Hongzhi Zhang

Zehao Wang, Guanglei Yang, Yihan Zeng et al.

Federated fine-tuning of foundation models with Low-Rank Adaptation (LoRA) provides an efficient solution for reducing communication and computation costs while preserving data locality. However, the direct combination of FedAvg and LoRA suffers from three key issues: limited update space, which restricts the model's effective learning capacity; inter-round state mismatch, which disrupts cross-round local optimization continuity; and a client-agnostic starting state, which slows local convergence on clients. Although recent methods mitigate the limited update space issue by merging LoRA updates into the backbone across communication rounds, inter-round state mismatch and the client-agnostic starting state remain insufficiently addressed. To address these issues, we propose FedSmoothLoRA, a federated LoRA tuning framework that preserves the enlarged update space, improves cross-round local optimization continuity, and provides a client-aware starting state for local training. At each communication round, FedSmoothLoRA constructs the local LoRA initialization using two matrices: a Round-Matching matrix that preserves cross-round local state continuity, and a Gradient-Aligned matrix that provides client-specific optimization guidance from gradient signals estimated on local data. Together, these designs enable smoother and faster convergence. Extensive experiments on image classification and natural language generation tasks demonstrate that FedSmoothLoRA consistently outperforms existing federated LoRA tuning methods. Code: https://github.com/wangzehao0704/FedSmoothLoRA

Zhilu Zhang, Ruohao Wang, Hongzhi Zhang et al.

In this paper, we consider two challenging issues in reference-based super-resolution (RefSR), (i) how to choose a proper reference image, and (ii) how to learn real-world RefSR in a self-supervised manner. Particularly, we present a novel self-supervised learning approach for real-world image SR from observations at dual camera zooms (SelfDZSR). Considering the popularity of multiple cameras in modern smartphones, the more zoomed (telephoto) image can be naturally leveraged as the reference to guide the SR of the lesser zoomed (short-focus) image. Furthermore, SelfDZSR learns a deep network to obtain the SR result of short-focus image to have the same resolution as the telephoto image. For this purpose, we take the telephoto image instead of an additional high-resolution image as the supervision information and select a center patch from it as the reference to super-resolve the corresponding short-focus image patch. To mitigate the effect of the misalignment between short-focus low-resolution (LR) image and telephoto ground-truth (GT) image, we design an auxiliary-LR generator and map the GT to an auxiliary-LR while keeping the spatial position unchanged. Then the auxiliary-LR can be utilized to deform the LR features by the proposed adaptive spatial transformer networks (AdaSTN), and match the Ref features to GT. During testing, SelfDZSR can be directly deployed to super-solve the whole short-focus image with the reference of telephoto image. Experiments show that our method achieves better quantitative and qualitative performance against state-of-the-arts. Codes are available at https://github.com/cszhilu1998/SelfDZSR.

Renlong Wu, Zhilu Zhang, Shuohao Zhang et al.

Burst super-resolution (BurstSR) aims at reconstructing a high-resolution (HR) image from a sequence of low-resolution (LR) and noisy images, which is conducive to enhancing the imaging effects of smartphones with limited sensors. The main challenge of BurstSR is to effectively combine the complementary information from input frames, while existing methods still struggle with it. In this paper, we suggest fusing cues frame-by-frame with an efficient and flexible recurrent network. In particular, we emphasize the role of the base-frame and utilize it as a key prompt to guide the knowledge acquisition from other frames in every recurrence. Moreover, we introduce an implicit weighting loss to improve the model's flexibility in facing input frames with variable numbers. Extensive experiments on both synthetic and real-world datasets demonstrate that our method achieves better results than state-of-the-art ones. Codes and pre-trained models are available at https://github.com/ZcsrenlongZ/RBSR.

Wei Shang, Dongwei Ren, Yi Yang et al.

