Haiyun Guo

Yandu Sun, Zhiyan Hou, Haokai Ma et al.

Model merging offers a training-free way to combine multiple post-trained expert models, but merging experts obtained through reinforcement learning (RL) remains challenging. Existing spectral merging methods often assume that leading singular directions contain the main task signal, while lower-energy residual components can be compressed, selected, or attenuated to reduce interference. We find that this assumption does not hold for RL task vectors: after decomposing each task vector into a leading spectral head and a residual component, both parts can independently recover substantial behavior knowledge, while exhibiting different merging properties. The head is highly concentrated and informative but more prone to sharp cross-expert conflicts, whereas the residual component is more dispersed and provides a more stable basis for aggregation. Based on this observation, we propose ResMerge, a residual-based spectral merging framework for RL experts. ResMerge first constructs a stable residual backbone with Spherical Residual Consensus Adaptation, which estimates a reliability-weighted consensus direction on the Frobenius sphere. It then reintroduces leading-head information through a Lightweight Head Correction module gated by positive cross-expert agreement. Experiments across multiple RL expert groups and capability domains show that ResMerge better preserves expert capabilities than representative task-vector and spectral merging baselines. The implementation of ResMerge is publicly available at https://github.com/sunyd0303-cpu/ResMerge-release.

Yongqi An, Xu Zhao, Tao Yu et al.

Background subtraction (BGS) aims to extract all moving objects in the video frames to obtain binary foreground segmentation masks. Deep learning has been widely used in this field. Compared with supervised-based BGS methods, unsupervised methods have better generalization. However, previous unsupervised deep learning BGS algorithms perform poorly in sophisticated scenarios such as shadows or night lights, and they cannot detect objects outside the pre-defined categories. In this work, we propose an unsupervised BGS algorithm based on zero-shot object detection called Zero-shot Background Subtraction (ZBS). The proposed method fully utilizes the advantages of zero-shot object detection to build the open-vocabulary instance-level background model. Based on it, the foreground can be effectively extracted by comparing the detection results of new frames with the background model. ZBS performs well for sophisticated scenarios, and it has rich and extensible categories. Furthermore, our method can easily generalize to other tasks, such as abandoned object detection in unseen environments. We experimentally show that ZBS surpasses state-of-the-art unsupervised BGS methods by 4.70% F-Measure on the CDnet 2014 dataset. The code is released at https://github.com/CASIA-IVA-Lab/ZBS.

Kuan Zhu, Haiyun Guo, Tianyi Yan et al.

In person re-identification (ReID), very recent researches have validated pre-training the models on unlabelled person images is much better than on ImageNet. However, these researches directly apply the existing self-supervised learning (SSL) methods designed for image classification to ReID without any adaption in the framework. These SSL methods match the outputs of local views (e.g., red T-shirt, blue shorts) to those of the global views at the same time, losing lots of details. In this paper, we propose a ReID-specific pre-training method, Part-Aware Self-Supervised pre-training (PASS), which can generate part-level features to offer fine-grained information and is more suitable for ReID. PASS divides the images into several local areas, and the local views randomly cropped from each area are assigned with a specific learnable [PART] token. On the other hand, the [PART]s of all local areas are also appended to the global views. PASS learns to match the output of the local views and global views on the same [PART]. That is, the learned [PART] of the local views from a local area is only matched with the corresponding [PART] learned from the global views. As a result, each [PART] can focus on a specific local area of the image and extracts fine-grained information of this area. Experiments show PASS sets the new state-of-the-art performances on Market1501 and MSMT17 on various ReID tasks, e.g., vanilla ViT-S/16 pre-trained by PASS achieves 92.2\%/90.2\%/88.5\% mAP accuracy on Market1501 for supervised/UDA/USL ReID. Our codes are available at https://github.com/CASIA-IVA-Lab/PASS-reID.

Tianyue Wang, Leigang Qu, Tianyu Yang et al.

Zero-Shot Composed Image Retrieval (ZS-CIR) aims to retrieve target images given a multimodal query (comprising a reference image and a modification text), without training on annotated triplets. Existing methods typically convert the multimodal query into a single modality-either as an edited caption for Text-to-Image retrieval (T2I) or as an edited image for Image-to-Image retrieval (I2I). However, each paradigm has inherent limitations: T2I often loses fine-grained visual details, while I2I struggles with complex semantic modifications. To effectively leverage their complementary strengths under diverse query intents, we propose WISER, a training-free framework that unifies T2I and I2I via a "retrieve-verify-refine" pipeline, explicitly modeling intent awareness and uncertainty awareness. Specifically, WISER first performs Wider Search by generating both edited captions and images for parallel retrieval to broaden the candidate pool. Then, it conducts Adaptive Fusion with a verifier to assess retrieval confidence, triggering refinement for uncertain retrievals, and dynamically fusing the dual-path for reliable ones. For uncertain retrievals, WISER generates refinement suggestions through structured self-reflection to guide the next retrieval round toward Deeper Thinking. Extensive experiments demonstrate that WISER significantly outperforms previous methods across multiple benchmarks, achieving relative improvements of 45% on CIRCO (mAP@5) and 57% on CIRR (Recall@1) over existing training-free methods. Notably, it even surpasses many training-dependent methods, highlighting its superiority and generalization under diverse scenarios. Code will be released at https://github.com/Physicsmile/WISER.

