Chenyang Wang

Xiaoyu Han, Shunyuan Zheng, Zonglin Li et al.

Image-based virtual try-on aims to seamlessly fit in-shop clothing to a person image while maintaining pose consistency. Existing methods commonly employ the thin plate spline (TPS) transformation or appearance flow to deform in-shop clothing for aligning with the person's body. Despite their promising performance, these methods often lack precise control over fine details, leading to inconsistencies in shape between clothing and the person's body as well as distortions in exposed limb regions. To tackle these challenges, we propose a novel shape-guided clothing warping method for virtual try-on, dubbed SCW-VTON, which incorporates global shape constraints and additional limb textures to enhance the realism and consistency of the warped clothing and try-on results. To integrate global shape constraints for clothing warping, we devise a dual-path clothing warping module comprising a shape path and a flow path. The former path captures the clothing shape aligned with the person's body, while the latter path leverages the mapping between the pre- and post-deformation of the clothing shape to guide the estimation of appearance flow. Furthermore, to alleviate distortions in limb regions of try-on results, we integrate detailed limb guidance by developing a limb reconstruction network based on masked image modeling. Through the utilization of SCW-VTON, we are able to generate try-on results with enhanced clothing shape consistency and precise control over details. Extensive experiments demonstrate the superiority of our approach over state-of-the-art methods both qualitatively and quantitatively. The code is available at https://github.com/xyhanHIT/SCW-VTON.

Jiarui Che, Yifei Chen, Zhixing Tian et al.

Search relevance modeling is a core task in e-commerce search systems, assessing how well a user query matches candidate products. Rather than relying on a single holistic matching signal, relevance judgment often requires structured reasoning over query understanding, product understanding, and facet-level matching. With large language models (LLMs), this process is increasingly formulated as chain-of-thought (CoT) reasoning and optimized with reinforcement learning (RL). However, existing RL methods mainly rely on outcome-level rewards and treat the entire reasoning chain as a single optimization unit. This makes it difficult to distinguish faulty reasoning steps from correct intermediate ones, leading to misaligned credit assignment. Although process-reward methods provide denser supervision, they often treat reasoning steps independently and ignore dependency-driven error propagation, making responsibility attribution difficult and limiting the optimization of structured relevance reasoning. We propose Graph-GRPO, a graph-structured extension of GRPO for multi-component relevance reasoning. Graph-GRPO constructs a relevance reasoning dependency graph, where CoT steps are modeled as nodes and their logical dependencies as edges. It propagates outcome-level rewards over the graph to derive step-level credit signals, enabling more accurate fine-grained credit assignment. We further introduce a main-loss-driven controller that adaptively adjusts edge-wise credit-propagation coefficients. Together with CoT random masking for supervised policy initialization and graph-node-based multi-head distillation, we build a trainable and deployable framework for generative relevance modeling. Extensive offline evaluations and online A/B tests on a leading e-commerce platform demonstrate that the Graph-GRPO-based framework improves relevance classification metrics and key engagement metrics.

Chenyang Wang, Junjun Jiang, Zhiwei Zhong et al.

Face super-resolution is a technology that transforms a low-resolution face image into the corresponding high-resolution one. In this paper, we build a novel parsing map guided face super-resolution network which extracts the face prior (i.e., parsing map) directly from low-resolution face image for the following utilization. To exploit the extracted prior fully, a parsing map attention fusion block is carefully designed, which can not only effectively explore the information of parsing map, but also combines powerful attention mechanism. Moreover, in light of that high-resolution features contain more precise spatial information while low-resolution features provide strong contextual information, we hope to maintain and utilize these complementary information. To achieve this goal, we develop a multi-scale refine block to maintain spatial and contextual information and take advantage of multi-scale features to refine the feature representations. Experimental results demonstrate that our method outperforms the state-of-the-arts in terms of quantitative metrics and visual quality. The source codes will be available at https://github.com/wcy-cs/FishFSRNet.

Chenyang Wang, Junjun Jiang, Xiong Zhou et al.

Data augmentation (DA) is a widely used technique for enhancing the training of deep neural networks. Recent DA techniques which achieve state-of-the-art performance always meet the need for diversity in augmented training samples. However, an augmentation strategy that has a high diversity usually introduces out-of-distribution (OOD) augmented samples and these samples consequently impair the performance. To alleviate this issue, we propose ReSmooth, a framework that firstly detects OOD samples in augmented samples and then leverages them. To be specific, we first use a Gaussian mixture model to fit the loss distribution of both the original and augmented samples and accordingly split these samples into in-distribution (ID) samples and OOD samples. Then we start a new training where ID and OOD samples are incorporated with different smooth labels. By treating ID samples and OOD samples unequally, we can make better use of the diverse augmented data. Further, we incorporate our ReSmooth framework with negative data augmentation strategies. By properly handling their intentionally created OOD samples, the classification performance of negative data augmentations is largely ameliorated. Experiments on several classification benchmarks show that ReSmooth can be easily extended to existing augmentation strategies (such as RandAugment, rotate, and jigsaw) and improve on them. Our code is available at https://github.com/Chenyang4/ReSmooth.

Kaituo Feng, Manyuan Zhang, Shuang Chen et al.

