Yuyang Liu

Ya-nan Guan, Shaonan Zhang, Hang Guo et al.

In this paper, we present a comprehensive overview of the NTIRE 2026 3rd Restore Any Image Model (RAIM) challenge, with a specific focus on Track 3: AI Flash Portrait. Despite significant advancements in deep learning for image restoration, existing models still encounter substantial challenges in real-world low-light portrait scenarios. Specifically, they struggle to achieve an optimal balance among noise suppression, detail preservation, and faithful illumination and color reproduction. To bridge this gap, this challenge aims to establish a novel benchmark for real-world low-light portrait restoration. We comprehensively evaluate the proposed algorithms utilizing a hybrid evaluation system that integrates objective quantitative metrics with rigorous subjective assessment protocols. For this competition, we provide a dataset containing 800 groups of real-captured low-light portrait data. Each group consists of a 1K-resolution low-light input image, a 1K ground truth (GT), and a 1K person mask. This challenge has garnered widespread attention from both academia and industry, attracting over 100 participating teams and receiving more than 3,000 valid submissions. This report details the motivation behind the challenge, the dataset construction process, the evaluation metrics, and the various phases of the competition. The released dataset and baseline code for this track are publicly available from the same \href{https://github.com/zsn1434/AI_Flash-BaseLine/tree/main}{GitHub repository}, and the official challenge webpage is hosted on \href{https://www.codabench.org/competitions/12885/}{CodaBench}.

Dipam Goswami, Yuyang Liu, Bartłomiej Twardowski et al.

Exemplar-free class-incremental learning (CIL) poses several challenges since it prohibits the rehearsal of data from previous tasks and thus suffers from catastrophic forgetting. Recent approaches to incrementally learning the classifier by freezing the feature extractor after the first task have gained much attention. In this paper, we explore prototypical networks for CIL, which generate new class prototypes using the frozen feature extractor and classify the features based on the Euclidean distance to the prototypes. In an analysis of the feature distributions of classes, we show that classification based on Euclidean metrics is successful for jointly trained features. However, when learning from non-stationary data, we observe that the Euclidean metric is suboptimal and that feature distributions are heterogeneous. To address this challenge, we revisit the anisotropic Mahalanobis distance for CIL. In addition, we empirically show that modeling the feature covariance relations is better than previous attempts at sampling features from normal distributions and training a linear classifier. Unlike existing methods, our approach generalizes to both many- and few-shot CIL settings, as well as to domain-incremental settings. Interestingly, without updating the backbone network, our method obtains state-of-the-art results on several standard continual learning benchmarks. Code is available at https://github.com/dipamgoswami/FeCAM.

Yangyuxuan Kang, Yuyang Liu, Anbang Yao et al.

Existing lifting networks for regressing 3D human poses from 2D single-view poses are typically constructed with linear layers based on graph-structured representation learning. In sharp contrast to them, this paper presents Grid Convolution (GridConv), mimicking the wisdom of regular convolution operations in image space. GridConv is based on a novel Semantic Grid Transformation (SGT) which leverages a binary assignment matrix to map the irregular graph-structured human pose onto a regular weave-like grid pose representation joint by joint, enabling layer-wise feature learning with GridConv operations. We provide two ways to implement SGT, including handcrafted and learnable designs. Surprisingly, both designs turn out to achieve promising results and the learnable one is better, demonstrating the great potential of this new lifting representation learning formulation. To improve the ability of GridConv to encode contextual cues, we introduce an attention module over the convolutional kernel, making grid convolution operations input-dependent, spatial-aware and grid-specific. We show that our fully convolutional grid lifting network outperforms state-of-the-art methods with noticeable margins under (1) conventional evaluation on Human3.6M and (2) cross-evaluation on MPI-INF-3DHP. Code is available at https://github.com/OSVAI/GridConv

Lai-Man Po, Yuyang Liu, Haoxuan Wu et al.

This paper introduces Standard Basis LoRA (SBoRA), a novel parameter-efficient fine-tuning approach for Large Language Models that builds upon the pioneering works of Low-Rank Adaptation (LoRA) and Orthogonal Adaptation. SBoRA reduces the number of trainable parameters by half or doubles the rank with the similar number of trainable parameters as LoRA, while improving learning performance. By utilizing orthogonal standard basis vectors to initialize one of the low-rank matrices (either $\mathbf{A}$ or $\mathbf{B}$), SBoRA facilitates regional weight updates and memory-efficient fine-tuning. This results in two variants, SBoRA-FA and SBoRA-FB, where only one of the matrices is updated, leading to a sparse update matrix $\mathrmΔ \mathbf{W}$ with predominantly zero rows or columns. Consequently, most of the fine-tuned model's weights $(\mathbf{W}_0+\mathrmΔ \mathbf{W})$ remain unchanged from the pre-trained weights, akin to the modular organization of the human brain, which efficiently adapts to new tasks. Our empirical results demonstrate the superiority of SBoRA-FA over LoRA in various fine-tuning tasks, including commonsense reasoning and arithmetic reasoning. Furthermore, we evaluate the effectiveness of QSBoRA on quantized LLaMA models of varying scales, highlighting its potential for efficient adaptation to new tasks. Code is available at https://github.com/cityuhkai/SBoRA

Chenhao Bai, Liqin Lu, Kaijun Wang et al.

Scaling robust robot policies requires more than broader randomization, because physical-domain experience must remain organized and learnable throughout training. We study when a policy can benefit from harder physics and identify recoverability as a central constraint in on-policy physical-domain scaling. In on-policy training, new dynamics are useful only insofar as they remain close enough to the current policy to generate corrective on-policy data, rather than collapsing rollouts into unrecoverable failures. Using quadruped locomotion as a physically demanding benchmark for embodied generalization, we introduce HORIZON, a checkpointed frontier curriculum that expands physical domains only within the current policy's recoverable boundary. HORIZON uses rollback and boundary refinement to govern each expansion step, turning fixed randomization into a continual process of physical-domain growth. Experiments reveal three regularities of physical-domain expansion. First, direct domain widening is uneven across physical axes and often unlearnable without staged ordering. Second, domain composition is non-monotonic, and adding more domains beyond a compact core can dilute recoverable joint samples and reduce overall robustness. Third, offline distillation of isolated experts cannot substitute for the joint interaction generated by on-policy curriculum. Together, these results frame physical-domain generalization as a continual growth problem for embodied control, with recoverability as the organizing principle for on-policy expansion.