Natural videos captured by consumer cameras often suffer from low framerate and motion blur due to the combination of dynamic scene complexity, lens and sensor imperfection, and less than ideal exposure setting. As a result, computational methods that jointly perform video frame interpolation and deblurring begin to emerge with the unrealistic assumption that the exposure time is known and fixed. In this work, we aim ambitiously for a more realistic and challenging task - joint video multi-frame interpolation and deblurring under unknown exposure time. Toward this goal, we first adopt a variant of supervised contrastive learning to construct an exposure-aware representation from input blurred frames. We then train two U-Nets for intra-motion and inter-motion analysis, respectively, adapting to the learned exposure representation via gain tuning. We finally build our video reconstruction network upon the exposure and motion representation by progressive exposure-adaptive convolution and motion refinement. Extensive experiments on both simulated and real-world datasets show that our optimized method achieves notable performance gains over the state-of-the-art on the joint video x8 interpolation and deblurring task. Moreover, on the seemingly implausible x16 interpolation task, our method outperforms existing methods by more than 1.5 dB in terms of PSNR.

Pengju Liu, Hongzhi Zhang, Jinghui Wang et al.

Recently, deep learning-based image denoising methods have achieved promising performance on test data with the same distribution as training set, where various denoising models based on synthetic or collected real-world training data have been learned. However, when handling real-world noisy images, the denoising performance is still limited. In this paper, we propose a simple yet effective Bayesian deep ensemble (BDE) method for real-world image denoising, where several representative deep denoisers pre-trained with various training data settings can be fused to improve robustness. The foundation of BDE is that real-world image noises are highly signal-dependent, and heterogeneous noises in a real-world noisy image can be separately handled by different denoisers. In particular, we take well-trained CBDNet, NBNet, HINet, Uformer and GMSNet into denoiser pool, and a U-Net is adopted to predict pixel-wise weighting maps to fuse these denoisers. Instead of solely learning pixel-wise weighting maps, Bayesian deep learning strategy is introduced to predict weighting uncertainty as well as weighting map, by which prediction variance can be modeled for improving robustness on real-world noisy images. Extensive experiments have shown that real-world noises can be better removed by fusing existing denoisers instead of training a big denoiser with expensive cost. On DND dataset, our BDE achieves +0.28~dB PSNR gain over the state-of-the-art denoising method. Moreover, we note that our BDE denoiser based on different Gaussian noise levels outperforms state-of-the-art CBDNet when applying to real-world noisy images. Furthermore, our BDE can be extended to other image restoration tasks, and achieves +0.30dB, +0.18dB and +0.12dB PSNR gains on benchmark datasets for image deblurring, image deraining and single image super-resolution, respectively.

Qi Wang, Hongzhi Zhang, Jia Fu et al.

Despite the proliferation of powerful agentic models, the lack of critical post-training details hinders the development of strong counterparts in the open-source community. In this study, we present a comprehensive and fully open-source pipeline for training a high-performance agentic model for interacting with external tools and environments, named Klear-Qwen3-AgentForge, starting from the Qwen3-8B base model. We design effective supervised fine-tuning (SFT) with synthetic data followed by multi-turn reinforcement learning (RL) to unlock the potential for multiple diverse agentic tasks. We perform exclusive experiments on various agentic benchmarks in both tool use and coding domains. Klear-Qwen3-AgentForge-8B achieves state-of-the-art performance among LLMs of similar size and remains competitive with significantly larger models.

Yuanze Li, Yiwen Guo, Qizhang Li et al.

The study of multi-task learning has drawn great attention from the community. Despite the remarkable progress, the challenge of optimally learning different tasks simultaneously remains to be explored. Previous works attempt to modify the gradients from different tasks. Yet these methods give a subjective assumption of the relationship between tasks, and the modified gradient may be less accurate. In this paper, we introduce Stochastic Task Allocation~(STA), a mechanism that addresses this issue by a task allocation approach, in which each sample is randomly allocated a subset of tasks. For further progress, we propose Interleaved Stochastic Task Allocation~(ISTA) to iteratively allocate all tasks to each example during several consecutive iterations. We evaluate STA and ISTA on various datasets and applications: NYUv2, Cityscapes, and COCO for scene understanding and instance segmentation. Our experiments show both STA and ISTA outperform current state-of-the-art methods. The code will be available.