Gengsheng Li, Jinghan He, Shijie Wang et al.

Self-play bootstraps LLM reasoning through an iterative Challenger-Solver loop: the Challenger is trained to generate questions that target the Solver's capabilities, and the Solver is optimized on the generated data to expand its reasoning skills. However, existing frameworks like R-Zero often exhibit non-sustained improvement, where early gains degrade as self-play continues. We identify a key failure mode, Diversity Illusion, where the Solver's training signals appear diverse yet collapse into recurring underlying patterns. It manifests as (1) Local Diversity Illusion, where diversity is enforced only within-batch, inducing cross-iteration mode cycling; and (2) Surface Diversity Illusion, where questions vary superficially but require near-identical reasoning skills. To mitigate them, we propose R-Diverse with two aligned innovations: Memory-Augmented Penalty (MAP), which uses a persistent memory bank to discourage recycling across iterations, and Skill-Aware Measurement (SAM), which evaluates diversity by the reasoning skills exercised rather than surface variation of questions. Across 10 math and general reasoning benchmarks, R-Diverse sustains gains over more iterations and consistently outperforms prior self-play methods. Code is available at https://github.com/Gengsheng-Li/R-Diverse.

Jinghan He, Haiyun Guo, Ming Tang et al.

Instruction tuning is now a widely adopted approach to aligning large multimodal models (LMMs) to follow human intent. It unifies the data format of vision-language tasks, enabling multi-task joint training. However, vision-language tasks are constantly being created in practice. Instead of always re-training LMMs when new tasks arrive, continual learning offers flexibility for models to continually and efficiently exploit the evolving data. This work aims to explore the following two questions: 1) Do LMMs still suffer from catastrophic forgetting in continual instruction tuning? 2) Are the existing three classes of continual learning methods still applicable to the continual instruction tuning of LMMs? An extensive study is conducted to address the above questions. First, we establish the first benchmark in this setting and reveal that catastrophic forgetting is still observed when continually instruction-tuning LMMs. However, the multi-task joint instruction tuning can facilitate the model's continual learning ability and mitigate forgetting. Second, we integrate and adapt classic continual learning methods to our context, demonstrating the efficacy of data replay and model expansion strategies across diverse scenarios. In contrast, regularization-based methods only perform well on models that have been jointly instruction-tuned on multiple tasks. Third, we delve into the correlation and forgetting dynamics between vision-language task pairs and propose task-similarity-informed regularization and model expansion methods for continual instruction tuning of LMMs. Experimental results show that our approach consistently boosts the model's performance.

Tianyi Yan, Kuan Zhu, Haiyun guo et al.

Clustering-based methods, which alternate between the generation of pseudo labels and the optimization of the feature extraction network, play a dominant role in both unsupervised learning (USL) and unsupervised domain adaptive (UDA) person re-identification (Re-ID). To alleviate the adverse effect of noisy pseudo labels, the existing methods either abandon unreliable labels or refine the pseudo labels via mutual learning or label propagation. However, a great many erroneous labels are still accumulated because these methods mostly adopt traditional unsupervised clustering algorithms which rely on certain assumptions on data distribution and fail to capture the distribution of complex real-world data. In this paper, we propose the plug-and-play graph-based pseudo label correction network (GLC) to refine the pseudo labels in the manner of supervised clustering. GLC is trained to perceive the varying data distribution at each epoch of the self-training with the supervision of initial pseudo labels generated by any clustering method. It can learn to rectify the initial noisy labels by means of the relationship constraints between samples on the k Nearest Neighbor (kNN) graph and early-stop training strategy. Specifically, GLC learns to aggregate node features from neighbors and predict whether the nodes should be linked on the graph. Besides, GLC is optimized with 'early stop' before the noisy labels are severely memorized to prevent overfitting to noisy pseudo labels. Consequently, GLC improves the quality of pseudo labels though the supervision signals contain some noise, leading to better Re-ID performance. Extensive experiments in USL and UDA person Re-ID on Market-1501 and MSMT17 show that our method is widely compatible with various clustering-based methods and promotes the state-of-the-art performance consistently.

Xiangzhao Hao, Shijie Wang, Tianyu Yang et al.

Universal Multimodal Retrieval requires unified embedding models capable of interpreting diverse user intents, ranging from simple keywords to complex compositional instructions. While Multimodal Large Language Models (MLLMs) possess strong reasoning capabilities, prevailing adaptations confine them to static encoders, underutilizing their generative potential. This encoder-only paradigm struggles with complex intents that demand logical deduction rather than superficial pattern matching. To address this, we introduce TRACE (Task-adaptive Reasoning And Compressing Embeddings). TRACE unifies generative reasoning with discriminative representation learning. It first generates a structured Chain-of-Thought (CoT) to explicitly reason about the query, and subsequently compresses this reasoning trace into a compact embedding via a dedicated token. To train this framework, we construct M-BEIR-CoT, a large-scale dataset featuring a difficulty-aware routing strategy. Experiments on the M-BEIR benchmark establish TRACE as the new state-of-the-art. Crucially, TRACE demonstrates a learned implicit routing behavior. It autonomously activates reasoning for complex queries while bypassing it for simpler ones, achieving an optimal balance between retrieval accuracy and inference throughput. Furthermore, by internalizing the deductive process, TRACE exhibits remarkable zero-shot transferability to unseen domains and novel constraints.

