论文检索

Ahyoung Oh, Wonseok Shin, Songkuk Kim

Sparse Autoencoders (SAEs) have demonstrated significant success in interpreting Large Language Models (LLMs) by decomposing dense representations into sparse, semantic components. However, their potential for analyzing Vision Transformers (ViTs) remains largely under-explored. In this work, we present the first application of SAEs to the ViT [CLS] token for out-of-distribution (OOD) detection, addressing the limitation of existing methods that rely on entangled feature representations. We propose a novel framework utilizing a Top-k SAE to disentangle the dense [CLS] features into a structured latent space. Through this analysis, we reveal that in-distribution (ID) data exhibits consistent, class-specific activation patterns, which we formalize as Class Activation Profiles (CAPs). Our study uncovers a key structural invariant: while ID samples preserve a stable pattern within CAPs, OOD samples systematically disrupt this structure. Leveraging this insight, we introduce a scoring function based on the divergence of core energy profiles to quantify the deviation from ideal activation profiles. Our method achieves strong results on the FPR95 metric--critical for safety-sensitive applications--across multiple benchmarks, while also achieving competitive AUROC. Overall, our findings demonstrate that the sparse, disentangled features revealed by SAEs can serve as a powerful, interpretable tool for robust OOD detection in vision models.

Di Yang, Yaohui Wang, Shuai Shao, François Brémond, Jiangtao Wang

Real-world human action understanding remains challenging due to long-tailed label distributions, compositional motion patterns, and viewpoint variations. Existing skeleton-based methods often lack a structured and transferable representation of motion, and task-specific models for generation, classification, and detection are usually trained independently, resulting in fragmented pipelines and limited cross-task generalization. We present PRISM, a PRImitive-centric Skeleton Modeling framework that learns a shared motion representation from a motion generation objective and transfers it to perception tasks. PRISM represents each action sequence as a trajectory in a primitive coefficient space, which captures how a set of learned atomic motion primitives contribute to the observed motion. A structured decomposition module learns this representation in a physically grounded and view-invariant manner via motion generation. Instead of enforcing joint or unified training across tasks, PRISM provides a single primitive-centric representation that can be sequentially transferred to downstream classification and frame-wise detection through lightweight task heads. This representation introduces structure, compositionality, and improved generalization across distinct supervisions. PRISM consistently improves performance on long-tailed and multi-label datasets and enables interpretable reasoning over compositional and rare actions. Extensive experimental results show that the structured primitive space serves as a transferable and robust foundation for diverse action understanding tasks in real-world datasets.

Jianfei Zhao, Feng Zhang, Xin Sun, Chong Feng, Zhixing Tan

Visual attention serves as the primary mechanism through which MLLMs interpret visual information; however, its limited localization capability often leads to hallucinations. We observe that although MLLMs can accurately extract visual semantics from visual tokens, they fail to fully leverage this advantage during subsequent inference. To address this limitation, we propose Vision-Guided Attention (VGA), a training-free method that first constructs precise visual grounding by exploiting the semantic content of visual tokens, and then uses this grounding to guide the model's focus toward relevant visual regions. In image captioning, VGA further refines this guidance dynamically during generation by suppressing regions that have already been described. In VGA, each token undergoes only a single forward pass, introducing a negligible latency overhead. In addition, VGA is fully compatible with efficient attention implementations such as FlashAttention. Extensive experiments across diverse MLLMs and multiple hallucination benchmarks demonstrate that VGA achieves state-of-the-art dehallucination performance. Further analysis confirms that explicit visual guidance plays a crucial role in enhancing the visual understanding capabilities of MLLMs.