Recent image generation models have shown strong capabilities in generating high-fidelity and photorealistic images. However, they are fundamentally constrained by frozen internal knowledge, thus often failing on real-world scenarios that are knowledge-intensive or require up-to-date information. In this paper, we present Gen-Searcher, as the first attempt to train a search-augmented image generation agent, which performs multi-hop reasoning and search to collect the textual knowledge and reference images needed for grounded generation. To achieve this, we construct a tailored data pipeline and curate two high-quality datasets, Gen-Searcher-SFT-10k and Gen-Searcher-RL-6k, containing diverse search-intensive prompts and corresponding ground-truth synthesis images. We further introduce KnowGen, a comprehensive benchmark that explicitly requires search-grounded external knowledge for image generation and evaluates models from multiple dimensions. Based on these resources, we train Gen-Searcher with SFT followed by agentic reinforcement learning with dual reward feedback, which combines text-based and image-based rewards to provide more stable and informative learning signals for GRPO training. Experiments show that Gen-Searcher brings substantial gains, improving Qwen-Image by around 16 points on KnowGen and 15 points on WISE. We hope this work can serve as an open foundation for search agents in image generation, and we fully open-source our data, models, and code.

Lifan Jiang, Tianrun Wu, Yuhang Pei et al.

The evaluation of visual editing models remains fragmented across methods and modalities. Existing benchmarks are often tailored to specific paradigms, making fair cross-paradigm comparisons difficult, while video editing lacks reliable evaluation benchmarks. Furthermore, common automatic metrics often misalign with human preference, yet directly deploying large multimodal models (MLLMs) as evaluators incurs prohibitive computational and financial costs. We present UniEditBench, a unified benchmark for image and video editing that supports reconstruction-based and instruction-driven methods under a shared protocol. UniEditBench includes a structured taxonomy of nine image operations (Add, Remove, Replace, Change, Stroke-based, Extract, Adjust, Count, Reorder) and eight video operations, with coverage of challenging compositional tasks such as counting and spatial reordering. To enable scalable evaluation, we distill a high-capacity MLLM judge (Qwen3-VL-235B-A22B Instruct) into lightweight 4B/8B evaluators that provide multi-dimensional scoring over structural fidelity, text alignment, background consistency, naturalness, and temporal-spatial consistency (for videos). Experiments show that the distilled evaluators maintain strong agreement with human judgments and substantially reduce deployment cost relative to the teacher model. UniEditBench provides a practical and reproducible protocol for benchmarking modern visual editing methods. Our benchmark and the associated reward models are publicly available at https://github.com/wesar1/UniEditBench.

Chenyang Wang, Weizhong Wang, Yinuo Ren et al.

iffusion-based generative models increasingly rely on inference-time guidance, adding a drift term or reweighting mixture of experts, to improve sample quality on task-specific objectives. However, most existing techniques require repeated score or gradient evaluations, introducing bias, high computational overhead, or both. We introduce \texttt{URGE}, Unbiased Resampling via Girsanov Estimation, a derivative-free inference-time scaling algorithm that performs path-wise importance reweighting via a Girsanov change of measure. Instead of computing gradient-based particle weights in previous work, \texttt{URGE} attaches a simple multiplicative weight to each simulated trajectory and periodically resamples. No score, no Hessian, and no PDE evaluation is required. We establish an equivalence between path-wise and particle-wise SMC: the Girsanov path weight admits a backward conditional expectation that recovers the previous particle-level weights, guaranteeing that both schemes produce the same unbiased terminal law. Empirically, \texttt{URGE} outperforms existing inference-time guidance baselines on synthetic tests and diffusion-model benchmarks, achieving better generation quality, while being significantly simpler to implement and fully gradient-free.

Tian Xu, Chenyang Wang, Xiaochen Zhai et al.

Adversarial imitation learning (AIL) achieves high-quality imitation by mitigating compounding errors in behavioral cloning (BC), but often exhibits training instability due to adversarial optimization. To avoid this issue, a class of non-adversarial Q-based imitation learning (IL) methods, represented by IQ-Learn, has emerged and is widely believed to outperform BC by leveraging online environment interactions. However, this paper revisits IQ-Learn and demonstrates that it provably reduces to BC and suffers from an imitation gap lower bound with quadratic dependence on horizon, therefore still suffering from compounding errors. Theoretical analysis reveals that, despite using online interactions, IQ-Learn uniformly suppresses the Q-values for all actions on states uncovered by demonstrations, thereby failing to generalize. To address this limitation, we introduce a primal-dual framework for distribution matching, yielding a new Q-based IL method, Dual Q-DM. The key mechanism in Dual Q-DM is incorporating Bellman constraints to propagate high Q-values from visited states to unvisited ones, thereby achieving generalization beyond demonstrations. We prove that Dual Q-DM is equivalent to AIL and can recover expert actions beyond demonstrations, thereby mitigating compounding errors. To the best of our knowledge, Dual Q-DM is the first non-adversarial IL method that is theoretically guaranteed to eliminate compounding errors. Experimental results further corroborate our theoretical results.

Peng-Yuan Wang, Ziniu Li, Tian Xu et al.