Junfeng Fang, Zhepei Hong, Mao Zheng et al.

On-policy distillation (OPD) is a powerful paradigm for model alignment, yet its reliance on teacher logits restricts its application to white-box scenarios. We contend that structured semantic rubrics can serve as a scalable alternative to teacher logits, enabling OPD using only teacher-generated responses. To prove it, we introduce ROPD, a simple yet foundational framework for rubric-based OPD. Specifically, ROPD induces prompt-specific rubrics from teacher-student contrasts, and then utilizes these rubrics to score the student rollouts for on-policy optimization. Empirically, ROPD outperforms the advanced logit-based OPD methods across most scenarios, and achieving up to a 10x gain in sample efficiency. These results position rubric-based OPD as a flexible, black-box-compatible alternative to the prevailing logit-based OPD, offering a simple yet strong baseline for scalable distillation across proprietary and open-source LLMs. Code is available at https://github.com/Peregrine123/ROPD_official.

Tianyu Yang, ChenWei He, Xiangzhao Hao et al.

Composed Image Retrieval (CIR) aims to retrieve target images based on a hybrid query comprising a reference image and a modification text. Early dual-tower Vision-Language Models (VLMs) struggle with cross-modality compositional reasoning required for this task. Recently, adapting generative Multimodal Large Language Models (MLLMs) for retrieval offers a promising direction. However, we identify that this adaptation strategy overlooks a fundamental issue: adapting a generative MLLM into a single-embedding discriminative retriever triggers a paradigm conflict, which leads to Capability Degradation - the deterioration of native fine-grained reasoning after retrieval adaptation. To address this challenge, we propose ReCALL (Recalibrating Capability Degradation), a model-agnostic framework that follows a diagnose-generate-refine pipeline: Firstly, we diagnose cognitive blind spots of the retriever via self-guided informative instance mining. Next, we generate corrective instructions and triplets by CoT prompting the foundation MLLM and conduct quality control with VQA-based consistency filtering. Finally, we refine the retriever through continual training on these triplets with a grouped contrastive scheme, thereby internalizing fine-grained visual-semantic distinctions and realigning the discriminative embedding space of retriever with intrinsic compositional reasoning within the MLLM. Extensive experiments on CIRR and FashionIQ show that ReCALL consistently recalibrates degraded capabilities and achieves state-of-the-art performance. Code will be released soon.

Yicheng Xiao, Yu Chen, Haoxuan Ma et al.

While the Contrastive Language-Image Pretraining(CLIP) model has achieved remarkable success in a variety of downstream vison language understanding tasks, enhancing its capability for fine-grained image-text alignment remains an active research focus. To this end, most existing works adopt the strategy of explicitly increasing the granularity of visual information processing, e.g., incorporating visual prompts to guide the model focus on specific local regions within the image. Meanwhile, researches on Multimodal Large Language Models(MLLMs) have demonstrated that training with long and detailed textual descriptions can effectively improve the model's fine-grained vision-language alignment. However, the inherent token length limitation of CLIP's text encoder fundamentally limits CLIP to process more granular textual information embedded in long text sequences. To synergistically leverage the advantages of enhancing both visual and textual content processing granularity, we propose PixCLIP, a novel framework designed to concurrently accommodate visual prompt inputs and process lengthy textual descriptions. Specifically, we first establish an automated annotation pipeline capable of generating pixel-level localized, long-form textual descriptions for images. Utilizing this pipeline, we construct LongGRIT, a high-quality dataset comprising nearly 1.5 million samples. Secondly, we replace CLIP's original text encoder with the LLM and propose a three-branch pixel-text alignment learning framework, facilitating fine-grained alignment between image regions and corresponding textual descriptions at arbitrary granularity. Experiments demonstrate that PixCLIP showcases breakthroughs in pixel-level interaction and handling long-form texts, achieving state-of-the-art performance.

Xin He, Haiyun Guo, Kuan Zhu et al.

Lane detection plays an important role in autonomous driving perception systems. As deep learning algorithms gain popularity, monocular lane detection methods based on them have demonstrated superior performance and emerged as a key research direction in autonomous driving perception. The core designs of these algorithmic frameworks can be summarized as follows: (1) Task paradigm, focusing on lane instance-level discrimination; (2) Lane modeling, representing lanes as a set of learnable parameters in the neural network; (3) Global context supplementation, enhancing inference on the obscure lanes; (4) Perspective effect elimination, providing accurate 3D lanes for downstream applications. From these perspectives, this paper presents a comprehensive overview of existing methods, encompassing both the increasingly mature 2D lane detection approaches and the developing 3D lane detection works. Besides, this paper compares the performance of mainstream methods on different benchmarks and investigates their inference speed under a unified setting for fair comparison. Moreover, we present some extended works on lane detection, including multi-task perception, video lane detection, online high-definition map construction, and lane topology reasoning, to offer readers a comprehensive roadmap for the evolution of lane detection. Finally, we point out some potential future research directions in this field. We exhaustively collect the papers and codes of existing works at https://github.com/Core9724/Awesome-Lane-Detection and will keep tracing the research.

