state, and editing should be the minimal sequence of actions that makes this state true while preserving everything else. This perspective motivates Edit-As-Act, a framework that performs open-vocabulary scene editing as goal-regressive planning in 3D space.Given a source scene and free-form instruction, Edit-As-Act predicts symbolic goal predicates and plans in EditLang, a PDDL-inspired action language that we design with explicit preconditions and effects encoding support, contact, collision, and other geometric relations. A language-driven planner proposes actions, and a validator enforces goal-directedness, monotonicity, and physical feasibility, producing interpretable and physically coherent transformations.By separating reasoning from low-level generation, Edit-As-Act achieves instruction fidelity, semantic consistency, and physical plausibility--three criteria that existing paradigms cannot satisfy together. On E2A-Bench, our benchmark of 63 editing tasks across 9 indoor environments, Edit-As-Act significantly outperforms prior approaches across all edit types and scene categories.

Sitong Gong, Yunzhi Zhuge, Lu Zhang, Jiazuo Yu, Pingping Zhang, Xu Jia, Huchuan Lu

Video reasoning segmentation (VRS) endeavors to delineate referred objects in videos guided by implicit instructions that encapsulate human intent and temporal logic. Previous approaches leverage large vision language models (LVLMs) to encode object semantics into \SEG tokens for mask prediction. However, this paradigm suffers from limited interpretability during inference and suboptimal performance due to inadequate spatiotemporal reasoning. Drawing inspiration from seminal breakthroughs in reinforcement learning, we introduce Veason-R1, a specialized LVLM for VRS that emphasizes structured reasoning in segmentation. Veason-R1 is trained through Group Relative Policy Optimization (GRPO) augmented with Chain-of-Thought (CoT) initialization. To begin with, we curate high-quality CoT training data to instill structured reasoning trajectories, bridging video-level semantics and frame-level spatial grounding, yielding the supervised fine-tuned model Veason-SFT. Subsequently, GRPO fine-tuning encourages efficient exploration of the reasoning space by optimizing reasoning chains. To this end, we incorporate a holistic reward mechanism that synergistically enhances spatial alignment and temporal consistency, bolstering keyframe localization and fine-grained grounding. Comprehensive empirical evaluations demonstrate that Veason-R1 achieves state-of-the-art performance on multiple benchmarks, surpassing prior art by significant margins (e.g., +1.3 \mathcal J &\mathcal F in ReVOS and +10.0 \mathcal J &\mathcal F in ReasonVOS), while exhibiting robustness to hallucinations (+8.8 \mathcal R ).

Siqi Liu, Xinyang Li, Bochao Zou, Junbao Zhuo, Huimin Ma, Jiansheng Chen

As large language models (LLMs) continue to advance, there is increasing interest in their ability to infer human mental states and demonstrate a human-like Theory of Mind (ToM). Most existing ToM evaluations, however, are centered on text-based inputs, while scenarios relying solely on visual information receive far less attention. This leaves a gap, since real-world human-AI interaction typically requires multimodal understanding. In addition, many current methods regard the model as a black box and rarely probe how its internal attention behaves in multiple-choice question answering (QA). The impact of LLM hallucinations on such tasks is also underexplored from an interpretability perspective. To address these issues, we introduce VisionToM, a vision-oriented intervention framework designed to strengthen task-aware reasoning. The core idea is to compute intervention vectors that align visual representations with the correct semantic targets, thereby steering the model's attention through different layers of visual features. This guidance reduces the model's reliance on spurious linguistic priors, leading to more reliable multimodal language model (MLLM) outputs and better QA performance. Experiments on the EgoToM benchmark--an egocentric, real-world video dataset for ToM with three multiple-choice QA settings--demonstrate that our method substantially improves the ToM abilities of MLLMs. Furthermore, results on an additional open-ended generation task show that VisionToM enables MLLMs to produce free-form explanations that more accurately capture agents' mental states, pushing machine-human collaboration toward greater alignment.

Anqi Zhang, Xiaokang Ji, Guangyu Gao, Jianbo Jiao, Chi Harold Liu, Yunchao Wei

Recent segmentation methods leveraging Multi-modal Large Language Models (MLLMs) have shown reliable object-level segmentation and enhanced spatial perception. However, almost all previous methods predominantly rely on specialist mask decoders to interpret masks from generated segmentation-related embeddings and visual features, or incorporate multiple additional tokens to assist. This paper aims to investigate whether and how we can unlock segmentation from MLLM itSELF with 1 segmentation Embedding (SELF1E) while achieving competitive results, which eliminates the need for external decoders. To this end, our approach targets the fundamental limitation of resolution reduction in pixel-shuffled image features from MLLMs. First, we retain image features at their original uncompressed resolution, and refill them with residual features extracted from MLLM-processed compressed features, thereby improving feature precision. Subsequently, we integrate pixel-unshuffle operations on image features with and without LLM processing, respectively, to unleash the details of compressed features and amplify the residual features under uncompressed resolution, which further enhances the resolution of refilled features. Moreover, we redesign the attention mask with dual perception pathways, i.e., image-to-image and image-to-segmentation, enabling rich feature interaction between pixels and the segmentation token. Comprehensive experiments across multiple segmentation tasks validate that SELF1E achieves performance competitive with specialist mask decoder-based methods, demonstrating the feasibility of decoder-free segmentation in MLLMs. Project page: https://github.com/ANDYZAQ/SELF1E.