Ruoyu Chen, Xiaoqing Guo, Kangwei Liu, Siyuan Liang, Shiming Liu, Qunli Zhang, Laiyuan Wang, Hua Zhang, Xiaochun Cao

Multimodal large language models (MLLMs) have demonstrated remarkable capabilities in aligning visual inputs with natural language outputs. Yet, the extent to which generated tokens depend on visual modalities remains poorly understood, limiting interpretability and reliability. In this work, we present EAGLE, a lightweight black-box framework for explaining autoregressive token generation in MLLMs. EAGLE attributes any selected tokens to compact perceptual regions while quantifying the relative influence of language priors and perceptual evidence. The framework introduces an objective function that unifies sufficiency (insight score) and indispensability (necessity score), optimized via greedy search over sparsified image regions for faithful and efficient attribution. Beyond spatial attribution, EAGLE performs modality-aware analysis that disentangles what tokens rely on, providing fine-grained interpretability of model decisions. Extensive experiments across open-source MLLMs show that EAGLE consistently outperforms existing methods in faithfulness, localization, and hallucination diagnosis, while requiring substantially less GPU memory. These results highlight its effectiveness and practicality for advancing the interpretability of MLLMs.

Jian-Yu Jiang-Lin, Kang-Yang Huang, Ling Zou, Ling Lo, Sheng-Ping Yang, Yu-Wen Tseng, Kun-Hsiang Lin, Chia-Ling Chen, Yu-Ting Ta, Yan-Tsung Wang 等

Advances in generative modeling have made it increasingly easy to fabricate realistic portrayals of individuals, creating serious risks for security, communication, and public trust. Detecting such person-driven manipulations requires systems that not only distinguish altered content from authentic media but also provide clear and reliable reasoning. In this paper, we introduce TriDF, a comprehensive benchmark for interpretable DeepFake detection. TriDF contains high-quality forgeries from advanced synthesis models, covering 16 DeepFake types across image, video, and audio modalities. The benchmark evaluates three key aspects: Perception, which measures the ability of a model to identify fine-grained manipulation artifacts using human-annotated evidence; Detection, which assesses classification performance across diverse forgery families and generators; and Hallucination, which quantifies the reliability of model-generated explanations. Experiments on state-of-the-art multimodal large language models show that accurate perception is essential for reliable detection, but hallucination can severely disrupt decision-making, revealing the interdependence of these three aspects. TriDF provides a unified framework for understanding the interaction between detection accuracy, evidence identification, and explanation reliability, offering a foundation for building trustworthy systems that address real-world synthetic media threats.

Yansong Li, Zhongxi Qiu, Yun Tian, Zheng Jinyu, Shuo Li

Cardiac magnetic resonance (CMR) is the clinical gold standard for assessing cardiovascular diseases, but its interpretation relies on expert experience and remains challenging, particularly for identifying rare diseases. Existing automated methods lack interpretable reasoning processes, limiting clinical adoption. Although vision-language models (VLMs) possess basic visual understanding and text generation capabilities, they still lack verifiable reasoning chains in medical diagnosis and underperform on minority classes in long-tail distributions. To address these challenges, we propose CMR-RD, to our knowledge the first VLM for interpretable diagnosis in CMR, capable of generating explicit diagnostic chains aligned with imaging evidence. We construct a CMR dataset that reflects real-world clinical distributions, comprising five disease categories (including two rare conditions) plus normal controls. Building on this, the general-purpose VLM is aligned to medical and CMR semantics using large-scale medical vision-text data, and cold-start training is used to enhance its understanding of medical concepts and basic reasoning. To enhance reasoning and performance on rare samples, we propose Group Phase Policy Optimization (GPPO), which combines online multi-stage reinforcement learning (RL) with adaptive sampling. GPPO enables the model to proactively explore rare and underperforming classes, thereby effectively mitigating long-tail bias. Experiments demonstrate that CMR-RD achieves highly competitive accuracy and reasoning-chain correctness compared with medical and general VLMs, shows stronger recognition of rare categories.

Zhongyu Xia, Wenhao Chen, Yongtao Wang, Ming-Hsuan Yang

Visual-language reasoning, driving knowledge, and value alignment are essential for advanced autonomous driving systems. However, existing approaches largely rely on data-driven learning, making it difficult to capture the complex logic underlying decision-making through imitation or limited reinforcement rewards. To address this, we propose KnowVal, a new autonomous driving system that enables visual-language reasoning through the synergistic integration of open-world perception and knowledge retrieval. Specifically, we construct a comprehensive driving knowledge graph that encodes traffic laws, defensive driving principles, and ethical norms, complemented by an efficient LLM-based retrieval mechanism tailored for driving scenarios. Furthermore, we develop a human-preference dataset and train a Value Model to guide interpretable, value-aligned trajectory assessment. Experimental results show that our method substantially improves planning performance while remaining compatible with existing architectures. Notably, KnowVal achieves the lowest collision rate on nuScenes and state-of-the-art results on Bench2Drive and NAVSIM.