Weiyun Jiang, Haiyun Guo, Christopher A. Metzler et al.

This work introduces the first closed-loop adaptive optics (AO) system capable of optically correcting aberrations in real-time without a guidestar or a wavefront sensor. Nearly 40 years ago, Cederquist et al. demonstrated that asymmetric apertures enable phase retrieval (PR) algorithms to perform fully computational wavefront sensing, albeit at a high computational cost. More recently, Chimitt et al. extended this approach with machine learning and demonstrated real-time wavefront sensing using only a single (guidestar-based) point-spread-function (PSF) measurement. Inspired by these works, we introduce a guidestar-free AO framework built around asymmetric apertures and machine learning. Our approach combines three key elements: (1) an asymmetric aperture placed at the system's pupil plane that enables PR-based wavefront sensing, (2) a pair of machine learning algorithms that estimate the PSF from natural scene measurements and reconstruct phase aberrations, and (3) a spatial light modulator that performs optical correction. We experimentally validate this framework on dense natural scenes imaged through unknown obscurants. Our method outperforms state-of-the-art guidestar-free wavefront shaping methods, using an order of magnitude fewer measurements and three orders of magnitude less computation.

Xiangzhao Hao, Kuan Zhu, Hongyu Guo et al.

Using natural language to query visual information is a fundamental need in real-world applications. Text-Image Retrieval (TIR) retrieves a target image from a gallery based on an image-level description, while Referring Expression Comprehension (REC) localizes a target object within a given image using an instance-level description. However, real-world applications often present more complex demands. Users typically query an instance-level description across a large gallery and expect to receive both relevant image and the corresponding instance location. In such scenarios, TIR struggles with fine-grained descriptions and object-level localization, while REC is limited in its ability to efficiently search large galleries and lacks an effective ranking mechanism. In this paper, we introduce a new task called \textbf{Referring Expression Instance Retrieval (REIR)}, which supports both instance-level retrieval and localization based on fine-grained referring expressions. First, we propose a large-scale benchmark for REIR, named REIRCOCO, constructed by prompting advanced vision-language models to generate high-quality referring expressions for instances in the MSCOCO and RefCOCO datasets. Second, we present a baseline method, Contrastive Language-Instance Alignment with Relation Experts (CLARE), which employs a dual-stream architecture to address REIR in an end-to-end manner. Given a referring expression, the textual branch encodes it into a query embedding. The visual branch detects candidate objects and extracts their instance-level visual features. The most similar candidate to the query is selected for bounding box prediction. CLARE is first trained on object detection and REC datasets to establish initial grounding capabilities, then optimized via Contrastive Language-Instance Alignment (CLIA) for improved retrieval across images. We will release our code and benchmark publicly.

Zhenglin Hua, Jinghan He, Zijun Yao et al.

Large vision-language models (LVLMs) have achieved remarkable performance on multimodal tasks. However, they still suffer from hallucinations, generating text inconsistent with visual input, posing significant risks in real-world applications. Existing approaches to address this issue focus on incorporating external knowledge bases, alignment training, or decoding strategies, all of which require substantial computational cost and time. Recent works try to explore more efficient alternatives by adjusting LVLMs' internal representations. Although promising, these methods may cause hallucinations to be insufficiently suppressed or lead to excessive interventions that negatively affect normal semantics. In this work, we leverage sparse autoencoders (SAEs) to identify semantic directions closely associated with faithfulness or hallucination, extracting more precise and disentangled hallucination-related representations. Our analysis demonstrates that interventions along the identified faithful direction can mitigate hallucinations, while those along the hallucinatory direction can exacerbate them. Building on these insights, we propose Steering LVLMs via SAE Latent Directions (SSL), a plug-and-play method based on SAE-derived latent directions to mitigate hallucinations in LVLMs. Extensive experiments demonstrate that SSL significantly outperforms existing decoding approaches in mitigating hallucinations, while maintaining transferability across different model architectures with negligible additional time overhead. The code is available at https://github.com/huazhenglin2003/SSL.

Kuan Zhu, Haiyun Guo, Zhiwei Liu et al.

Existing alignment-based methods have to employ the pretrained human parsing models to achieve the pixel-level alignment, and cannot identify the personal belongings (e.g., backpacks and reticule) which are crucial to person re-ID. In this paper, we propose the identity-guided human semantic parsing approach (ISP) to locate both the human body parts and personal belongings at pixel-level for aligned person re-ID only with person identity labels. We design the cascaded clustering on feature maps to generate the pseudo-labels of human parts. Specifically, for the pixels of all images of a person, we first group them to foreground or background and then group the foreground pixels to human parts. The cluster assignments are subsequently used as pseudo-labels of human parts to supervise the part estimation and ISP iteratively learns the feature maps and groups them. Finally, local features of both human body parts and personal belongings are obtained according to the selflearned part estimation, and only features of visible parts are utilized for the retrieval. Extensive experiments on three widely used datasets validate the superiority of ISP over lots of state-of-the-art methods. Our code is available at https://github.com/CASIA-IVA-Lab/ISP-reID.

Gengsheng Li, Tianyu Yang, Junfeng Fang et al.