Wei Cong, Yang Cong, Gan Sun et al.

Continual learning algorithms which keep the parameters of new tasks close to that of previous tasks, are popular in preventing catastrophic forgetting in sequential task learning settings. However, 1) the performance for the new continual learner will be degraded without distinguishing the contributions of previously learned tasks; 2) the computational cost will be greatly increased with the number of tasks, since most existing algorithms need to regularize all previous tasks when learning new tasks. To address the above challenges, we propose a self-paced Weight Consolidation (spWC) framework to attain robust continual learning via evaluating the discriminative contributions of previous tasks. To be specific, we develop a self-paced regularization to reflect the priorities of past tasks via measuring difficulty based on key performance indicator (i.e., accuracy). When encountering a new task, all previous tasks are sorted from "difficult" to "easy" based on the priorities. Then the parameters of the new continual learner will be learned via selectively maintaining the knowledge amongst more difficult past tasks, which could well overcome catastrophic forgetting with less computational cost. We adopt an alternative convex search to iteratively update the model parameters and priority weights in the bi-convex formulation. The proposed spWC framework is plug-and-play, which is applicable to most continual learning algorithms (e.g., EWC, MAS and RCIL) in different directions (e.g., classification and segmentation). Experimental results on several public benchmark datasets demonstrate that our proposed framework can effectively improve performance when compared with other popular continual learning algorithms.

Yipu Wang, Yuheng Ji, Yuyang Liu et al.

Cross-view correspondence is a fundamental capability for spatial understanding and embodied AI. However, it is still far from being realized in Vision-Language Models (VLMs), especially in achieving precise point-level correspondence, which is crucial for precise affordance interaction. So we propose the Cross-View Point Correspondence (CVPC) task and CrossPoint-Bench, a comprehensive benchmark with hierarchical design, inspired by the human cognitive process of "perceive", "reason", and "correspond". Our evaluation shows the state-of-the-art models (e.g., Gemini-2.5-Pro) still fall far behind humans, with a gap of over 54.65% in overall accuracy, exposing a challenge in transitioning from coarse-grained judgement to fine-grained coordinate prediction. To address this problem, we construct CrossPoint-378K, a dataset with 378K question-answering pairs across 900 scenes, focused on actionable affordance regions that better reflect real-world manipulation and interaction scenarios. Furthermore, we propose CroPond that trained on the CrossPoint-378K dataset. Our CroPond achieves state-of-the-art performance on CrossPoint-Bench, surpassing Gemini-2.5-Pro by 39.7% accuracy, which offers a foundation for advancing future work on cross-view correspondence. The benchmark, dataset, and model are publicly available at https://github.com/WangYipu2002/CrossPoint.

Yufei Huang, Siyuan Li, Jin Su et al.

Protein structure-based property prediction has emerged as a promising approach for various biological tasks, such as protein function prediction and sub-cellular location estimation. The existing methods highly rely on experimental protein structure data and fail in scenarios where these data are unavailable. Predicted protein structures from AI tools (e.g., AlphaFold2) were utilized as alternatives. However, we observed that current practices, which simply employ accurately predicted structures during inference, suffer from notable degradation in prediction accuracy. While similar phenomena have been extensively studied in general fields (e.g., Computer Vision) as model robustness, their impact on protein property prediction remains unexplored. In this paper, we first investigate the reason behind the performance decrease when utilizing predicted structures, attributing it to the structure embedding bias from the perspective of structure representation learning. To study this problem, we identify a Protein 3D Graph Structure Learning Problem for Robust Protein Property Prediction (PGSL-RP3), collect benchmark datasets, and present a protein Structure embedding Alignment Optimization framework (SAO) to mitigate the problem of structure embedding bias between the predicted and experimental protein structures. Extensive experiments have shown that our framework is model-agnostic and effective in improving the property prediction of both predicted structures and experimental structures. The benchmark datasets and codes will be released to benefit the community.

Ziheng Qin, Yuheng Ji, Renshuai Tao et al.

The pursuit of a universal AI-generated image (AIGI) detector often relies on aggregating data from numerous generators to improve generalization. However, this paper identifies a paradoxical phenomenon we term the Benefit then Conflict dilemma, where detector performance stagnates and eventually degrades as source diversity expands. Our systematic analysis, diagnoses this failure by identifying two core issues: severe data-level heterogeneity, which causes the feature distributions of real and synthetic images to increasingly overlap, and a critical model-level bottleneck from fixed, pretrained encoders that cannot adapt to the rising complexity. To address these challenges, we propose Generator-Aware Prototype Learning (GAPL), a framework that constrain representation with a structured learning paradigm. GAPL learns a compact set of canonical forgery prototypes to create a unified, low-variance feature space, effectively countering data heterogeneity.To resolve the model bottleneck, it employs a two-stage training scheme with Low-Rank Adaptation, enhancing its discriminative power while preserving valuable pretrained knowledge. This approach establishes a more robust and generalizable decision boundary. Through extensive experiments, we demonstrate that GAPL achieves state-of-the-art performance, showing superior detection accuracy across a wide variety of GAN and diffusion-based generators. Code is available at https://github.com/UltraCapture/GAPL

Linhan Wu, Chenxi Wang, Chuhan Yang et al.

Inverse design of materials has significantly advanced target-driven formulation optimization, yet existing materials machine learning benchmarks remain limited to forward property prediction, failing to systematically evaluate inverse optimization and generation algorithms, a critical gap that hinders the progress of target-driven materials design. To address this limitation, we propose MatFormBench, a novel benchmarking ecosystem tailored to evaluate and guide generative strategies for target-driven formulation. MatFormBench integrates a physics-driven formulation generation scheme to generate synthetic samples that faithfully emulate realistic materials structure-property response relationships, complemented by five escalating difficulty levels to quantify the complexity of these relationships. To rigorously assess algorithm performance, we further propose MatFormScore, a multi-dimensional metric that comprehensively quantifies performance across five critical axes: target success, search efficiency, exploratory capacity, robustness, and stability. We validate MatFormBench by evaluating 39 diverse inverse design algorithms, covering classical surrogate-assisted black-box search, state-of-the-art deep generative models, and increasingly popular Large Language Model (LLM)-based recommendation strategies. Across 1170 standardized algorithm-task evaluations, diffusion-based models demonstrate the strongest overall performance, while Variational Autoencoder (VAE)-based and Genetic Algorithm (GA)-based methods exhibit distinct advantages in specific scenarios. By establishing a unified evaluation standard for target-driven materials formulation, MatFormBench enables reproducible benchmarking, principled algorithm comparison, and diagnostic analysis of inverse design strategies, providing a foundational tool for advancing materials inverse design.