Subin Kim, Sangwoo Mo, Mamshad Nayeem Rizve, Yiran Xu, Difan Liu, Jinwoo Shin, Tobias Hinz

Achieving precise alignment between user intent and generated visuals remains a central challenge in text-to-visual generation, as a single attempt often fails to produce the desired output. To handle this, prior approaches mainly scale the visual generation process (e.g., increasing sampling steps or seeds), but this quickly leads to a quality plateau. This limitation arises because prompt adaptation does not scale along with the increasing population of generated samples. To address this, we propose Prompt Redesign for Inference-time Scaling, coined PRIS, a framework that adaptively revises the prompt during inference in response to the scaled visual generations. The core idea of PRIS is to review the generated visuals, identify recurring failure patterns across visuals, and redesign the prompt accordingly before regenerating the visuals with the revised prompt. To provide precise alignment feedback for prompt revision, we introduce a new verifier, element-level factual correction, which evaluates the alignment between prompt attributes and generated visuals at a fine-grained level, achieving more accurate and interpretable assessments than holistic measures. Extensive experiments on both text-to-image and text-to-video benchmarks demonstrate the effectiveness of our approach, including a 15% gain on VBench 2.0. These results highlight that jointly scaling prompts and visuals is key to fully leveraging scaling laws at inference-time.

Songyuan Yang, Weijiang Yu, Jilin Ma, Ziyu Liu, Guijian Tang, Wenjing Yang, Huibin Tan, Nong Xiao

Video reasoning has advanced with large multimodal models (LMMs), yet their inference is often a single pass that returns an answer without verifying whether the reasoning is evidence-aligned. We introduce **Reinforce to Learn, Elect to Reason (RLER)**, a dual paradigm that decouples learning to produce evidence from obtaining a reliable answer. In **RLER-Training**, we optimize the policy with group-relative reinforcement learning (RL) and 3 novel task-driven rewards: Frame-sensitive reward grounds reasoning on explicit key frames, Think-transparency reward shapes readable and parsable reasoning traces, and Anti-repetition reward boosts information density. These signals teach the model to emit structured, machine-checkable evidence and potentiate reasoning capabilities. In **RLER-Inference**, we apply a train-free orchestrator that generates a small set of diverse candidates, parses their answers and cited frames, scores them by evidence consistency, confidence, transparency, and non-redundancy, and then performs a robust evidence-weighted election. This closes the loop between producing and using evidence, improving reliability and interpretability without enlarging the model. We comprehensively evaluate RLER against various open-source and RL-based LMMs on 8 representative benchmarks. RLER achieves state of the art across all benchmarks and delivers an average improvement of 6.3 % over base models, while using on average 3.1 candidates per question, indicating a favorable balance between compute and quality. The results support a simple thesis: making evidence explicit during learning and electing by evidence during inference is a robust path to trustworthy video reasoning.

Qinfu Xu, Liyuan Pan, Yiwei Wei, Shaozu Yuan, Jiaqi Chen, Tianyu Liu

Multimodal Emotion Analysis (MEA) is crucial for human-centric AI, yet current methods struggle with two core challenges: the sparse nature of emotional cues across modalities and their inherent temporal asynchrony. Existing approaches, which often rely on implicit fusion, consequently suffer from diluted salient features and entangled representations. To address this issue, we propose EmoThinker, a new framework that advances MEA through explicit, structured reasoning. Our method introduces a structural token selection mechanism to concentrate on pivotal facial regions while refining background context, enhancing visual saliency and efficiency. For audio, an audio evidence extractor aggregates critical paralinguistic features into compact, emotion-rich tokens. More importantly, we enable step-by-step reasoning by constructing a Chain-of-Emotion-Thought dataset, which provides fine-grained annotations for disentangling asynchronous cues and resolving inter-modal conflicts. By decoupling evidence acquisition from reasoning, EmoThinker achieves a more interpretable and robust emotion analysis. Extensive experiments on multiple benchmarks demonstrate that our framework achieves new state-of-the-art performance.