Improving data utilization efficiency is critical for scaling reinforcement learning (RL) for long-horizon tasks where generating trajectories is expensive. However, the dominant RL methods for LLMs are largely on-policy: they update each batch of data only once, discard it, and then collect fresh samples, resulting in poor sample efficiency. In this work, we explore an alternative value-based RL framework for LLMs that naturally enables off-policy learning. We propose ReVal, a Bellman-update-based method that combines stepwise signals capturing internal consistency with trajectory-level signals derived from outcome verification. ReVal naturally supports replay-buffer-based training, allowing efficient reuse of past trajectories. Experiments on standard mathematical reasoning benchmarks show that ReVal not only converges faster but also outperforms GRPO in final performance. On DeepSeek-R1-Distill-1.5B, ReVal improves training efficiency and achieves improvement of 2.7% in AIME24 and 4.5% in out-of-domain benchmark GPQA over GRPO. These results suggest that value-based RL is a practical alternative to policy-based methods for LLM training.

Xiaoyu Han, Shengping Zhang, Qinglin Liu et al.

Existing image-based virtual try-on methods directly transfer specific clothing to a human image without utilizing clothing attributes to refine the transferred clothing geometry and textures, which causes incomplete and blurred clothing appearances. In addition, these methods usually mask the limb textures of the input for the clothing-agnostic person representation, which results in inaccurate predictions for human limb regions (i.e., the exposed arm skin), especially when transforming between long-sleeved and short-sleeved garments. To address these problems, we present a progressive virtual try-on framework, named PL-VTON, which performs pixel-level clothing warping based on multiple attributes of clothing and embeds explicit limb-aware features to generate photo-realistic try-on results. Specifically, we design a Multi-attribute Clothing Warping (MCW) module that adopts a two-stage alignment strategy based on multiple attributes to progressively estimate pixel-level clothing displacements. A Human Parsing Estimator (HPE) is then introduced to semantically divide the person into various regions, which provides structural constraints on the human body and therefore alleviates texture bleeding between clothing and limb regions. Finally, we propose a Limb-aware Texture Fusion (LTF) module to estimate high-quality details in limb regions by fusing textures of the clothing and the human body with the guidance of explicit limb-aware features. Extensive experiments demonstrate that our proposed method outperforms the state-of-the-art virtual try-on methods both qualitatively and quantitatively. The code is available at https://github.com/xyhanHIT/PL-VTON.

Jie Ying, Zihong Chen, Zhefan Wang et al.

Seed science is essential for modern agriculture, directly influencing crop yields and global food security. However, challenges such as interdisciplinary complexity and high costs with limited returns hinder progress, leading to a shortage of experts and insufficient technological support. While large language models (LLMs) have shown promise across various fields, their application in seed science remains limited due to the scarcity of digital resources, complex gene-trait relationships, and the lack of standardized benchmarks. To address this gap, we introduce SeedBench -- the first multi-task benchmark specifically designed for seed science. Developed in collaboration with domain experts, SeedBench focuses on seed breeding and simulates key aspects of modern breeding processes. We conduct a comprehensive evaluation of 26 leading LLMs, encompassing proprietary, open-source, and domain-specific fine-tuned models. Our findings not only highlight the substantial gaps between the power of LLMs and the real-world seed science problems, but also make a foundational step for research on LLMs for seed design.

Haoyu Wang, Chunyu Qiang, Tianrui Wang et al.

Recent advancements in speech synthesis models, trained on extensive datasets, have demonstrated remarkable zero-shot capabilities. These models can control content, timbre, and emotion in generated speech based on prompt inputs. Despite these advancements, the choice of prompts significantly impacts the output quality, yet most existing selection schemes do not adequately address the control of emotional intensity. To address this question, this paper proposes a two-stage prompt selection strategy EmoPro, which is specifically designed for emotionally controllable speech synthesis. This strategy focuses on selecting highly expressive and high-quality prompts by evaluating them from four perspectives: emotional expression strength, speech quality, text-emotion consistency, and model generation performance. Experimental results show that prompts selected using the proposed method result in more emotionally expressive and engaging synthesized speech compared to those obtained through baseline. Audio samples and codes will be available at https://whyrrrrun.github.io/EmoPro/.

Chenyang Wang, Liang Wen, Shousheng Jia et al.

While advancements in the reasoning abilities of LLMs have significantly enhanced their performance in solving mathematical problems, coding tasks, and general puzzles, their effectiveness in accurately adhering to instructions remains inconsistent, particularly with more complex directives. Our investigation identifies lazy reasoning during the thinking stage as the primary factor contributing to poor instruction adherence. To mitigate this issue, we propose a comprehensive framework designed to enable rigorous reasoning processes involving preview and self-checking, essential for satisfying strict instruction constraints. Specifically, we first generate instructions with complex constraints and apply a filtering process to obtain valid prompts, resulting in three distinct prompt datasets categorized as hard, easy, and pass. Then, we employ rejection sampling on the pass prompts to curate a small yet high-quality dataset, enabling a cold-start initialization of the model and facilitating its adaptation to effective reasoning patterns. Subsequently, we employ an entropy-preserving supervised fine-tuning (Entropy-SFT) strategy coupled with token-wise entropy-adaptive (TEA-RL) reinforcement learning guided by rule-based dense rewards. This approach encourages the model to transform its reasoning mechanism, ultimately fostering generalizable reasoning abilities that encompass preview and self-checking. Extensive experiments conducted on instruction-following benchmarks demonstrate remarkable performance improvements across various model scales. Notably, our Light-IF-32B model surpasses both larger open-source models such as DeepSeek-R1 and closed-source models like Doubao-1.6.

Xiao Wei, Xiaobao Wang, Ning Zhuang et al.