Yuyang Liu, Jingya Wang, Liuzhenghao Lv et al.

Large language models (LLMs) have demonstrated significant reasoning capabilities in scientific discovery but struggle to bridge the gap to physical execution in wet-labs. In these irreversible environments, probabilistic hallucinations are not merely incorrect, but also cause equipment damage or experimental failure. To address this, we propose \textbf{BioProAgent}, a neuro-symbolic framework that anchors probabilistic planning in a deterministic Finite State Machine (FSM). We introduce a State-Augmented Planning mechanism that enforces a rigorous \textit{Design-Verify-Rectify} workflow, ensuring hardware compliance before execution. Furthermore, we address the context bottleneck inherent in complex device schemas by \textit{Semantic Symbol Grounding}, reducing token consumption by $\sim$6$\times$ through symbolic abstraction. In the extended BioProBench benchmark, BioProAgent achieves 95.6\% physical compliance (compared to 21.0\% for ReAct), demonstrating that neuro-symbolic constraints are essential for reliable autonomy in irreversible physical environments. \footnote{Code at https://github.com/YuyangSunshine/bioproagent and project at https://yuyangsunshine.github.io/BioPro-Project/}

Zhihui Ke, Yuyang Liu, Xiaobo Zhou et al.

Streaming free-viewpoint video~(FVV) in real-time still faces significant challenges, particularly in training, rendering, and transmission efficiency. Harnessing superior performance of 3D Gaussian Splatting~(3DGS), recent 3DGS-based FVV methods have achieved notable breakthroughs in both training and rendering. However, the storage requirements of these methods can reach up to $10$MB per frame, making stream FVV in real-time impossible. To address this problem, we propose a novel FVV representation, dubbed StreamSTGS, designed for real-time streaming. StreamSTGS represents a dynamic scene using canonical 3D Gaussians, temporal features, and a deformation field. For high compression efficiency, we encode canonical Gaussian attributes as 2D images and temporal features as a video. This design not only enables real-time streaming, but also inherently supports adaptive bitrate control based on network condition without any extra training. Moreover, we propose a sliding window scheme to aggregate adjacent temporal features to learn local motions, and then introduce a transformer-guided auxiliary training module to learn global motions. On diverse FVV benchmarks, StreamSTGS demonstrates competitive performance on all metrics compared to state-of-the-art methods. Notably, StreamSTGS increases the PSNR by an average of $1$dB while reducing the average frame size to just $170$KB. The code is publicly available on https://github.com/kkkzh/StreamSTGS.

Yuyang Liu, Weijun Dong, Yingdong Hu et al.

Humans often acquire new skills through observation and imitation. For robotic agents, learning from the plethora of unlabeled video demonstration data available on the Internet necessitates imitating the expert without access to its action, presenting a challenge known as Imitation Learning from Observations (ILfO). A common approach to tackle ILfO problems is to convert them into inverse reinforcement learning problems, utilizing a proxy reward computed from the agent's and the expert's observations. Nonetheless, we identify that tasks characterized by a progress dependency property pose significant challenges for such approaches; in these tasks, the agent needs to initially learn the expert's preceding behaviors before mastering the subsequent ones. Our investigation reveals that the main cause is that the reward signals assigned to later steps hinder the learning of initial behaviors. To address this challenge, we present a novel ILfO framework that enables the agent to master earlier behaviors before advancing to later ones. We introduce an Automatic Discount Scheduling (ADS) mechanism that adaptively alters the discount factor in reinforcement learning during the training phase, prioritizing earlier rewards initially and gradually engaging later rewards only when the earlier behaviors have been mastered. Our experiments, conducted on nine Meta-World tasks, demonstrate that our method significantly outperforms state-of-the-art methods across all tasks, including those that are unsolvable by them.

Marwa Dhiaf, Mohamed Ali Souibgui, Kai Wang et al.

Self-supervised learning has recently emerged as a strong alternative in document analysis. These approaches are now capable of learning high-quality image representations and overcoming the limitations of supervised methods, which require a large amount of labeled data. However, these methods are unable to capture new knowledge in an incremental fashion, where data is presented to the model sequentially, which is closer to the realistic scenario. In this paper, we explore the potential of continual self-supervised learning to alleviate the catastrophic forgetting problem in handwritten text recognition, as an example of sequence recognition. Our method consists in adding intermediate layers called adapters for each task, and efficiently distilling knowledge from the previous model while learning the current task. Our proposed framework is efficient in both computation and memory complexity. To demonstrate its effectiveness, we evaluate our method by transferring the learned model to diverse text recognition downstream tasks, including Latin and non-Latin scripts. As far as we know, this is the first application of continual self-supervised learning for handwritten text recognition. We attain state-of-the-art performance on English, Italian and Russian scripts, whilst adding only a few parameters per task. The code and trained models will be publicly available.

Wei Cong, Yang Cong, Yuyang Liu et al.

Incremental semantic segmentation endeavors to segment newly encountered classes while maintaining knowledge of old classes. However, existing methods either 1) lack guidance from class-specific knowledge (i.e., old class prototypes), leading to a bias towards new classes, or 2) constrain class-shared knowledge (i.e., old model weights) excessively without discrimination, resulting in a preference for old classes. In this paper, to trade off model performance, we propose the Class-specific and Class-shared Knowledge (Cs2K) guidance for incremental semantic segmentation. Specifically, from the class-specific knowledge aspect, we design a prototype-guided pseudo labeling that exploits feature proximity from prototypes to correct pseudo labels, thereby overcoming catastrophic forgetting. Meanwhile, we develop a prototype-guided class adaptation that aligns class distribution across datasets via learning old augmented prototypes. Moreover, from the class-shared knowledge aspect, we propose a weight-guided selective consolidation to strengthen old memory while maintaining new memory by integrating old and new model weights based on weight importance relative to old classes. Experiments on public datasets demonstrate that our proposed Cs2K significantly improves segmentation performance and is plug-and-play.