Muhammad Atif Butt, Alexandra Gomez-Villa, Tao Wu, Javier Vazquez-Corral, Joost Van De Weijer, Kai Wang

Recent years have seen impressive advances in text-to-image generation, with image generative or unified models, generating high-quality images from text. Yet these models still struggle with fine-grained color control, often failing to accurately match colors specified in text prompts. While existing benchmarks evaluate compositional reasoning and prompt adherence, none systematically assess the color precision. Color is fundamental to human visual perception and communication, and critical for applications from art to design workflows requiring brand consistency. However, current benchmarks either neglect color or rely on coarse assessments, missing key capabilities like interpreting RGB values or aligning with human expectations. To this end, we propose GenColorBench, the first comprehensive benchmark for T2I color generation, grounded in color systems like ISCC-NBS and CSS3/X11, including numerical colors which are absent elsewhere. With 44K color-focused prompts covering 400+ colors, it reveals models' true capabilities via perceptual and automated assessments. Evaluations of popular T2I models on GenColorBench reveal significant performance variation, indicating which color conventions models understand the best and exposing their failure modes. Furthermore, GenColorBench provides insights to guide future improvements in precise color generation. The benchmark is available at https://moatifbutt.github.io/gencolorbench/.

Hiran Sarkar, Liming Kuang, Yordanka Velikova, Benjamin Busam

Predicting scene dynamics from visual observations is challenging. Existing methods capture dynamics only within observed boundaries failing to extrapolate far beyond the training sequence. Node-RF (Neural ODE-based NeRF) overcomes this limitation by integrating Neural Ordinary Differential Equations (NODEs) with dynamic Neural Radiance Fields (NeRFs), enabling a continuous-time, spatiotemporal representation that generalizes beyond observed trajectories at constant memory cost. From visual input, Node-RF learns an implicit scene state that evolves over time via an ODE solver, propagating feature embeddings via differential calculus. A NeRF-based renderer interprets calculated embeddings to synthesize arbitrary views for long-range extrapolation. Training on multiple motion sequences with shared dynamics allows for generalization to unseen conditions. Our experiments demonstrate that Node-RF can characterize abstract system behavior without explicit model to identify critical points for future predictions. Our code will be made publicly available.

ZeBin Ji, Yang Hu, Xiuli Bi, Bo Liu, Bin Xiao

It is essential for understanding neural network decisions to interpret the functionality (also known as concepts) of neurons. Existing approaches describe neuron concepts by generating natural language descriptions, thereby advancing the understanding of the neural network's decision-making mechanism. However, these approaches assume that each neuron has well-defined functions and provides discriminative features for neural network decision-making. In fact, some neurons may be redundant or may offer misleading concepts. Thus, the descriptions for such neurons may cause misinterpretations of the factors driving the neural network's decisions. To address the issue, we introduce a verification of neuron functions, which checks whether the generated concept highly activates the corresponding neuron. Furthermore, we propose a Select-Hypothesize-Verify framework for interpreting neuron functionality. This framework consists of: 1) selecting activation samples that best capture a neuron's well-defined functional behavior through activation-distribution analysis; 2) forming hypotheses about concepts for the selected neurons; and 3) verifying whether the generated concepts accurately reflect the functionality of the neuron. Extensive experiments show that our method produces more accurate neuron concepts. Our generated concepts activate the corresponding neurons with a probability approximately 1.5 times that of the current state-of-the-art method.