Intent detection aims to identify user intents from natural language inputs, where supervised methods rely heavily on labeled in-domain (IND) data and struggle with out-of-domain (OOD) intents, limiting their practical applicability. Generalized Intent Discovery (GID) addresses this by leveraging unlabeled OOD data to discover new intents without additional annotation. However, existing methods focus solely on clustering unsupervised data while neglecting domain adaptation. Therefore, we propose a consistency-driven prototype-prompting framework for GID from the perspective of integrating old and new knowledge, which includes a prototype-prompting framework for transferring old knowledge from external sources, and a hierarchical consistency constraint for learning new knowledge from target domains. We conducted extensive experiments and the results show that our method significantly outperforms all baseline methods, achieving state-of-the-art results, which strongly demonstrates the effectiveness and generalization of our methods. Our source code is publicly available at https://github.com/smileix/cpp.

Xingyu Hu, Junjun Jiang, Chenyang Wang et al.

Unified image fusion aims to integrate complementary information from multi-source images, enhancing image quality through a unified framework applicable to diverse fusion tasks. While treating all fusion tasks as a unified problem facilitates task-invariant knowledge sharing, it often overlooks task-specific characteristics, thereby limiting the overall performance. Existing general image fusion methods incorporate explicit task identification to enable adaptation to different fusion tasks. However, this dependence during inference restricts the model's generalization to unseen fusion tasks. To address these issues, we propose a novel unified image fusion framework named "TITA", which dynamically balances both Task-invariant Interaction and Task-specific Adaptation. For task-invariant interaction, we introduce the Interaction-enhanced Pixel Attention (IPA) module to enhance pixel-wise interactions for better multi-source complementary information extraction. For task-specific adaptation, the Operation-based Adaptive Fusion (OAF) module dynamically adjusts operation weights based on task properties. Additionally, we incorporate the Fast Adaptive Multitask Optimization (FAMO) strategy to mitigate the impact of gradient conflicts across tasks during joint training. Extensive experiments demonstrate that TITA not only achieves competitive performance compared to specialized methods across three image fusion scenarios but also exhibits strong generalization to unseen fusion tasks. The source codes are released at https://github.com/huxingyuabc/TITA.

Benxue Gong, Zhenyu Zhao, Chenyang Wang

We propose a local Legendre frame (LLF) method for function approximation from equispaced data on a finite interval. Motivated by the difficulty of stable high-order polynomial approximation at equispaced points, especially in the presence of the Runge phenomenon, the method partitions the interval into subintervals, maps each subinterval to a common reference interval, and computes local coefficients by a truncated singular value decomposition (TSVD) regularization. Since all subintervals share the same local sampling matrix, the method admits a natural offline--online implementation. We establish a quasi-optimal estimate for the regularized reconstruction and discuss practical parameter selection. Numerical results show that LLF attains high accuracy for relatively smooth and moderately oscillatory functions, while it remains applicable to highly oscillatory functions, although comparable accuracy generally requires more sampling points. For continuous piecewise smooth functions with derivative singularities, the method also provides an effective detect--localize--correct strategy based on one-sided coefficient-energy indicators. These results indicate that LLF provides a stable and flexible local approximation framework for equispaced data.

Xiong Zhou, Xianming Liu, Chenyang Wang et al.

Learning with noisy labels is an important and challenging task for training accurate deep neural networks. Some commonly-used loss functions, such as Cross Entropy (CE), suffer from severe overfitting to noisy labels. Robust loss functions that satisfy the symmetric condition were tailored to remedy this problem, which however encounter the underfitting effect. In this paper, we theoretically prove that \textbf{any loss can be made robust to noisy labels} by restricting the network output to the set of permutations over a fixed vector. When the fixed vector is one-hot, we only need to constrain the output to be one-hot, which however produces zero gradients almost everywhere and thus makes gradient-based optimization difficult. In this work, we introduce the sparse regularization strategy to approximate the one-hot constraint, which is composed of network output sharpening operation that enforces the output distribution of a network to be sharp and the $\ell_p$-norm ($p\le 1$) regularization that promotes the network output to be sparse. This simple approach guarantees the robustness of arbitrary loss functions while not hindering the fitting ability. Experimental results demonstrate that our method can significantly improve the performance of commonly-used loss functions in the presence of noisy labels and class imbalance, and outperform the state-of-the-art methods. The code is available at https://github.com/hitcszx/lnl_sr.

Junjun Jiang, Chenyang Wang, Xianming Liu et al.

Face super-resolution (FSR), also known as face hallucination, which is aimed at enhancing the resolution of low-resolution (LR) face images to generate high-resolution (HR) face images, is a domain-specific image super-resolution problem. Recently, FSR has received considerable attention and witnessed dazzling advances with the development of deep learning techniques. To date, few summaries of the studies on the deep learning-based FSR are available. In this survey, we present a comprehensive review of deep learning-based FSR methods in a systematic manner. First, we summarize the problem formulation of FSR and introduce popular assessment metrics and loss functions. Second, we elaborate on the facial characteristics and popular datasets used in FSR. Third, we roughly categorize existing methods according to the utilization of facial characteristics. In each category, we start with a general description of design principles, then present an overview of representative approaches, and then discuss the pros and cons among them. Fourth, we evaluate the performance of some state-of-the-art methods. Fifth, joint FSR and other tasks, and FSR-related applications are roughly introduced. Finally, we envision the prospects of further technological advancement in this field. A curated list of papers and resources to face super-resolution are available at \url{https://github.com/junjun-jiang/Face-Hallucination-Benchmark}

Kazuya Nishimura, Junya Hayashida, Chenyang Wang et al.