Damian Sójka, Yuyang Liu, Dipam Goswami et al.

The goal of the challenge is to develop a test-time adaptation (TTA) method, which could adapt the model to gradually changing domains in video sequences for semantic segmentation task. It is based on a synthetic driving video dataset - SHIFT. The source model is trained on images taken during daytime in clear weather. Domain changes at test-time are mainly caused by varying weather conditions and times of day. The TTA methods are evaluated in each image sequence (video) separately, meaning the model is reset to the source model state before the next sequence. Images come one by one and a prediction has to be made at the arrival of each frame. Each sequence is composed of 401 images and starts with the source domain, then gradually drifts to a different one (changing weather or time of day) until the middle of the sequence. In the second half of the sequence, the domain gradually shifts back to the source one. Ground truth data is available only for the validation split of the SHIFT dataset, in which there are only six sequences that start and end with the source domain. We conduct an analysis specifically on those sequences. Ground truth data for test split, on which the developed TTA methods are evaluated for leader board ranking, are not publicly available. The proposed solution secured a 3rd place in a challenge and received an innovation award. Contrary to the solutions that scored better, we did not use any external pretrained models or specialized data augmentations, to keep the solutions as general as possible. We have focused on analyzing the distributional shift and developing a method that could adapt to changing data dynamics and generalize across different scenarios.

Shuo Yang, Jinda Lu, Kexin Huang et al.

Reinforcement Learning with Verifiable Rewards (RLVR) has become a promising paradigm for scaling reasoning capabilities of Large Language Models (LLMs). However, the sparsity of binary verifier rewards often leads to low efficiency and optimization instability. To stabilize training, existing methods typically impose token-level constraints relative to a reference policy. We identify that such constraints penalize deviations indiscriminately; this can flip verifier-determined direction when the policy attempts to outperform the reference, thereby suppressing gains. To resolve this, we propose One-Way Policy Optimization (OWPO), a method based on the principle of decoupling optimization direction from update magnitude. In OWPO, the verifier dictates the update direction, while the reference policy serves only to adjust the magnitude. Specifically, OWPO applies asymmetric reweighting: it performs Accelerated Alignment for inferior deviations (where the policy lags behind the reference) and Gain Locking for superior deviations (where the policy surpasses the reference). Furthermore, by incorporating iterative reference updates, OWPO creates a ``Ratchet Effect'' that continuously consolidates gains. Experimental results demonstrate that OWPO outperforms strong baselines, including DAPO, OPD, and MOPD, breaking the bottleneck of fixed priors to enable continuous self-evolution without reliance on external reference models.

Shuo Yang, Jinda Lu, Chiyu Ma et al.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a central paradigm for scaling LLM reasoning, yet its optimization often suffers from training instability and suboptimal convergence. Through a systematic dissection of clipping-based GRPO-style objectives, we identify the rigid clipping decision induced by hard clipping as a key practical bottleneck in the studied RLVR setups. Specifically, our analysis suggests that informative signals can lie in the near-boundary region just beyond the clipping threshold, and are therefore discarded by the standard hard-clipping rule. Notably, once this bottleneck is precisely identified, even simple stochastic perturbations at the boundary can recover meaningful performance gains. Building on this finding, we propose Near-boundary Stochastic Rescue (NSR), a minimal, plug-and-play modification that stochastically retains these slightly out-of-bound tokens to recover lost signals. While NSR, via stochastic sampling, can be interpreted as inducing an implicit gradient decay in expectation, our ablations reveal that its stochastic, boundary-local rescue mechanism is consistently more effective than deterministic gradient decay. Validated by extensive experiments across model sizes from 7B to 30B and both dense and MoE architectures, as a plug-and-play solution, NSR substantially improves training stability and delivers consistent gains over strong baselines such as DAPO and GSPO.

Yuheng Ji, Yuyang Liu, Huajie Tan et al.

Current robotic evaluation is still largely dominated by binary success rates, which collapse rich execution processes into a single outcome and obscure critical qualities such as progress, efficiency, and stability. To address this limitation, we propose PRM-as-a-Judge, a dense evaluation paradigm that leverages Process Reward Models (PRMs) to audit policy execution directly from trajectory videos by estimating task progress from observation sequences. Central to this paradigm is the OPD (Outcome-Process-Diagnosis) metric system, which explicitly formalizes execution quality via a task-aligned progress potential. We characterize dense robotic evaluation through two axiomatic properties: macro-consistency, which requires additive and path-consistent aggregation, and micro-resolution, which requires sensitivity to fine-grained physical evolution. Under this formulation, potential-based PRM judges provide a natural instantiation of dense evaluation, with macro-consistency following directly from the induced scalar potential. We empirically validate the micro-resolution property using RoboPulse, a diagnostic benchmark specifically designed for probing micro-scale progress discrimination, where several trajectory-trained PRM judges outperform discriminative similarity-based methods and general-purpose foundation-model judges. Finally, leveraging PRM-as-a-Judge and the OPD metric system, we conduct a structured audit of mainstream policy paradigms across long-horizon tasks, revealing behavioral signatures and failure modes that are invisible to outcome-only metrics.

Shuo Yang, Qihui Zhang, Yuyang Liu et al.

Large language models (LLMs) are vulnerable to safety risks during fine-tuning, where small amounts of malicious or harmless data can compromise safeguards. In this paper, building on the concept of alignment direction -- defined by the weight difference between aligned and unaligned models -- we observe that perturbations along this direction preserve model safety. In contrast, perturbations along directions orthogonal to this alignment are strongly linked to harmful direction perturbations, rapidly degrading safety and framing the parameter space as a narrow safety basin. Based on this insight, we propose a methodology for safety fine-tuning called AsFT (Anchoring Safety in Fine-Tuning), which integrates a regularization term into the training objective. This term uses the alignment direction as an anchor to suppress updates in harmful directions, ensuring that fine-tuning is constrained within the narrow safety basin. Extensive experiments on multiple datasets show that AsFT outperforms Safe LoRA, reducing harmful behavior by 7.60 percent, improving model performance by 3.44 percent, and maintaining robust performance across various experimental settings. Code is available at https://github.com/PKU-YuanGroup/AsFT