Ruichao Yang, Wei Gao, Xiaobin Zhu, Jing Ma, Hongzhan Lin, Ziyang Luo, Bo-Wen Zhang, Xu-Cheng Yin

Multimodal misinformation poses an escalating challenge that often evades traditional detectors, which are opaque black boxes and fragile against new manipulation tactics. We present Probabilistic Concept Graph Reasoning (PCGR), an interpretable, modular, and evolvable framework that reframes multimodal misinformation detection (MMD) as structured, concept-based reasoning. PCGR follows a build-then-infer paradigm, which first constructs a graph of human-understandable concept nodes, including novel high-level concepts automatically discovered and validated by multimodal large language models (MLLMs), and then applies hierarchical attention over this concept graph to infer claim veracity. This design produces interpretable reasoning chains linking evidence to conclusions. Experiments demonstrate that PCGR achieves state-of-the-art MMD accuracy and robustness to emerging manipulation types, outperforming prior methods in both coarse detection and fine-grained manipulation recognition.

Zhenbin Wang, Lei Zhang, Lituan Wang, Zhenwei Zhang, Guangwu Qian, Yan Wang, Wei Huang

This work tackles a key challenge in test-time energy adaptation: prohibitive time overhead arising from recent state-of-the-art test-time adaptation (TTA) methods, which are built on energy models relying on iterative Monte Carlo or Langevin dynamics sampling with multiple stochastic updates per test instance to approximate energy gradients. We tackle the problem from an innovative control system perspective by i) describing the energy as a complex-valued wave, where the amplitude encodes energy uncertainty and the phase characterizes its evolution, and ii) maintaining a time-dependent wave equation that interprets TTA as a control system evolution process. By enforcing the control system law of probability current conservation, our method directs probability current away from high-energy (error-prone) regions toward low-energy (accurate) ones, achieving adaptive energy redistribution without additional stochastic sampling while preserving the overall normalization of the energy landscape. Experimentally, the proposed method significantly outperforms baseline methods across several public benchmark datasets, with adaptive time being only 1/3 1/7 of that required by the compared Top-1 to Top-3 baselines. Code is available at https://github.com/wongzbb/APT.

Dongsheng Wang, Jinsen Zhang, Dawei Su, Hui Huang

Recently, sparse autoencoders (SAEs) have emerged as a promising technique for interpreting activations in foundation models by disentangling features into a sparse set of concepts. However, identifying the optimal level of sparsity for each neuron remains challenging in practice: excessive sparsity might lead to poor reconstruction, whereas insufficient sparsity harms interpretability. While existing activation functions such as ReLU and TopK provide certain sparsity guarantees, they typically require additional sparsity regularization or cherry-picked hyperparameters. We show in this paper that adaptive sparse attention mechanisms using sparsemax can bridge this trade-off, due to their ability to determine the number of concepts in a data-dependent manner.Specifically, we first explore a new class of SAEs based on the cross-attention architecture with the latent features as queries and the learnable dictionary as the key and value matrices. To encourage sparse pattern learning, we employ a sparsemax-based attention strategy that automatically infers a sparse set of concepts according to the complexity of each neuron, resulting in a more flexible and efficient activation function. Through comprehensive evaluation and visualization, we show that our approach successfully achieves lower reconstruction loss while producing high-quality concepts. Moreover, the sparsity level automatically determined by our approach can serve as tuning guidance to improve existing SAEs. The code is available at https://github.com/qyj-bkjx/Sparsemax-SAE.

Yuan Wang, Borui Liao, Huijuan Huang, Jinda Lu, Ouxiang Li, Kuien Liu, Meng Wang, Xiang Wang

Recent advances in video reward models and post-training strategies have improved text-to-video (T2V) generation. While these models typically assess visual quality, motion quality, and text alignment, they often overlook key structural distortions, such as abnormal object appearances and interactions, which can degrade the overall quality of the generative video.To address this gap, we introduce REACT, a frame-level reward model designed specifically for structural distortions evaluation in generative videos. REACT assigns point-wise scores and attribution labels by reasoning over video frames, focusing on recognizing distortions. To support this, we construct a large-scale human preference dataset, annotated based on our proposed taxonomy of structural distortions, and generate additional data using a efficient Chain-of-Thought (CoT) synthesis pipeline. REACT is trained with a two-stage framework: (1) supervised fine-tuning with masked loss for domain knowledge injection, followed by (2) reinforcement learning with Group Relative Policy Optimization (GRPO) and pairwise rewards to enhance reasoning capability and align output scores with human preferences. During inference, a dynamic sampling mechanism is introduced to focus on frames most likely to exhibit distortion.We also present REACT-Bench, a benchmark for generative video distortion evaluation. Experimental results demonstrate that REACT complements existing reward models in assessing structural distortion, achieving both accurate quantitative evaluations and interpretable attribution analysis.