Reinforcement learning with verifiable rewards (RLVR) has become a standard paradigm for post-training large language models. While Group Relative Policy Optimization (GRPO) is widely adopted, its coarse credit assignment uniformly penalizes failed rollouts, lacking the token-level focus needed to efficiently address specific deviations. Self-Distillation Policy Optimization (SDPO) addresses this by providing denser, more targeted logit-level supervision that facilitates rapid early improvement, yet it frequently collapses during prolonged training. We trace this late-stage instability to two intrinsic flaws: self-distillation on already-correct samples introduces optimization ambiguity, and the self-teacher's signal reliability progressively degrades. To resolve these issues, we propose Sample-Routed Policy Optimization (SRPO), a unified on-policy framework that routes correct samples to GRPO's reward-aligned reinforcement and failed samples to SDPO's targeted logit-level correction. SRPO further incorporates an entropy-aware dynamic weighting mechanism to suppress high-entropy, unreliable distillation targets while emphasizing confident ones. Evaluated across five benchmarks and two model scales, SRPO achieves both the rapid early improvement of SDPO and the long-horizon stability of GRPO. It consistently surpasses the peak performance of both baselines, raising the five-benchmark average on Qwen3-8B by 3.4% over GRPO and 6.3% over SDPO, while simultaneously yielding moderate response lengths and lowering per-step compute cost by up to 17.2%.

Lin Sha, Haiyun Guo, Tao Wang et al.

Vision-Language Models (VLMs) have become central to autonomous driving systems, yet their deployment is severely bottlenecked by the massive computational overhead of multi-view camera and multi-frame video input. Existing token pruning methods, primarily designed for single-image inputs, treat each frame or view in isolation and thus fail to exploit the inherent spatio-temporal redundancies in driving scenarios. To bridge this gap, we propose ST-Prune, a training-free, plug-and-play framework comprising two complementary modules: Motion-aware Temporal Pruning (MTP) and Ring-view Spatial Pruning (RSP). MTP addresses temporal redundancy by encoding motion volatility and temporal recency as soft constraints within the diversity selection objective, prioritizing dynamic trajectories and current-frame content over static historical background. RSP further resolves spatial redundancy by exploiting the ring-view camera geometry to penalize bilateral cross-view similarity, eliminating duplicate projections and residual background that temporal pruning alone cannot suppress. These two modules together constitute a complete spatio-temporal pruning process, preserving key scene information under strict compression. Validated across four benchmarks spanning perception, prediction, and planning, ST-Prune establishes new state-of-the-art for training-free token pruning. Notably, even at 90\% token reduction, ST-Prune achieves near-lossless performance with certain metrics surpassing the full-model baseline, while maintaining inference speeds comparable to existing pruning approaches.

Jinghan He, Kuan Zhu, Haiyun Guo et al.

Large vision-language models (LVLMs) have made substantial progress in integrating large language models (LLMs) with visual inputs, enabling advanced multimodal reasoning. Despite their success, a persistent challenge is hallucination-where generated text fails to accurately reflect visual content-undermining both accuracy and reliability. Existing methods focus on alignment training or decoding refinements but primarily address symptoms at the generation stage without probing the underlying causes. In this work, we investigate the internal mechanisms driving hallucination in LVLMs, with an emphasis on the multi-head attention module. Specifically, we introduce Vision-aware Head Divergence (VHD), a metric that quantifies the sensitivity of attention head outputs to visual context. Based on this, our findings reveal the presence of vision-aware attention heads that are more attuned to visual information; however, the model's overreliance on its prior language patterns is closely related to hallucinations. Building on these insights, we propose Vision-aware Head Reinforcement (VHR), a training-free approach to mitigate hallucination by enhancing the role of vision-aware attention heads. Extensive experiments demonstrate that our method achieves superior performance compared to state-of-the-art approaches in mitigating hallucinations, while maintaining high efficiency with negligible additional time overhead.

Weijie Zhou, Manli Tao, Chaoyang Zhao et al.

Understanding the environment and a robot's physical reachability is crucial for task execution. While state-of-the-art vision-language models (VLMs) excel in environmental perception, they often generate inaccurate or impractical responses in embodied visual reasoning tasks due to a lack of understanding of robotic physical reachability. To address this issue, we propose a unified representation of physical reachability across diverse robots, i.e., Space-Physical Reachability Map (S-P Map), and PhysVLM, a vision-language model that integrates this reachability information into visual reasoning. Specifically, the S-P Map abstracts a robot's physical reachability into a generalized spatial representation, independent of specific robot configurations, allowing the model to focus on reachability features rather than robot-specific parameters. Subsequently, PhysVLM extends traditional VLM architectures by incorporating an additional feature encoder to process the S-P Map, enabling the model to reason about physical reachability without compromising its general vision-language capabilities. To train and evaluate PhysVLM, we constructed a large-scale multi-robot dataset, Phys100K, and a challenging benchmark, EQA-phys, which includes tasks for six different robots in both simulated and real-world environments. Experimental results demonstrate that PhysVLM outperforms existing models, achieving a 14\% improvement over GPT-4o on EQA-phys and surpassing advanced embodied VLMs such as RoboMamba and SpatialVLM on the RoboVQA-val and OpenEQA benchmarks. Additionally, the S-P Map shows strong compatibility with various VLMs, and its integration into GPT-4o-mini yields a 7.1\% performance improvement.