Yuyang Liu

Combinatorial optimization problems arise in logistics, scheduling, and resource allocation, yet existing approaches face a fundamental trade-off among generality, performance, and usability. We present cuGenOpt, a GPU-accelerated general-purpose metaheuristic framework that addresses all three dimensions simultaneously. At the engine level, cuGenOpt adopts a "one block evolves one solution" CUDA architecture with a unified encoding abstraction (permutation, binary, integer), a two-level adaptive operator selection mechanism, and hardware-aware resource management. At the extensibility level, a user-defined operator registration interface allows domain experts to inject problem-specific CUDA search operators. At the usability level, a JIT compilation pipeline exposes the framework as a pure-Python API, and an LLM-based modeling assistant converts natural-language problem descriptions into executable solver code. Experiments across five thematic suites on three GPU architectures (T4, V100, A800) show that cuGenOpt outperforms general MIP solvers by orders of magnitude, achieves competitive quality against specialized solvers on instances up to n=150, and attains 4.73% gap on TSP-442 within 30s. Twelve problem types spanning five encoding variants are solved to optimality. Framework-level optimizations cumulatively reduce pcb442 gap from 36% to 4.73% and boost VRPTW throughput by 75-81%. Code: https://github.com/L-yang-yang/cugenopt

Yajie Bao, Yuyang Liu

Linear discriminant analysis (LDA) is an important classification tool in statistics and machine learning. This paper investigates the varying coefficient LDA model for dynamic data, with Bayes' discriminant direction being a function of some exposure variable to address the heterogeneity. We propose a new least-square estimation method based on the B-spline approximation. The data-driven discriminant procedure is more computationally efficient than the dynamic linear programming rule \citep{jiang2020dynamic}. We also establish the convergence rates for the corresponding estimation error bound and the excess misclassification risk. The estimation error in $L_2$ distance is optimal for the low-dimensional regime and is near optimal for the high-dimensional regime. Numerical experiments on synthetic data and real data both corroborate the superiority of our proposed classification method.

Yitian Zheng, Zhangchen Ye, Weijun Dong et al.

Recent advances in hierarchical robot systems leverage a high-level planner to propose task plans and a low-level policy to generate robot actions. This design allows training the planner on action-free or even non-robot data sources (e.g., videos), providing transferable high-level guidance. Nevertheless, grounding these high-level plans into executable actions remains challenging, especially with the limited availability of high-quality robot data. To this end, we propose to improve the low-level policy through online interactions. Specifically, our approach collects online rollouts, retrospectively annotates the corresponding high-level goals from achieved outcomes, and aggregates these hindsight-relabeled experiences to update a goal-conditioned imitation policy. Our method, Hindsight Flow-conditioned Online Imitation (HinFlow), instantiates this idea with 2D point flows as the high-level planner. Across diverse manipulation tasks in both simulation and physical world, our method achieves more than $2\times$ performance improvement over the base policy, significantly outperforming the existing methods. Moreover, our framework enables policy acquisition from planners trained on cross-embodiment video data, demonstrating its potential for scalable and transferable robot learning.

Qiuhe Hong, Yuyang Liu, Shuo Yang et al.

Vision-Language Models in Continual Learning (VLM-CL) aim to continuously adapt to new multimodal tasks while retaining prior knowledge. The emerging paradigm that couples Multimodal Large Language Models (MLLMs) with Reinforcement Learning with Verifiable Rewards (RLVR) calls for a new pattern to guide continual adaptation. Advances in reasoning capability now make it feasible to impose constraints at the reasoning level. We formalize portability, a sample-level measure of how reusable the previous policy's behavior is on a new task, and empirically show that reasoning-level signals remain reliable on out-of-distribution samples while answer-level signals do not. We instantiate this as Reasoning Portability (RP) and propose Reasoning-based Dynamic Balance Continual Learning (RDB-CL), which modulates the per-sample Kullback-Leibler regularization in RLVR according to RP: a tight anchor preserves reusable reasoning on high-RP samples, while a relaxed anchor on low-RP samples permits exploration of new reasoning pathways. Experiments show that RDB-CL consistently outperforms baselines, improving Last accuracy by +12.0% over the vanilla RLVR baseline.

Yuheng Ji, Yipu Wang, Yuyang Liu et al.

Visual transformation reasoning (VTR) is a vital cognitive capability that empowers intelligent agents to understand dynamic scenes, model causal relationships, and predict future states, and thereby guiding actions and laying the foundation for advanced intelligent systems. However, existing benchmarks suffer from a sim-to-real gap, limited task complexity, and incomplete reasoning coverage, limiting their practical use in real-world scenarios. To address these limitations, we introduce VisualTrans, the first comprehensive benchmark specifically designed for VTR in real-world human-object interaction scenarios. VisualTrans encompasses 12 semantically diverse manipulation tasks and systematically evaluates three essential reasoning dimensions - spatial, procedural, and quantitative - through 6 well-defined subtask types. The benchmark features 472 high-quality question-answer pairs in various formats, including multiple-choice, open-ended counting, and target enumeration. We introduce a scalable data construction pipeline built upon first-person manipulation videos, which integrates task selection, image pair extraction, automated metadata annotation with large multimodal models, and structured question generation. Human verification ensures the final benchmark is both high-quality and interpretable. Evaluations of various state-of-the-art vision-language models show strong performance in static spatial tasks. However, they reveal notable shortcomings in dynamic, multi-step reasoning scenarios, particularly in areas like intermediate state recognition and transformation sequence planning. These findings highlight fundamental weaknesses in temporal modeling and causal reasoning, providing clear directions for future research aimed at developing more capable and generalizable VTR systems. The dataset and code are available at https://github.com/WangYipu2002/VisualTrans.

Yuyang Liu, Liuzhenghao Lv, Xiancheng Zhang et al.