We propose a weakly-supervised cell tracking method that can train a convolutional neural network (CNN) by using only the annotation of "cell detection" (i.e., the coordinates of cell positions) without association information, in which cell positions can be easily obtained by nuclear staining. First, we train co-detection CNN that detects cells in successive frames by using weak-labels. Our key assumption is that co-detection CNN implicitly learns association in addition to detection. To obtain the association, we propose a backward-and-forward propagation method that analyzes the correspondence of cell positions in the outputs of co-detection CNN. Experiments demonstrated that the proposed method can associate cells by analyzing co-detection CNN. Even though the method uses only weak supervision, the performance of our method was almost the same as the state-of-the-art supervised method. Code is publicly available in https://github.com/naivete5656/WSCTBFP

Weizhi Ma, Min Zhang, Yue Cao et al.

Explainability and effectiveness are two key aspects for building recommender systems. Prior efforts mostly focus on incorporating side information to achieve better recommendation performance. However, these methods have some weaknesses: (1) prediction of neural network-based embedding methods are hard to explain and debug; (2) symbolic, graph-based approaches (e.g., meta path-based models) require manual efforts and domain knowledge to define patterns and rules, and ignore the item association types (e.g. substitutable and complementary). In this paper, we propose a novel joint learning framework to integrate \textit{induction of explainable rules from knowledge graph} with \textit{construction of a rule-guided neural recommendation model}. The framework encourages two modules to complement each other in generating effective and explainable recommendation: 1) inductive rules, mined from item-centric knowledge graphs, summarize common multi-hop relational patterns for inferring different item associations and provide human-readable explanation for model prediction; 2) recommendation module can be augmented by induced rules and thus have better generalization ability dealing with the cold-start issue. Extensive experiments\footnote{Code and data can be found at: \url{https://github.com/THUIR/RuleRec}} show that our proposed method has achieved significant improvements in item recommendation over baselines on real-world datasets. Our model demonstrates robust performance over "noisy" item knowledge graphs, generated by linking item names to related entities.

Chenyang Wang, Yun Yang

We derive explicit non-asymptotic PAC-Bayes generalization bounds for Gibbs posteriors, that is, data-dependent distributions over model parameters obtained by exponentially tilting a prior with the empirical risk. Unlike classical worst-case complexity bounds based on uniform laws of large numbers, which require explicit control of the model space in terms of metric entropy (integrals), our analysis yields posterior-averaged risk bounds that can be applied to overparameterized models and adapt to the data structure and the intrinsic model complexity. The bound involves a marginal-type integral over the parameter space, which we analyze using tools from singular learning theory to obtain explicit and practically meaningful characterizations of the posterior risk. Applications to low-rank matrix completion and ReLU neural network regression and classification show that the resulting bounds are analytically tractable and substantially tighter than classical complexity-based bounds. Our results highlight the potential of PAC-Bayes analysis for precise finite-sample generalization guarantees in modern overparameterized and singular models.

Lei Zhu, Xiaobao Wang, Jianbiao Yang et al.

Factual Error Correction (FEC) aims to revise inaccurate text into statements that are factually consistent with external evidence. Although recent methods perform well on single-hop correction, they often treat claims as atomic units and struggle with multi-hop cases that require compositional reasoning across multiple evidence sources. This challenge is further amplified by limited paired data and difficulties in locating semantic errors within complex reasoning chains. We present CECoR (Compositional Error Correction via Reasoning-aware Synthesis), a reasoning-aware framework that introduces a Decomposition and Injection paradigm for compositional error correction. CECoR decomposes multi-hop claims into interpretable reasoning steps and injects controlled perturbations to synthesize high-quality training pairs. A two-stage learning strategy combining supervised fine-tuning and reinforcement learning improves factual accuracy and robustness. Comprehensive evaluations show that CECoR achieves strong performance on multi-hop benchmarks, outperforming both distantly supervised methods and few-shot LLM baselines. It also generalizes effectively to single-hop correction and remains stable under noisy evidence, demonstrating its versatility for real-world factual correction.

Shi Qiu, Shaoyang Guo, Zhuo-Yang Song et al.

Current benchmarks for evaluating the reasoning capabilities of Large Language Models (LLMs) face significant limitations: task oversimplification, data contamination, and flawed evaluation items. These deficiencies necessitate more rigorous assessment methods. To address these limitations, we introduce PHYBench, a benchmark of 500 original physics problems ranging from high school to Physics Olympiad difficulty. PHYBench addresses data contamination through original content and employs a systematic curation pipeline to eliminate flawed items. Evaluations show that PHYBench activates more tokens and provides stronger differentiation between reasoning models compared to other baselines like AIME 2024, OlympiadBench and GPQA. Even the best-performing model, Gemini 2.5 Pro, achieves only 36.9% accuracy compared to human experts' 61.9%. To further enhance evaluation precision, we introduce the Expression Edit Distance (EED) Score for mathematical expression assessment, which improves sample efficiency by 204% over binary scoring. Moreover, PHYBench effectively elicits multi-step and multi-condition reasoning, providing a platform for examining models' reasoning robustness, preferences, and deficiencies. The benchmark results and dataset are publicly available at https://www.phybench.cn/.