Zehao Wang, Yihan Zeng, Zidong Gong et al.

Post-training via Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) is crucial for enhancing reasoning in Multimodal Large Language Models (MLLMs), yet existing paradigms often reach a performance bottleneck due to the limitations of static data. While current methods leverage self-reflection or self-evolution to push these boundaries, they still suffer from cognitive drift and hallucinated reasoning paths caused by low-quality synthetic data. To address these challenges, we propose Anchor Evolution (AnE), a new paradigm that integrates truth-anchored data curation and model evolution, achieving faithful and steady performance gains at the reasoning frontier. Specifically, we propose Truth Anchor Expansion, which pinpoints the model failing frontier via trajectory rollouts and leverages ground-truth databases to retrieve high-fidelity anchors for faithful data curation. Subsequently, we introduce the Scaffold-Stripping Mechanism to internalize reasoning capabilities. This mechanism first anchors reasoning paths via scaffold-augmented supervision to mitigate the learning complexity and distribution drift of direct SFT on raw data, then leverages RL to strip the scaffold template, thereby effectively transitioning the reasoning paths into intrinsic model capabilities. Experimental results on multimodal reasoning benchmarks show that our method substantially advances the model performance frontier, improving the base model by 10.3\% across eight multimodal benchmarks and achieving state-of-the-art results. The code will be made publicly available.

Sirui Wang, Kaiwen Wei, Hongzhi Zhang et al.

Demonstration learning aims to guide the prompt prediction via providing answered demonstrations in the few shot settings. Despite achieving promising results, existing work only concatenates the answered examples as demonstrations to the prompt template (including the raw context) without any additional operation, neglecting the prompt-demonstration dependencies. Besides, prior research found that randomly replacing the labels of demonstrations marginally hurts performance, illustrating that the model could not properly learn the knowledge brought by the demonstrations. Inspired by the human learning process, in this paper, we introduce Imitation DEMOnstration Learning (Imitation-Demo) to strengthen demonstration learning via explicitly imitating human review behaviour, which includes: (1) contrastive learning mechanism to concentrate on the similar demonstrations. (2) demonstration-label re-prediction method to consolidate known knowledge. Experiment results show that our proposed method achieves state-of-the-art performance on 11 out of 14 classification corpora. Further studies also prove that Imitation-Demo strengthen the association between prompt and demonstrations, which could provide the basis for exploring how demonstration learning works.

Yupeng Zhang, Hongzhi Zhang, Sirui Wang et al.

A wide range of NLP tasks benefit from the fine-tuning of pretrained language models (PLMs). However, a number of redundant parameters which contribute less to the downstream task are observed in a directly fine-tuned model. We consider the gap between pretraining and downstream tasks hinders the training of these redundant parameters, and results in a suboptimal performance of the overall model. In this paper, we present PATS (Perturbation According To Sensitivity), a noisy training mechanism which considers each parameter's importance in the downstream task to help fine-tune PLMs. The main idea of PATS is to add bigger noise to parameters with lower sensitivity and vice versa, in order to activate more parameters' contributions to downstream tasks without affecting the sensitive ones much. Extensive experiments conducted on different tasks of the GLUE benchmark show PATS can consistently empower the fine-tuning of different sizes of PLMs, and the parameters in the well-performing models always have more concentrated distributions of sensitivities, which experimentally proves the effectiveness of our method.

Hongzhi Zhang, Yuanze Hu, Tinghai Zhang et al.