Biological protocols are fundamental to reproducibility and safety in life science research. While large language models (LLMs) perform well on general tasks, their systematic evaluation on these highly specialized, accuracy-critical, and inherently procedural texts remains limited. In this work, we present BioProBench, the first large-scale, multi-task benchmark for biological protocol understanding and reasoning. While there are several benchmark tasks involving protocol question answering, BioProBench provides a comprehensive suite of five core tasks: Protocol Question Answering, Step Ordering, Error Correction, Protocol Generation, and Protocol Reasoning, enabling a holistic evaluation of LLMs on procedural biological texts. Built upon 27K original protocols, it yields nearly 556K high-quality structured instances. We evaluate 12 mainstream open/closed-source LLMs. Experimental results reveal that some models perform well on basic understanding tasks (e.g., \sim70% PQA-Acc., >64% ERR F1), but struggle significantly with deep reasoning and structured generation tasks like ordering and generation. Furthermore, model comparisons show diverse performance: certain open-source models approach closed-source levels on some tasks, yet bio-specific small models lag behind general LLMs, indicating limitations on complex procedural content. Overall, BioProBench, through its task design and experimental findings, systematically reveals the fundamental challenges for current LLMs in procedural knowledge understanding, deep adaptability to specific domains, reliability of structured reasoning, and handling of sophisticated precision and safety constraints, providing key directions for future AI in the field of scientific experiment automation. The code and data are available at: https://github.com/YuyangSunshine/bioprotocolbench and https://huggingface.co/datasets/BioProBench/BioProBench.

Yuyang Liu, Xinyuan Shi, Xiaondan Liang

The success of Large Language Models (LLMs) has significantly propelled the research of video understanding. To harvest the benefits of well-trained expert models (i.e., tools), video LLMs prioritize the exploration of tool usage capabilities. Existing methods either prompt closed-source LLMs or employ the instruction tuning paradigm for tool-use fine-tuning. These methods, however, assume an established repository of fixed tools and struggle to generalize to real-world environments where tool data is perpetually evolving and streaming in. To this end, we propose to enhance open-source video LLMs with COntinuaL Tool usage (termed COLT), which automatically acquires tool-use ability in a successive tool stream without suffering 'catastrophic forgetting' of the past learned tools. Specifically, our COLT incorporates a learnable tool codebook as a tool-specific memory system. Then relevant tools are dynamically selected based on the similarity between user instruction and tool features within the codebook. To unleash the tool usage potential of video LLMs, we collect a video-centric tool-use instruction tuning dataset VideoToolBench. Extensive experiments on both previous video LLM benchmarks and the tool-use-specific VideoToolBench dataset demonstrate the state-of-the-art performance of our proposed COLT.

Yuyang Liu, Qiuhe Hong, Linlan Huang et al.

Vision-language models (VLMs) have achieved impressive performance across diverse multimodal tasks by leveraging large-scale pre-training. However, enabling them to learn continually from non-stationary data remains a major challenge, as their cross-modal alignment and generalization capabilities are particularly vulnerable to catastrophic forgetting. Unlike traditional unimodal continual learning (CL), VLMs face unique challenges such as cross-modal feature drift, parameter interference due to shared architectures, and zero-shot capability erosion. This survey offers the first focused and systematic review of continual learning for VLMs (VLM-CL). We begin by identifying the three core failure modes that degrade performance in VLM-CL. Based on these, we propose a challenge-driven taxonomy that maps solutions to their target problems: (1) \textit{Multi-Modal Replay Strategies} address cross-modal drift through explicit or implicit memory mechanisms; (2) \textit{Cross-Modal Regularization} preserves modality alignment during updates; and (3) \textit{Parameter-Efficient Adaptation} mitigates parameter interference with modular or low-rank updates. We further analyze current evaluation protocols, datasets, and metrics, highlighting the need for better benchmarks that capture VLM-specific forgetting and compositional generalization. Finally, we outline open problems and future directions, including continual pre-training and compositional zero-shot learning. This survey aims to serve as a comprehensive and diagnostic reference for researchers developing lifelong vision-language systems. All resources are available at: https://github.com/YuyangSunshine/Awesome-Continual-learning-of-Vision-Language-Models.

Zhiqing Zhang, Guojia Fan, Tianyong Liu et al.

Medical image segmentation is critical for diagnosing and treating spinal disorders. However, the presence of high noise, ambiguity, and uncertainty makes this task highly challenging. Factors such as unclear anatomical boundaries, inter-class similarities, and irrational annotations contribute to this challenge. Achieving both accurate and diverse segmentation templates is essential to support radiologists in clinical practice. In recent years, denoising diffusion probabilistic modeling (DDPM) has emerged as a prominent research topic in computer vision. It has demonstrated effectiveness in various vision tasks, including image deblurring, super-resolution, anomaly detection, and even semantic representation generation at the pixel level. Despite the robustness of existing diffusion models in visual generation tasks, they still struggle with discrete masks and their various effects. To address the need for accurate and diverse spine medical image segmentation templates, we propose an end-to-end framework called VerseDiff-UNet, which leverages the denoising diffusion probabilistic model (DDPM). Our approach integrates the diffusion model into a standard U-shaped architecture. At each step, we combine the noise-added image with the labeled mask to guide the diffusion direction accurately towards the target region. Furthermore, to capture specific anatomical a priori information in medical images, we incorporate a shape a priori module. This module efficiently extracts structural semantic information from the input spine images. We evaluate our method on a single dataset of spine images acquired through X-ray imaging. Our results demonstrate that VerseDiff-UNet significantly outperforms other state-of-the-art methods in terms of accuracy while preserving the natural features and variations of anatomy.

Meng Cao, Yuyang Liu, Yingfei Liu et al.

Instruction tuning constitutes a prevalent technique for tailoring Large Vision Language Models (LVLMs) to meet individual task requirements. To date, most of the existing approaches are confined to single-task adaptation, whereas the requirements in real-world scenarios are inherently varied and continually evolving. Thus an ideal LVLM should sustain continual instruction tuning in the face of stream-task distributions (i.e., different domains, emerging capabilities, and new datasets) while minimizing the forgetting of previously acquired knowledge. To achieve this, we propose a new benchmark for COntinuAl inStruction Tuning on LVLMs (COAST), which encompasses the aforementioned domain-incremental, capability-incremental, and dataset-incremental configurations. In terms of methodology, we propose Continual LLaVA, a rehearsal-free method tailored for continual instruction tuning in LVLMs. To circumvent the additional overhead associated with experience replay, we freeze LVLMs and construct the dual increment embeddings for each input instruction to facilitate parameter-efficient tuning. Specifically, the increment embeddings can be decomposed into two principal components: 1) intrinsic increment embeddings to encode task-specific characteristics. To achieve this, we set up a low-rank pool containing candidate embeddings, from which we select the relevant ones based on their similarity with the user instructions; 2) contextual increment embeddings to investigate the inter-dependencies across tasks. In this regard, the low-rank embeddings chosen in the previous tasks are aggregated via learnable weighted sum to provide complementary hints. Extensive experiments indicate that the proposed Continual LLaVA outperforms previous methods by significantly reducing the forgetting during the continual instruction tuning process.