Jinghan He, Haiyun Guo, Kuan Zhu et al.

Continual learning (CL) is crucial for language models to dynamically adapt to the evolving real-world demands. To mitigate the catastrophic forgetting problem in CL, data replay has been proven a simple and effective strategy, and the subsequent data-replay-based distillation can further enhance the performance. However, existing methods fail to fully exploit the knowledge embedded in models from previous tasks, resulting in the need for a relatively large number of replay samples to achieve good results. In this work, we first explore and emphasize the importance of attention weights in knowledge retention, and then propose a SElective attEntion-guided Knowledge Retention method (SEEKR) for data-efficient replay-based continual learning of large language models (LLMs). Specifically, SEEKR performs attention distillation on the selected attention heads for finer-grained knowledge retention, where the proposed forgettability-based and task-sensitivity-based measures are used to identify the most valuable attention heads. Experimental results on two continual learning benchmarks for LLMs demonstrate the superiority of SEEKR over the existing methods on both performance and efficiency. Explicitly, SEEKR achieves comparable or even better performance with only 1/10 of the replayed data used by other methods, and reduces the proportion of replayed data to 1%.

Mingyang Xie, Haiyun Guo, Brandon Y. Feng et al.

Imaging through scattering media is a fundamental and pervasive challenge in fields ranging from medical diagnostics to astronomy. A promising strategy to overcome this challenge is wavefront modulation, which induces measurement diversity during image acquisition. Despite its importance, designing optimal wavefront modulations to image through scattering remains under-explored. This paper introduces a novel learning-based framework to address the gap. Our approach jointly optimizes wavefront modulations and a computationally lightweight feedforward "proxy" reconstruction network. This network is trained to recover scenes obscured by scattering, using measurements that are modified by these modulations. The learned modulations produced by our framework generalize effectively to unseen scattering scenarios and exhibit remarkable versatility. During deployment, the learned modulations can be decoupled from the proxy network to augment other more computationally expensive restoration algorithms. Through extensive experiments, we demonstrate our approach significantly advances the state of the art in imaging through scattering media. Our project webpage is at https://wavemo-2024.github.io/.

Chenwei He, Xiangzhao Hao, Tianyu Yang et al.

Universal multimodal embedding (UME) maps heterogeneous inputs into a shared retrieval space with a single model. Recent approaches improve UME by generating explicit chain-of-thought (CoT) rationales before extracting embeddings, enabling multimodal large language models to better infer complex query intent. However, explicit CoT incurs substantial inference overhead and can compress rich multimodal evidence into a narrow textual bottleneck. We propose PLUME, a latent reasoning framework that advances UME by replacing verbalized CoT with a short autoregressive rollout of continuous latent states. To support diverse multimodal queries, PLUME further introduces a semantic-anchor-guided transition adapter that steers latent rollout along different reasoning trajectories under the same fixed computation budget. To stabilize training, PLUME adopts a progressive explicit-to-latent curriculum that uses verbalized reasoning only as a temporary training scaffold and gradually transfers this behavior into hidden-state computation, eliminating explicit CoT at inference. On the 78-task MMEB-v2 benchmark, PLUME outperforms strong explicit-CoT UME baselines while reducing reasoning from hundreds of generated tokens to fewer than 10 latent steps, delivering over 30x faster inference. PLUME is especially well suited to retrieval settings where relevant evidence is dense, structurally complex, and difficult to organize through verbalized intermediate rationales, such as video and visual document retrieval. These results show that structured latent computation can preserve the benefits of intermediate reasoning without the overhead of explicit rationale generation, providing a stronger and more efficient paradigm for practical retrieval systems.

Yuzhe Zhou, Zhenglin Hua, Haiyun Guo et al.

Dynamic data selection accelerates training by sampling a changing subset of the dataset while preserving accuracy. We rethink two core notions underlying sample evaluation: representativeness and diversity. Instead of local geometric centrality, we define representativeness as coverage of dataset-level common or high-frequency feature factors. Instead of within-subset dispersion, we define diversity at the process level, requiring the selection trajectory to gradually include complementary rare factors over training. Based on this view, we propose a dynamic selection framework with three components. First, we score representativeness in a plug-in feature space to prioritize samples covering frequent factors. We instantiate this with a sparse autoencoder trained on the target dataset, using sparse unit activations to summarize both individual samples and dataset-wide factor statistics. Second, we realize process-level diversity by combining rare-factor sampling with a Usage-Frequency Penalty that promotes sample rotation, provably discourages monopoly, and reduces gradient bias. Third, we couple the two-dimensional scoring with a smooth scheduler that transitions selection from core-pattern consolidation to rare-factor exploration, without extra gradients, influence estimates, or second-order computations on the training model. Extensive experiments on five benchmarks across vision and text tasks demonstrate improved accuracy-efficiency trade-offs across models. Our method matches or exceeds full-data accuracy with over 2x training acceleration. Code will be released.

Xiangzhao Hao, Zefeng Zhang, Zhenyu Zhang et al.