Zheng Chen, Jingkai Wang, Kai Liu et al.

This paper provides a review of the NTIRE 2025 challenge on real-world face restoration, highlighting the proposed solutions and the resulting outcomes. The challenge focuses on generating natural, realistic outputs while maintaining identity consistency. Its goal is to advance state-of-the-art solutions for perceptual quality and realism, without imposing constraints on computational resources or training data. The track of the challenge evaluates performance using a weighted image quality assessment (IQA) score and employs the AdaFace model as an identity checker. The competition attracted 141 registrants, with 13 teams submitting valid models, and ultimately, 10 teams achieved a valid score in the final ranking. This collaborative effort advances the performance of real-world face restoration while offering an in-depth overview of the latest trends in the field.

Peng-Yuan Wang, Tian-Shuo Liu, Chenyang Wang et al.

Mathematical reasoning has long represented one of the most fundamental and challenging frontiers in artificial intelligence research. In recent years, large language models (LLMs) have achieved significant advances in this area. This survey examines the development of mathematical reasoning abilities in LLMs through two high-level cognitive phases: comprehension, where models gain mathematical understanding via diverse pretraining strategies, and answer generation, which has progressed from direct prediction to step-by-step Chain-of-Thought (CoT) reasoning. We review methods for enhancing mathematical reasoning, ranging from training-free prompting to fine-tuning approaches such as supervised fine-tuning and reinforcement learning, and discuss recent work on extended CoT and "test-time scaling". Despite notable progress, fundamental challenges remain in terms of capacity, efficiency, and generalization. To address these issues, we highlight promising research directions, including advanced pretraining and knowledge augmentation techniques, formal reasoning frameworks, and meta-generalization through principled learning paradigms. This survey tries to provide some insights for researchers interested in enhancing reasoning capabilities of LLMs and for those seeking to apply these techniques to other domains.

Chenyang Wang, Junjun Jiang, Xingyu Hu et al.

Deep learning systems are prone to catastrophic forgetting when learning from a sequence of tasks, as old data from previous tasks is unavailable when learning a new task. To address this, some methods propose replaying data from previous tasks during new task learning, typically using extra memory to store replay data. However, it is not expected in practice due to memory constraints and data privacy issues. Instead, data-free replay methods invert samples from the classification model. While effective, these methods face inconsistencies between inverted and real training data, overlooked in recent works. To that effect, we propose to measure the data consistency quantitatively by some simplification and assumptions. Using this measurement, we gain insight to develop a novel loss function that reduces inconsistency. Specifically, the loss minimizes the KL divergence between distributions of inverted and real data under a tied multivariate Gaussian assumption, which is simple to implement in continual learning. Additionally, we observe that old class weight norms decrease continually as learning progresses. We analyze the reasons and propose a regularization term to balance class weights, making old class samples more distinguishable. To conclude, we introduce Consistency-enhanced data replay with a Debiased classifier for class incremental learning (CwD). Extensive experiments on CIFAR-100, Tiny-ImageNet, and ImageNet100 show consistently improved performance of CwD compared to previous approaches.

Tian Tang, Zhixing Tian, Zhenyu Zhu et al.

Query and product relevance prediction is a critical component for ensuring a smooth user experience in e-commerce search. Traditional studies mainly focus on BERT-based models to assess the semantic relevance between queries and products. However, the discriminative paradigm and limited knowledge capacity of these approaches restrict their ability to comprehend the relevance between queries and products fully. With the rapid advancement of Large Language Models (LLMs), recent research has begun to explore their application to industrial search systems, as LLMs provide extensive world knowledge and flexible optimization for reasoning processes. Nonetheless, directly leveraging LLMs for relevance prediction tasks introduces new challenges, including a high demand for data quality, the necessity for meticulous optimization of reasoning processes, and an optimistic bias that can result in over-recall. To overcome the above problems, this paper proposes a novel framework called the LLM-based RElevance Framework (LREF) aimed at enhancing e-commerce search relevance. The framework comprises three main stages: supervised fine-tuning (SFT) with Data Selection, Multiple Chain of Thought (Multi-CoT) tuning, and Direct Preference Optimization (DPO) for de-biasing. We evaluate the performance of the framework through a series of offline experiments on large-scale real-world datasets, as well as online A/B testing. The results indicate significant improvements in both offline and online metrics. Ultimately, the model was deployed in a well-known e-commerce application, yielding substantial commercial benefits.

Chenyang Wang, Wenjie An, Kui Jiang et al.

Existing face super-resolution (FSR) methods have made significant advancements, but they primarily super-resolve face with limited visual information, original pixel-wise space in particular, commonly overlooking the pluralistic clues, like the higher-order depth and semantics, as well as non-visual inputs (text caption and description). Consequently, these methods struggle to produce a unified and meaningful representation from the input face. We suppose that introducing the language-vision pluralistic representation into unexplored potential embedding space could enhance FSR by encoding and exploiting the complementarity across language-vision prior. This motivates us to propose a new framework called LLV-FSR, which marries the power of large vision-language model and higher-order visual prior with the challenging task of FSR. Specifically, besides directly absorbing knowledge from original input, we introduce the pre-trained vision-language model to generate pluralistic priors, involving the image caption, descriptions, face semantic mask and depths. These priors are then employed to guide the more critical feature representation, facilitating realistic and high-quality face super-resolution. Experimental results demonstrate that our proposed framework significantly improves both the reconstruction quality and perceptual quality, surpassing the SOTA by 0.43dB in terms of PSNR on the MMCelebA-HQ dataset.