Shuo Yang, Yuwei Niu, Yuyang Liu et al.

Multimodal Large Language Models (MLLMs) have achieved remarkable progress in multimodal reasoning. However, they often excessively rely on textual information during the later stages of inference, neglecting the crucial integration of visual input. Current methods typically address this by explicitly injecting visual information to guide the reasoning process. In this work, through an analysis of MLLM attention patterns, we made an intriguing observation: with appropriate guidance, MLLMs can spontaneously re-focus their attention on visual inputs during the later stages of reasoning, even without explicit visual information injection. This spontaneous shift in focus suggests that MLLMs are intrinsically capable of performing visual fusion reasoning. Building on this insight, we introduce Look-Back, an implicit approach designed to guide MLLMs to ``look back" at visual information in a self-directed manner during reasoning. Look-Back empowers the model to autonomously determine when, where, and how to re-focus on visual inputs, eliminating the need for explicit model-structure constraints or additional input. We demonstrate that Look-Back significantly enhances the model's reasoning and perception capabilities, as evidenced by extensive empirical evaluations on multiple multimodal benchmarks.

Tiantian Peng, Yuyang Liu, Shuo Yang et al.

Contrastive Language-Image Pretraining has demonstrated remarkable zero-shot generalization by aligning visual and textual modalities in a shared embedding space. However, when continuously fine-tuned on diverse tasks, CLIP suffers from catastrophic forgetting and degradation of its embedding alignment, undermining its zero-shot capabilities. In this work, we propose Gradient Null Space Projection (GNSP), an efficient continual learning method that projects task-specific gradients onto the null space of previously learned knowledge. This orthogonal projection mathematically prevents interference with previous tasks without relying on rehearsal or architectural modification. Furthermore, to preserve the inherent generalization property of CLIP, we introduce knowledge distillation and combine it with a modality alignment preservation loss inspired by CLIP pre-training to stabilize the structure of the multimodal embedding space during fine-tuning. On the MTIL benchmark consisting of 11 tasks, our method achieved SOTA performance on both the Average and Last key metrics. More importantly, experiments show that our method successfully maintains the original modality gap and cross-modal retrieval performance of CLIP, confirming its effectiveness in maintaining a robust visual-language space throughout the continual learning process.

Xiaoyu Zhang, Mingtao Hu, Sen Lu et al.

State space models (SSMs) have recently emerged as a powerful framework for long sequence processing, outperforming traditional methods on diverse benchmarks. Fundamentally, SSMs can generalize both recurrent and convolutional networks and have been shown to even capture key functions of biological systems. Here we report an approach to implement SSMs in energy-efficient compute-in-memory (CIM) hardware to achieve real-time, event-driven processing. Our work re-parameterizes the model to function with real-valued coefficients and shared decay constants, reducing the complexity of model mapping onto practical hardware systems. By leveraging device dynamics and diagonalized state transition parameters, the state evolution can be natively implemented in crossbar-based CIM systems combined with memristors exhibiting short-term memory effects. Through this algorithm and hardware co-design, we show the proposed system offers both high accuracy and high energy efficiency while supporting fully asynchronous processing for event-based vision and audio tasks.

Yuyang Liu, Jingjing Cai, Jiayi Ren et al.

Anomaly troubleshooting for large model distributed inference (LMDI) remains a critical challenge. Resolving anomalies such as inference performance degradation or latency jitter in distributed system demands significant manual efforts from domain experts, resulting in extremely time-consuming diagnosis processes with relatively low accuracy. In this paper, we introduce Kunlun Anomaly Troubleshooter (KAT), the first anomaly troubleshooting framework tailored for LMDI. KAT addresses this problem through two core innovations. First, KAT exploits the synchronicity and consistency of GPU workers, innovatively leverages function trace data to precisely detect kernel-level anomalies and associated hardware components at nanosecond resolution. Second, KAT integrates these detection results into a domain-adapted LLM, delivering systematic causal reasoning and natural language interpretation of complex anomaly symptoms. Evaluations conducted in Alibaba Cloud Service production environment indicate that KAT achieves over 0.884 precision and 0.936 recall in anomaly detection, providing detail anomaly insights that significantly narrow down the diagnostic scope and improve both the efficiency and success rate of troubleshooting.

Yuyang Liu, Chuan Wen, Yihang Hu et al.

Designing dense rewards is crucial for reinforcement learning (RL), yet in robotics it often demands extensive manual effort and lacks scalability. One promising solution is to view task progress as a dense reward signal, as it quantifies the degree to which actions advance the system toward task completion over time. We present TimeRewarder, a simple yet effective reward learning method that derives progress estimation signals from passive videos, including robot demonstrations and human videos, by modeling temporal distances between frame pairs. We then demonstrate how TimeRewarder can supply step-wise proxy rewards to guide reinforcement learning. In our comprehensive experiments on ten challenging Meta-World tasks, we show that TimeRewarder dramatically improves RL for sparse-reward tasks, achieving nearly perfect success in 9/10 tasks with only 200,000 interactions per task with the environment. This approach outperformed previous methods and even the manually designed environment dense reward on both the final success rate and sample efficiency. Moreover, we show that TimeRewarder pretraining can exploit real-world human videos, highlighting its potential as a scalable approach path to rich reward signals from diverse video sources.

Egor Ershov, Artyom Panshin, Oleg Karasev et al.

This paper presents a review of the NTIRE 2024 challenge on night photography rendering. The goal of the challenge was to find solutions that process raw camera images taken in nighttime conditions, and thereby produce a photo-quality output images in the standard RGB (sRGB) space. Unlike the previous year's competition, the challenge images were collected with a mobile phone and the speed of algorithms was also measured alongside the quality of their output. To evaluate the results, a sufficient number of viewers were asked to assess the visual quality of the proposed solutions, considering the subjective nature of the task. There were 2 nominations: quality and efficiency. Top 5 solutions in terms of output quality were sorted by evaluation time (see Fig. 1). The top ranking participants' solutions effectively represent the state-of-the-art in nighttime photography rendering. More results can be found at https://nightimaging.org.