Image degradation from blur, noise, compression, and poor illumination severely undermines multimodal understanding in real-world settings. Unified multimodal models that combine understanding and generation within a single architecture are a natural fit for this challenge, as their generative pathway can model the fine-grained visual structure that degradation destroys. Yet these models fail to leverage their own generative capacity on degraded inputs. We trace this disconnect to two compounding factors: existing training regimes never ask the model to invoke generation during reasoning, and the standard decode-reencode pathway does not support effective joint optimization. We present CLEAR, a framework that connects the two capabilities through three progressive steps: (1) supervised fine-tuning on a degradation-aware dataset to establish the generate-then-answer reasoning pattern; (2) a Latent Representation Bridge that replaces the decode-reencode detour with a direct, optimizable connection between generation and reasoning; (3) Interleaved GRPO, a reinforcement learning method that jointly optimizes text reasoning and visual generation under answer-correctness rewards. We construct MMD-Bench, covering three degradation severity levels across six standard multimodal benchmarks. Experiments show that CLEAR substantially improves robustness on degraded inputs while preserving clean-image performance. Our analysis further reveals that removing pixel-level reconstruction supervision leads to intermediate visual states with higher perceptual quality, suggesting that task-driven optimization and visual quality are naturally aligned.

Hongyan An, Kuan Zhu, Xin He et al.

Pedestrian attribute recognition (PAR) is a fundamental perception task in intelligent transportation and security. To tackle this fine-grained task, most existing methods focus on extracting regional features to enrich attribute information. However, a regional feature is typically used to predict a fixed set of pre-defined attributes in these methods, which limits the performance and practicality in two aspects: 1) Regional features may compromise fine-grained patterns unique to certain attributes in favor of capturing common characteristics shared across attributes. 2) Regional features cannot generalize to predict unseen attributes in the test time. In this paper, we propose the \textbf{F}ine-grained \textbf{O}ptimization with semanti\textbf{C} g\textbf{U}ided under\textbf{S}tanding (FOCUS) approach for PAR, which adaptively extracts fine-grained attribute-level features for each attribute individually, regardless of whether the attributes are seen or not during training. Specifically, we propose the Multi-Granularity Mix Tokens (MGMT) to capture latent features at varying levels of visual granularity, thereby enriching the diversity of the extracted information. Next, we introduce the Attribute-guided Visual Feature Extraction (AVFE) module, which leverages textual attributes as queries to retrieve their corresponding visual attribute features from the Mix Tokens using a cross-attention mechanism. To ensure that textual attributes focus on the appropriate Mix Tokens, we further incorporate a Region-Aware Contrastive Learning (RACL) method, encouraging attributes within the same region to share consistent attention maps. Extensive experiments on PA100K, PETA, and RAPv1 datasets demonstrate the effectiveness and strong generalization ability of our method.

Zhiyan Hou, Haiyun Guo, Haokai Ma et al.

Continual instruction tuning (CIT) requires multimodal large language models (MLLMs) to adapt to a stream of tasks without forgetting prior capabilities. A common strategy is to isolate updates by routing inputs to different LoRA experts. However, existing LoRA-based Mixture-of-Experts (MoE) methods often jointly update the router and experts in an indiscriminate way, causing the router's preferences to co-drift with experts' adaptation pathways and gradually deviate from early-stage input-expert specialization. We term this phenomenon Misaligned Co-drift, which blurs expert responsibilities and exacerbates forgetting.To address this, we introduce the pathway activation subspace (PASs), a LoRA-induced subspace that reflects which low-rank pathway directions an input activates in each expert, providing a capability-aligned coordinate system for routing and preservation. Based on PASs, we propose a fixed-capacity PASs-based MoE-LoRA method with two components: PAS-guided Reweighting, which calibrates routing using each expert's pathway activation signals, and PAS-aware Rank Stabilization, which selectively stabilizes rank directions important to previous tasks. Experiments on a CIT benchmark show that our approach consistently outperforms a range of conventional continual learning baselines and MoE-LoRA variants in both accuracy and anti-forgetting without adding parameters. Our code will be released upon acceptance.

Jinghan He, Junfeng Fang, Feng Xiong et al.

Self-play has enabled large language models to autonomously improve through self-generated challenges. However, existing self-play methods for vision-language models rely on passive interaction with static image collections, resulting in strong dependence on initial datasets and inefficient learning. Without the ability to actively seek visual data tailored to their evolving capabilities, agents waste computational effort on samples that are either trivial or beyond their current skill level. To address these limitations, we propose Active-Zero, a framework that shifts from passive interaction to active exploration of visual environments. Active-Zero employs three co-evolving agents: a Searcher that retrieves images from open-world repositories based on the model's capability frontier, a Questioner that synthesizes calibrated reasoning tasks, and a Solver refined through accuracy rewards. This closed loop enables self-scaffolding auto-curricula where the model autonomously constructs its learning trajectory. On Qwen2.5-VL-7B-Instruct across 12 benchmarks, Active-Zero achieves 53.97 average accuracy on reasoning tasks (5.7% improvement) and 59.77 on general understanding (3.9% improvement), consistently outperforming existing self-play baselines. These results highlight active exploration as a key ingredient for scalable and adaptive self-evolving vision-language systems.

Hongyu Guo, Kuan Zhu, Xiangzhao Hao et al.