The evolution of Large Language Models (LLMs) towards autonomous agents has catalyzed progress in Deep Research. While retrieval capabilities are well-benchmarked, the post-retrieval synthesis stage--where agents must digest massive amounts of context and consolidate fragmented evidence into coherent, long-form reports--remains under-evaluated due to the subjectivity of open-ended writing. To bridge this gap, we introduce DeepSynth-Eval, a benchmark designed to objectively evaluate information consolidation capabilities. We leverage high-quality survey papers as gold standards, reverse-engineering research requests and constructing "Oracle Contexts" from their bibliographies to isolate synthesis from retrieval noise. We propose a fine-grained evaluation protocol using General Checklists (for factual coverage) and Constraint Checklists (for structural organization), transforming subjective judgment into verifiable metrics. Experiments across 96 tasks reveal that synthesizing information from hundreds of references remains a significant challenge. Our results demonstrate that agentic plan-and-write workflows significantly outperform single-turn generation, effectively reducing hallucinations and improving adherence to complex structural constraints.

Zihao Wang, Yuxiang Wei, Xinpeng Zhou et al.

Text-to-image generation has advanced rapidly, yet it still struggles to capture the nuanced user preferences. Existing approaches typically rely on multimodal large language models to infer user preferences, but the derived prompts or latent codes rarely reflect them faithfully, leading to suboptimal personalization. We present Premier, a novel preference modulation framework for personalized image generation. Premier represents each user's preference as a learnable embedding and introduces a preference adapter that fuses the user embedding with the text prompt. To enable accurate and fine-grained preference control, the fused preference embedding is further used to modulate the generative process. To enhance the distinctness of individual preference and improve alignment between outputs and user-specific styles, we incorporate a dispersion loss that enforces separation among user embeddings. When user data are scarce, new users are represented as linear combinations of existing preference embeddings learned during training, enabling effective generalization. Experiments show that Premier outperforms prior methods under the same history length, achieving stronger preference alignment and superior performance on text consistency, ViPer proxy metrics, and expert evaluations.

Hongzhi Zhang, Jia Fu, Jingyuan Zhang et al.

Reinforcement learning (RL) for large language models is an energy-intensive endeavor: training can be unstable, and the policy may gradually drift away from its pretrained weights. We present \emph{RLEP}\, -- \,Reinforcement Learning with Experience rePlay\, -- \,a two-phase framework that first collects verified trajectories and then replays them during subsequent training. At every update step, the policy is optimized on mini-batches that blend newly generated rollouts with these replayed successes. By replaying high-quality examples, RLEP steers the model away from fruitless exploration, focuses learning on promising reasoning paths, and delivers both faster convergence and stronger final performance. On the Qwen2.5-Math-7B base model, RLEP reaches baseline peak accuracy with substantially fewer updates and ultimately surpasses it, improving accuracy on AIME-2024 from 38.2% to 39.9%, on AIME-2025 from 19.8% to 22.3%, and on AMC-2023 from 77.0% to 82.2%. Our code, datasets, and checkpoints are publicly available at https://github.com/Kwai-Klear/RLEP to facilitate reproducibility and further research.

Yida Xiong, Kun Li, Jiameng Chen et al.

Molecular optimization (MO) is a crucial stage in drug discovery in which task-oriented generated molecules are optimized to meet practical industrial requirements. Existing mainstream MO approaches primarily utilize external property predictors to guide iterative property optimization. However, learning all molecular samples in the vast chemical space is unrealistic for predictors. As a result, errors and noise are inevitably introduced during property prediction due to the nature of approximation. This leads to discrepancy accumulation, generalization reduction and suboptimal molecular candidates. In this paper, we propose a text-guided multi-property molecular optimization method utilizing transformer-based diffusion language model (TransDLM). TransDLM leverages standardized chemical nomenclature as semantic representations of molecules and implicitly embeds property requirements into textual descriptions, thereby mitigating error propagation during diffusion process. By fusing physically and chemically detailed textual semantics with specialized molecular representations, TransDLM effectively integrates diverse information sources to guide precise optimization, which enhances the model's ability to balance structural retention and property enhancement. Additionally, the success of a case study further demonstrates TransDLM's ability to solve practical problems. Experimentally, our approach surpasses state-of-the-art methods in maintaining molecular structural similarity and enhancing chemical properties on the benchmark dataset. The code is available at: https://github.com/Cello2195/TransDLM.