Yisi Liu, Chenyang Wang, Hanjo Kim et al.

Voice conversion has emerged as a pivotal technology in numerous applications ranging from assistive communication to entertainment. In this paper, we present RT-VC, a zero-shot real-time voice conversion system that delivers ultra-low latency and high-quality performance. Our approach leverages an articulatory feature space to naturally disentangle content and speaker characteristics, facilitating more robust and interpretable voice transformations. Additionally, the integration of differentiable digital signal processing (DDSP) enables efficient vocoding directly from articulatory features, significantly reducing conversion latency. Experimental evaluations demonstrate that, while maintaining synthesis quality comparable to the current state-of-the-art (SOTA) method, RT-VC achieves a CPU latency of 61.4 ms, representing a 13.3\% reduction in latency.

Yipengjing Sun, Chenyang Wang, Shunyuan Zheng et al.

We propose GRGS, a generalizable and relightable 3D Gaussian framework for high-fidelity human novel view synthesis under diverse lighting conditions. Unlike existing methods that rely on per-character optimization or ignore physical constraints, GRGS adopts a feed-forward, fully supervised strategy that projects geometry, material, and illumination cues from multi-view 2D observations into 3D Gaussian representations. Specifically, to reconstruct lighting-invariant geometry, we introduce a Lighting-aware Geometry Refinement (LGR) module trained on synthetically relit data to predict accurate depth and surface normals. Based on the high-quality geometry, a Physically Grounded Neural Rendering (PGNR) module is further proposed to integrate neural prediction with physics-based shading, supporting editable relighting with shadows and indirect illumination. Besides, we design a 2D-to-3D projection training scheme that leverages differentiable supervision from ambient occlusion, direct, and indirect lighting maps, which alleviates the computational cost of explicit ray tracing. Extensive experiments demonstrate that GRGS achieves superior visual quality, geometric consistency, and generalization across characters and lighting conditions.

Weize Xie, Yi Ding, Ying He et al.

Diffusion strategies have advanced visual motor control by progressively denoising high-dimensional action sequences, providing a promising method for robot manipulation. However, as task complexity increases, the success rate of existing baseline models decreases considerably. Analysis indicates that current diffusion strategies are confronted with two limitations. First, these strategies only rely on short-term observations as conditions. Second, the training objective remains limited to a single denoising loss, which leads to error accumulation and causes grasping deviations. To address these limitations, this paper proposes Foresight-Conditioned Diffusion (ForeDiffusion), by injecting the predicted future view representation into the diffusion process. As a result, the policy is guided to be forward-looking, enabling it to correct trajectory deviations. Following this design, ForeDiffusion employs a dual loss mechanism, combining the traditional denoising loss and the consistency loss of future observations, to achieve the unified optimization. Extensive evaluation on the Adroit suite and the MetaWorld benchmark demonstrates that ForeDiffusion achieves an average success rate of 80% for the overall task, significantly outperforming the existing mainstream diffusion methods by 23% in complex tasks, while maintaining more stable performance across the entire tasks.

Weijia Feng, Jingyu Yang, Ruojia Zhang et al.

Micro-gestures are subtle and transient movements triggered by unconscious neural and emotional activities, holding great potential for human-computer interaction and clinical monitoring. However, their low amplitude, short duration, and strong inter-subject variability make existing deep models prone to degradation under low-sample, noisy, and cross-subject conditions. This paper presents an active inference-based framework for micro-gesture recognition, featuring Expected Free Energy (EFE)-guided temporal sampling and uncertainty-aware adaptive learning. The model actively selects the most discriminative temporal segments under EFE guidance, enabling dynamic observation and information gain maximization. Meanwhile, sample weighting driven by predictive uncertainty mitigates the effects of label noise and distribution shift. Experiments on the SMG dataset demonstrate the effectiveness of the proposed method, achieving consistent improvements across multiple mainstream backbones. Ablation studies confirm that both the EFE-guided observation and the adaptive learning mechanism are crucial to the performance gains. This work offers an interpretable and scalable paradigm for temporal behavior modeling under low-resource and noisy conditions, with broad applicability to wearable sensing, HCI, and clinical emotion monitoring.

Hanzhang Wang, Zonglin Liu, Jingyi Xu et al.

Proximal gradient algorithms (PGA), while foundational for inverse problems like image reconstruction, often yield unstable convergence and suboptimal solutions by violating the critical non-negativity constraint. We identify the gradient descent step as the root cause of this issue, which introduces negative values and induces high sensitivity to hyperparameters. To overcome these limitations, we propose a novel multiplicative update proximal gradient algorithm (SSO-PGA) with convergence guarantees, which is designed for robustness in non-negative inverse problems. Our key innovation lies in superseding the gradient descent step with a learnable sigmoid-based operator, which inherently enforces non-negativity and boundedness by transforming traditional subtractive updates into multiplicative ones. This design, augmented by a sliding parameter for enhanced stability and convergence, not only improves robustness but also boosts expressive capacity and noise immunity. We further formulate a degradation model for multi-modal restoration and derive its SSO-PGA-based optimization algorithm, which is then unfolded into a deep network to marry the interpretability of optimization with the power of deep learning. Extensive numerical and real-world experiments demonstrate that our method significantly surpasses traditional PGA and other state-of-the-art algorithms, ensuring superior performance and stability.