Few-shot fine-grained visual classification (FGVC) aims to leverage limited data to enable models to discriminate subtly distinct categories. Recent works mostly finetuned the pre-trained visual language models to achieve performance gain, yet suffering from overfitting and weak generalization. To deal with this, we introduce UniFGVC, a universal training-free framework that reformulates few-shot FGVC as multimodal retrieval. First, we propose the Category-Discriminative Visual Captioner (CDV-Captioner) to exploit the open-world knowledge of multimodal large language models (MLLMs) to generate a structured text description that captures the fine-grained attribute features distinguishing closely related classes. CDV-Captioner uses chain-of-thought prompting and visually similar reference images to reduce hallucination and enhance discrimination of generated captions. Using it we can convert each image into an image-description pair, enabling more comprehensive feature representation, and construct the multimodal category templates using few-shot samples for the subsequent retrieval pipeline. Then, off-the-shelf vision and text encoders embed query and template pairs, and FGVC is accomplished by retrieving the nearest template in the joint space. UniFGVC ensures broad compatibility with diverse MLLMs and encoders, offering reliable generalization and adaptability across few-shot FGVC scenarios. Extensive experiments on 12 FGVC benchmarks demonstrate its consistent superiority over prior few-shot CLIP-based methods and even several fully-supervised MLLMs-based approaches.

Haiyun Guo, ZhiYan Hou, Yu Chen et al.

Multimodal large language models (MLLMs) require continual instruction tuning during their post-training phase to adapt to the dynamic real-world demands. However, the absence of rigorous and systematic benchmarks has hindered progress in this area. To bridge this gap, we introduce \textbf{MLLM-CTBench}, a dataset curating seven challenging tasks from six diverse domains with three contributions. First,to enable fine-grained analysis of continual learning ability, we introduce \textbf{multidimensional evaluation metrics}, which combines final answer accuracy with Chain-of-Thought (CoT) reasoning quality assessment through a carefully trained MLLM evaluator. Then, we conduct a \textbf{comprehensive evaluation of continual learning algorithms}, systematically assessing eight algorithms from four major categories to provide actionable insights for algorithm design and adoption. Finally ,we evaluate the efficacy of \textbf{Reinforcement Fine-tuning (RFT) versus Supervised Fine-tuning (SFT)} in maintaining model performance across sequential tasks during continual instruction tuning. Our experiments demonstrate that reasoning processes in MLLMs exhibit greater resilience than final outputs to forgetting during continual learning, aligning with cognitive theories of hierarchical forgetting. We further show that both model capability and task sequence significantly influence continual learning outcomes, with stronger baseline models exhibiting greater resistance to forgetting. Notably, properly regularized RFT emerges as a more robust approach than SFT for maintaining performance across tasks.One of the key contributing factors is KL-divergence regularization, without which RFT leads to even worse forgetting than SFT on old tasks though may perform better on new tasks.

Kuan Zhu, Haiyun Guo, Shiliang Zhang et al.

In person re-identification (re-ID), extracting part-level features from person images has been verified to be crucial to offer fine-grained information. Most of the existing CNN-based methods only locate the human parts coarsely, or rely on pretrained human parsing models and fail in locating the identifiable nonhuman parts (e.g., knapsack). In this article, we introduce an alignment scheme in transformer architecture for the first time and propose the auto-aligned transformer (AAformer) to automatically locate both the human parts and nonhuman ones at patch level. We introduce the "Part tokens ([PART]s)", which are learnable vectors, to extract part features in the transformer. A [PART] only interacts with a local subset of patches in self-attention and learns to be the part representation. To adaptively group the image patches into different subsets, we design the auto-alignment. Auto-alignment employs a fast variant of optimal transport (OT) algorithm to online cluster the patch embeddings into several groups with the [PART]s as their prototypes. AAformer integrates the part alignment into the self-attention and the output [PART]s can be directly used as part features for retrieval. Extensive experiments validate the effectiveness of [PART]s and the superiority of AAformer over various state-of-the-art methods.

Zhiwei Liu, Xiangyu Zhu, Guosheng Hu et al.

Recently, deep learning based facial landmark detection has achieved great success. Despite this, we notice that the semantic ambiguity greatly degrades the detection performance. Specifically, the semantic ambiguity means that some landmarks (e.g. those evenly distributed along the face contour) do not have clear and accurate definition, causing inconsistent annotations by annotators. Accordingly, these inconsistent annotations, which are usually provided by public databases, commonly work as the ground-truth to supervise network training, leading to the degraded accuracy. To our knowledge, little research has investigated this problem. In this paper, we propose a novel probabilistic model which introduces a latent variable, i.e. the 'real' ground-truth which is semantically consistent, to optimize. This framework couples two parts (1) training landmark detection CNN and (2) searching the 'real' ground-truth. These two parts are alternatively optimized: the searched 'real' ground-truth supervises the CNN training; and the trained CNN assists the searching of 'real' ground-truth. In addition, to recover the unconfidently predicted landmarks due to occlusion and low quality, we propose a global heatmap correction unit (GHCU) to correct outliers by considering the global face shape as a constraint. Extensive experiments on both image-based (300W and AFLW) and video-based (300-VW) databases demonstrate that our method effectively improves the landmark detection accuracy and achieves the state of the art performance.