Zhilu Zhang, Ruohao Wang, Hongzhi Zhang et al.

In this paper, we consider two challenging issues in reference-based super-resolution (RefSR) for smartphone, (i) how to choose a proper reference image, and (ii) how to learn RefSR in a self-supervised manner. Particularly, we propose a novel self-supervised learning approach for real-world RefSR from observations at dual and multiple camera zooms. Firstly, considering the popularity of multiple cameras in modern smartphones, the more zoomed (telephoto) image can be naturally leveraged as the reference to guide the super-resolution (SR) of the lesser zoomed (ultra-wide) image, which gives us a chance to learn a deep network that performs SR from the dual zoomed observations (DZSR). Secondly, for self-supervised learning of DZSR, we take the telephoto image instead of an additional high-resolution image as the supervision information, and select a center patch from it as the reference to super-resolve the corresponding ultra-wide image patch. To mitigate the effect of the misalignment between ultra-wide low-resolution (LR) patch and telephoto ground-truth (GT) image during training, we first adopt patch-based optical flow alignment and then design an auxiliary-LR to guide the deforming of the warped LR features. To generate visually pleasing results, we present local overlapped sliced Wasserstein loss to better represent the perceptual difference between GT and output in the feature space. During testing, DZSR can be directly deployed to super-solve the whole ultra-wide image with the reference of the telephoto image. In addition, we further take multiple zoomed observations to explore self-supervised RefSR, and present a progressive fusion scheme for the effective utilization of reference images. Experiments show that our methods achieve better quantitative and qualitative performance against state-of-the-arts. Codes are available at https://github.com/cszhilu1998/SelfDZSR_PlusPlus.

Hongzhi Zhang, Xiuwen Gong, Shirui Pan et al.

Drug-target interaction (DTI) prediction is a critical component of the drug discovery process. In the drug development engineering field, predicting novel drug-target interactions is extremely crucial.However, although existing methods have achieved high accuracy levels in predicting known drugs and drug targets, they fail to utilize global protein information during DTI prediction. This leads to an inability to effectively predict interaction the interactions between novel drugs and their targets. As a result, the cross-field information fusion strategy is employed to acquire local and global protein information. Thus, we propose the siamese drug-target interaction SiamDTI prediction method, which utilizes a double channel network structure for cross-field supervised learning.Experimental results on three benchmark datasets demonstrate that SiamDTI achieves higher accuracy levels than other state-of-the-art (SOTA) methods on novel drugs and targets.Additionally, SiamDTI's performance with known drugs and targets is comparable to that of SOTA approachs. The code is available at https://anonymous.4open.science/r/DDDTI-434D.

Zhuo Zhang, Jingyuan Zhang, Jintao Huang et al.

Instruction tuning has been identified as a crucial technique for optimizing the performance of large language models (LLMs) in generating human-aligned responses. Nonetheless, gathering diversified and superior-quality instruction data for such tuning presents notable obstacles, especially in domains with rigid privacy provisions. Federated instruction tuning (FedIT) has emerged as a promising solution, by consolidating collaborative training across multiple data owners, thereby resulting in a privacy-preserving learning model. However, FedIT encounters limitations such as scarcity of instructional data and risk of exposure to training data extraction attacks. In this paper, we propose a novel federated algorithm, FewFedPIT, designed to simultaneously enhance privacy protection and model performance of federated few-shot learning. FewFedPITcomprises three vital components on the client side: (1) synthetic data generation, which utilizes LLMs' in-context learning capacity to generate synthetic data autonomously, thus expanding the local database; (2) parameter isolation training, which individually updates the public parameters in the synthetic data and the private parameters in the local data, consequently mitigating the noise impact of the synthetic data; (3) local aggregation sharing, which mixes public and private parameters before uploading, effectively preventing data extraction attacks. Extensive experiments on three open-source datasets demonstrate the effectiveness of FewFedPITin, enhancing privacy preservation and improving federated few-shot performance.