Hao Lu, Jiahao Wang, Yaolun Zhang, Ruohui Wang, Xuanyu Zheng, Yepeng Tang, Dahua Lin, Lewei Lu

We revisit video hallucination in multimodal large language models (Video-MLLMs) from a semantic aggregation perspective. While prior work attributes hallucinations to language priors, missing frames, or visual encoder biases, these explanations overlook errors arising during the aggregation of correct frame-level semantics into event-level interpretations. We term this phenomenon Semantic Aggregation Hallucination (SAH), which becomes increasingly prevalent in complex, multi-event video understanding tasks with rich temporal dependencies. To systematically study SAH, we introduce ELV-Halluc, the first benchmark designed for fine-grained evaluation of semantic aggregation errors. Our experiments reveal that SAH correlates with both semantic complexity and rapid semantic transitions. We further propose mitigation strategies: improved positional encoding preserves temporal structure, and reinforcement learning such as DPO enhances the model's ability to distinguish semantics within and across events. Using a curated 8K adversarial video-text pair dataset, our approach achieves consistent gains across benchmarks, including a 27.7% reduction in SAH rate on ELV-Halluc and Video-MME. Data and code are available at https://github.com/hlsv02/ELV-Halluc.

Jing Zuo, Lingzhou Mu, Fan Jiang, Chengcheng Ma, Mu Xu, Yonggang Qi

Achieving human-level performance in Vision-and-Language Navigation (VLN) requires an embodied agent to understand textual instructions, perceive visual observations, and reason over long action sequences. Recent works, such as NavCoT and NavGPT-2, demonstrate the potential of Chain-of-Thought (CoT) reasoning for improving interpretability and long-horizon planning. Moreover, multimodal extensions like OctoNav-R1 and CoT-VLA further validate CoT as a promising pathway toward human-like navigation reasoning. However, existing approaches face critical drawbacks: purely textual CoTs lack visual perception and easily overfit to sparse annotated reasoning steps, while multimodal CoTs incur severe token inflation by generating imagined visual observations, making real-time navigation impractical. In this work, we propose FantasyVLN, a unified implicit reasoning framework that preserves the benefits of CoT reasoning without explicit token overhead. Specifically, imagined visual tokens are encoded into a compact latent space using a pretrained Visual AutoRegressor (VAR) during CoT reasoning training, and the model jointly learns from textual, visual, and multimodal CoT modes under a unified multi-CoT strategy. At inference, our model performs direct instruction-to-action mapping while still enjoying reasoning-aware representations. Extensive experiments on LH-VLN show that our approach achieves reasoning-aware yet real-time navigation, improving success rates and efficiency while reducing inference latency by an order of magnitude compared to explicit CoT methods. Code is available at https://github.com/Fantasy-AMAP/fantasy-vln.

Li Jin, Weikai Chen, Yujie Wang, Yingda Yin, Zeyu Hu, Runze Zhang, Keyang Luo, Shengju Qian, Xin Wang, Xueying Qin

Open-world promptable 3D semantic segmentation remains brittle as semantics are inferred in the input sensor coordinates. Yet, humans, in contrast, interpret parts via functional roles in a canonical space -- wings extend laterally, handles protrude to the side, and legs support from below. Psychophysical evidence shows that we mentally rotate objects into canonical frames to reveal these roles. To fill this gap, we propose CoSMo3D, which attains canonical space perception by inducing a latent canonical reference frame learned directly from data. By construction, we create a unified canonical dataset through LLM-guided intra- and cross-category alignment, exposing canonical spatial regularities across 200 categories. By induction, we realize canonicality inside the model through a dual-branch architecture with canonical map anchoring and canonical box calibration, collapsing pose variation and symmetry into a stable canonical embedding. This shift from input pose space to canonical representation yields far more stable and transferable part semantics. Experimental results show that CoSMo3D establishes new state of the art in open-world promptable 3D segmentation.