Chenyang Wang, Roger Olsson, Stefan Forsström et al.

Empowered by deep learning, semantic communication marks a paradigm shift from transmitting raw data to conveying task-relevant meaning, enabling more efficient and intelligent wireless systems. In this study, we explore a deep learning-based task-oriented communication framework that jointly considers classification performance, computational latency, and communication cost. We adopt ResNets-based models and evaluate them on the CIFAR-10 and CIFAR-100 datasets to simulate real-world classification tasks in wireless environments. We partition the model at various points to simulate split inference across a wireless channel. By varying the split location and the size of the transmitted semantic feature vector, we systematically analyze the trade-offs between task accuracy and resource efficiency. Experimental results show that, with appropriate model partitioning and semantic feature compression, the system can retain over 85\% of baseline accuracy while significantly reducing both computational load and communication overhead.

Weijia Feng, Yichen Zhu, Ruojia Zhang et al.

Owing to its rapid progress and broad application prospects, few-shot action recognition has attracted considerable interest. However, current methods are predominantly based on limited single-modal data, which does not fully exploit the potential of multimodal information. This paper presents a novel framework that actively identifies reliable modalities for each sample using task-specific contextual cues, thus significantly improving recognition performance. Our framework integrates an Active Sample Inference (ASI) module, which utilizes active inference to predict reliable modalities based on posterior distributions and subsequently organizes them accordingly. Unlike reinforcement learning, active inference replaces rewards with evidence-based preferences, making more stable predictions. Additionally, we introduce an active mutual distillation module that enhances the representation learning of less reliable modalities by transferring knowledge from more reliable ones. Adaptive multimodal inference is employed during the meta-test to assign higher weights to reliable modalities. Extensive experiments across multiple benchmarks demonstrate that our method significantly outperforms existing approaches.

Junjun Jiang, Chenyang Wang, Xianming Liu et al.

High-resolution (HR) hyperspectral face image plays an important role in face related computer vision tasks under uncontrolled conditions, such as low-light environment and spoofing attacks. However, the dense spectral bands of hyperspectral face images come at the cost of limited amount of photons reached a narrow spectral window on average, which greatly reduces the spatial resolution of hyperspectral face images. In this paper, we investigate how to adapt the deep learning techniques to hyperspectral face image super-resolution (HFSR), especially when the training samples are very limited. Benefiting from the amount of spectral bands, in which each band can be seen as an image, we present a spectral splitting and aggregation network (SSANet) for HFSR with limited training samples. In the shallow layers, we split the hyperspectral image into different spectral groups. Then, we gradually aggregate the neighbor bands at deeper layers to exploit spectral correlations. By this spectral splitting and aggregation strategy (SSAS), we can divide the original hyperspectral image into multiple samples (\emph{from less to more}) to support the efficient training of the network and effectively exploit the spectral correlations among spectrum. To cope with the challenge of small training sample size (S3) problem, we propose to expand the training samples by a self-representation model and symmetry-induced augmentation. Experiments show that SSANet can well model the joint correlations of spatial and spectral information. By expanding the training samples, SSANet can effectively alleviate the S3 problem.

Zongyong Deng, Hao Liu, Yaoxing Wang et al.

In this paper, we propose a progressive margin loss (PML) approach for unconstrained facial age classification. Conventional methods make strong assumption on that each class owns adequate instances to outline its data distribution, likely leading to bias prediction where the training samples are sparse across age classes. Instead, our PML aims to adaptively refine the age label pattern by enforcing a couple of margins, which fully takes in the in-between discrepancy of the intra-class variance, inter-class variance and class center. Our PML typically incorporates with the ordinal margin and the variational margin, simultaneously plugging in the globally-tuned deep neural network paradigm. More specifically, the ordinal margin learns to exploit the correlated relationship of the real-world age labels. Accordingly, the variational margin is leveraged to minimize the influence of head classes that misleads the prediction of tailed samples. Moreover, our optimization carefully seeks a series of indicator curricula to achieve robust and efficient model training. Extensive experimental results on three face aging datasets demonstrate that our PML achieves compelling performance compared to state of the arts. Code will be made publicly.

Xiaofei Wang, Yiwen Han, Chenyang Wang et al.

Recently, along with the rapid development of mobile communication technology, edge computing theory and techniques have been attracting more and more attentions from global researchers and engineers, which can significantly bridge the capacity of cloud and requirement of devices by the network edges, and thus can accelerate the content deliveries and improve the quality of mobile services. In order to bring more intelligence to the edge systems, compared to traditional optimization methodology, and driven by the current deep learning techniques, we propose to integrate the Deep Reinforcement Learning techniques and Federated Learning framework with the mobile edge systems, for optimizing the mobile edge computing, caching and communication. And thus, we design the "In-Edge AI" framework in order to intelligently utilize the collaboration among devices and edge nodes to exchange the learning parameters for a better training and inference of the models, and thus to carry out dynamic system-level optimization and application-level enhancement while reducing the unnecessary system communication load. "In-Edge AI" is evaluated and proved to have near-optimal performance but relatively low overhead of learning, while the system is cognitive and adaptive to the mobile communication systems. Finally, we discuss several related challenges and opportunities for unveiling a promising upcoming future of "In-Edge AI".