Tao Zhang, Yang Cong, Gan Sun et al.

Visual-tactile fused sensing for object clustering has achieved significant progresses recently, since the involvement of tactile modality can effectively improve clustering performance. However, the missing data (i.e., partial data) issues always happen due to occlusion and noises during the data collecting process. This issue is not well solved by most existing partial multi-view clustering methods for the heterogeneous modality challenge. Naively employing these methods would inevitably induce a negative effect and further hurt the performance. To solve the mentioned challenges, we propose a Generative Partial Visual-Tactile Fused (i.e., GPVTF) framework for object clustering. More specifically, we first do partial visual and tactile features extraction from the partial visual and tactile data, respectively, and encode the extracted features in modality-specific feature subspaces. A conditional cross-modal clustering generative adversarial network is then developed to synthesize one modality conditioning on the other modality, which can compensate missing samples and align the visual and tactile modalities naturally by adversarial learning. To the end, two pseudo-label based KL-divergence losses are employed to update the corresponding modality-specific encoders. Extensive comparative experiments on three public visual-tactile datasets prove the effectiveness of our method.

Jiahua Dong, Yang Cong, Gan Sun et al.

Unsupervised domain adaptation without consuming annotation process for unlabeled target data attracts appealing interests in semantic segmentation. However, 1) existing methods neglect that not all semantic representations across domains are transferable, which cripples domain-wise transfer with untransferable knowledge; 2) they fail to narrow category-wise distribution shift due to category-agnostic feature alignment. To address above challenges, we develop a new Critical Semantic-Consistent Learning (CSCL) model, which mitigates the discrepancy of both domain-wise and category-wise distributions. Specifically, a critical transfer based adversarial framework is designed to highlight transferable domain-wise knowledge while neglecting untransferable knowledge. Transferability-critic guides transferability-quantizer to maximize positive transfer gain under reinforcement learning manner, although negative transfer of untransferable knowledge occurs. Meanwhile, with the help of confidence-guided pseudo labels generator of target samples, a symmetric soft divergence loss is presented to explore inter-class relationships and facilitate category-wise distribution alignment. Experiments on several datasets demonstrate the superiority of our model.

Yuyang Liu, Yang Cong, Gan Sun

3D object classification has been widely-applied into both academic and industrial scenarios. However, most state-of-the-art algorithms are facing with a fixed 3D object classification task set, which cannot well tackle the new coming data with incremental tasks as human ourselves. Meanwhile, the performance of most state-of-the-art lifelong learning models can be deteriorated easily on previously learned classification tasks, due to the existing of unordered, large-scale, and irregular 3D geometry data. To address this challenge, in this paper, we propose a Lifelong 3D Object Classification (i.e., L3DOC) framewor, which can consecutively learn new 3D object classification tasks via imitating 'human learning'. Specifically, the core idea of our proposed L3DOC model is to factorize PointNet in a perspective of lifelong learning, while capturing and storing the shared point-knowledge in a perspective of layer-wise tensor factorization architecture. To further transfer the task-specific knowledge from previous tasks to the new coming classification task, a memory attention mechanism is proposed to connect the current task with relevant previously tasks, which can effectively prevent catastrophic forgetting via soft-transferring previous knowledge. To our best knowledge, this is the first work about using lifelong learning to handle 3D object classification task without model fine-tuning or retraining. Furthermore, our L3DOC model can also be extended to other backbone network (e.g., PointNet++). To the end, comparisons on several point cloud datasets validate that our L3DOC model can reduce averaged 1.68~3.36 times parameters for the overall model, without sacrificing classification accuracy of each task.

Tai-Long He, Dylan B. A. Jones, Binxuan Huang et al.

We use a hybrid deep learning model to predict June-July-August (JJA) daily maximum 8-h average (MDA8) surface ozone concentrations in the US. A set of meteorological fields from the ERA-Interim reanalysis as well as monthly mean NO$_x$ emissions from the Community Emissions Data System (CEDS) inventory are selected as predictors. Ozone measurements from the US Environmental Protection Agency (EPA) Air Quality System (AQS) from 1980 to 2009 are used to train the model, whereas data from 2010 to 2014 are used to evaluate the performance of the model. The model captures well daily, seasonal and interannual variability in MDA8 ozone across the US. Feature maps show that the model captures teleconnections between MDA8 ozone and the meteorological fields, which are responsible for driving the ozone dynamics. We used the model to evaluate recent trends in NO$_x$ emissions in the US and found that the trend in the EPA emission inventory produced the largest negative bias in MDA8 ozone between 2010-2016. The top-down emission trends from the Tropospheric Chemistry Reanalysis (TCR-2), which is based on satellite observations, produced predictions in best agreement with observations. In urban regions, the trend in AQS NO$_2$ observations provided ozone predictions in agreement with observations, whereas in rural regions the satellite-derived trends produced the best agreement. In both rural and urban regions the EPA trend resulted in the largest negative bias in predicted ozone. Our results suggest that the EPA inventory is overestimating the reductions in NO$_x$ emissions and that the satellite-derived trend reflects the influence of reductions in NO$_x$ emissions as well as changes in background NO$_x$. Our results demonstrate the significantly greater predictive capability that the deep learning model provides over conventional atmospheric chemical transport models for air quality analyses.

Stephen Zhen Gou, Yuyang Liu

We propose Deep Q-Networks (DQN) with model-based exploration, an algorithm combining both model-free and model-based approaches that explores better and learns environments with sparse rewards more efficiently. DQN is a general-purpose, model-free algorithm and has been proven to perform well in a variety of tasks including Atari 2600 games since it's first proposed by Minh et el. However, like many other reinforcement learning (RL) algorithms, DQN suffers from poor sample efficiency when rewards are sparse in an environment. As a result, most of the transitions stored in the replay memory have no informative reward signal, and provide limited value to the convergence and training of the Q-Network. However, one insight is that these transitions can be used to learn the dynamics of the environment as a supervised learning problem. The transitions also provide information of the distribution of visited states. Our algorithm utilizes these two observations to perform a one-step planning during exploration to pick an action that leads to states least likely to be seen, thus improving the performance of exploration. We demonstrate our agent's performance in two classic environments with sparse rewards in OpenAI gym: Mountain Car and Lunar Lander.