Xiangyu Sun, Haoyi Jiang, Liu Liu, Seungtae Nam, Gyeongjin Kang, Xinjie Wang, Wei Sui, Zhizhong Su, Wenyu Liu, Xinggang Wang 等

Reconstructing and semantically interpreting 3D scenes from sparse 2D views remains a fundamental challenge in computer vision. Conventional methods often decouple semantic understanding from reconstruction or necessitate costly per-scene optimization, thereby restricting their scalability and generalizability. In this paper, we introduce Uni3R, a novel feed-forward framework that jointly reconstructs a unified 3D scene representation enriched with open-vocabulary semantics, directly from unposed multi-view images. Our approach leverages a Cross-View Transformer to robustly integrate information across arbitrary multi-view inputs, which then regresses a set of 3D Gaussian primitives endowed with semantic feature fields. This unified representation facilitates high-fidelity novel view synthesis, open-vocabulary 3D semantic segmentation, and depth prediction--all within a single, feed-forward pass. Extensive experiments demonstrate that Uni3R sets a new state of the art across multiple benchmarks, including in-domain datasets such as RE10K and ScanNet, as well as the out-of-domain dataset Mip-NeRF360. This work represents a new paradigm toward generalizable and unified 3D scene reconstruction and understanding.

Hani Alomari, Ali Asgarov, Chris Thomas

Most vision-language models assume images have a single literal meaning, even though images are inherently polysemous. We propose a retrieval paradigm that models many-to-many relationships between images and text using interpretive lenses and introduce Lenses, a multi-prompt embedding model and dataset for polysemous image-text retrieval. The Lenses dataset contains 105,669 images and 732,405 captions, with each image paired with multiple captions and image-side prompts annotated across five categories: Literal, Figurative, Abstract, Background, and Emotional. Building on a multimodal large language model, the Lenses model uses learned lens tokens to extract lens-specific embeddings for every image and caption and compares these using a lens-masking similarity function with a global fallback that prioritizes same-lens matches while retaining a global pathway. Training uses a category-aware multi-positive contrastive loss and intra-set diversity regularization to align corresponding perspectives while preventing semantic collapse across lenses. We further propose lens-aware evaluation protocols, including category-aware ranking, that better reflect how humans match images and text. Experiments on the Lenses dataset and public benchmarks show that our model outperforms baselines on literal and non-literal retrieval and reduces over-reliance on literal cues.

Samarth Chopra, Jing Liang, Gershom Seneviratne, Dinesh Manocha

Understanding physical properties such as friction, stiffness, hardness, and material composition is essential for enabling robots to interact safely and effectively with their surroundings. However, existing 3D reconstruction methods focus on geometry and appearance and cannot infer these underlying physical properties. We present PhysGS, a Bayesian-inferred extension of 3D Gaussian Splatting that estimates dense, per-point physical properties from visual cues and vision--language priors. We formulate property estimation as Bayesian inference over Gaussian splats, where material and property beliefs are iteratively refined as new observations arrive. PhysGS also models aleatoric and epistemic uncertainties, enabling uncertainty-aware object and scene interpretation. Across object-scale (ABO-500), indoor, and outdoor real-world datasets, PhysGS improves accuracy of the mass estimation by up to 22.8%, reduces Shore hardness error by up to 61.2%, and lowers kinetic friction error by up to 18.1% compared to deterministic baselines. Our results demonstrate that PhysGS unifies 3D reconstruction, uncertainty modeling, and physical reasoning in a single, spatially continuous framework for dense physical property estimation.

Ming Wang, Haoxuan Qu, Qiuhong Ke, Wei Zhou, Hossein Rahmani, Jun Liu

Surface electromyography (sEMG) signal-based activity recognition has attracted increasing research attention in recent years. To develop accurate sEMG signal-based activity recognizers, numerous approaches have been proposed. Some studies focus on designing larger and more expressive model architectures to enhance the representational capacity of sEMG signals, while others aim to enrich model priors through large-scale pretraining, thereby improving recognition performance. Recently, large language models (LLMs) have shown remarkable generalization and reasoning capabilities in natural language processing, whose implicit knowledge, learned from extensive linguistic descriptions of actions, opens new possibilities for interpreting sEMG signals and inferring activity intentions. Motivated by this, we propose LLM-sEMG, a novel framework that leverages LLMs as sEMG activity recognizers. Within this framework, we design a language-oriented mapping mechanism that converts continuous sEMG sequences into "sEMG language," integrating several strategies to further facilitate the signal-to-language mapping process. Extensive experiments demonstrate that the proposed framework achieves highly accurate sEMG signal-based activity recognition using large language models.