Jingyuan Zhang, Hongzhi Zhang, Zhou Haonan et al.

Data curation plays a crucial role in training powerful Visual Language Models (VLMs). In this work, we introduce the concept of Data Metabolism and present our data-centric framework to build VLMs throughout the development lifecycle. Starting from a standard model architecture, we discuss and provide insights into two crucial development steps: data curation and iteration, forming a closed-loop system that continuously improves model performance. We show a detailed codebook on how to process existing massive datasets and build user-specific data flywheel. As a demonstration, we release a VLM, named Capybara-VL, which excels in typical multimodal tasks (e.g. , visual question answering, scientific reasoning, and text-rich tasks). Despite its relatively compact size, Capybara-VL surpasses several open-source models that are up to 10 times larger in size. Moreover, it achieves results that are on par with those of several leading proprietary models, demonstrating its remarkable competitiveness. These results highlight the power of our data-centric framework and the potential of training smaller and more efficient VLMs.

Jia Fu, Xinyu Yang, Hongzhi Zhang et al.

Precise, correct feedback is crucial for effectively training large language models (LLMs) in code reinforcement learning. However, synthesizing high-quality test cases remains a profoundly challenging and unsolved problem. In this work, we present Klear-CodeTest, a comprehensive test case synthesis framework featuring rigorous verification to ensure quality and reliability of test cases. Our approach achieves broad coverage of programming problems via a novel Generator-Validation (G-V) framework, ensuring correctness through a consistency validation mechanism that verifies outputs against gold solutions. The proposed G-V framework generates comprehensive test cases including both regular and corner cases, enhancing test coverage and discriminative power for solution correctness assessment in code reinforcement learning. In addition, we design a multi-layered security sandbox system optimized for online verification platforms, guaranteeing safe and reliable code execution. Through comprehensive experiments, we demonstrate the effectiveness of our curated dataset, showing significant improvements in model performance and training stability. The source codes, curated dataset and sandbox system are available at: https://github.com/Kwai-Klear/CodeTest.

Hongzhi Zhang, Zhonglie Liu, Kun Meng et al.

Given the vastness of chemical space and the ongoing emergence of previously uncharacterized proteins, zero-shot compound-protein interaction (CPI) prediction better reflects the practical challenges and requirements of real-world drug development. Although existing methods perform adequately during certain CPI tasks, they still face the following challenges: (1) Representation learning from local or complete protein sequences often overlooks the complex interdependencies between subsequences, which are essential for predicting spatial structures and binding properties. (2) Dependence on large-scale or scarce multimodal protein datasets demands significant training data and computational resources, limiting scalability and efficiency. To address these challenges, we propose a novel approach that pretrains protein representations for CPI prediction tasks using subsequence reordering, explicitly capturing the dependencies between protein subsequences. Furthermore, we apply length-variable protein augmentation to ensure excellent pretraining performance on small training datasets. To evaluate the model's effectiveness and zero-shot learning ability, we combine it with various baseline methods. The results demonstrate that our approach can improve the baseline model's performance on the CPI task, especially in the challenging zero-shot scenario. Compared to existing pre-training models, our model demonstrates superior performance, particularly in data-scarce scenarios where training samples are limited. Our implementation is available at https://github.com/Hoch-Zhang/PSRP-CPI.

Linhao Yu, Xinguang Ji, Yahui Liu et al.

Video captioning can be used to assess the video understanding capabilities of Multimodal Large Language Models (MLLMs). However, existing benchmarks and evaluation protocols suffer from crucial issues, such as inadequate or homogeneous creation of key points, exorbitant cost of data creation, and limited evaluation scopes. To address these issues, we propose an automatic framework, named AutoCaption, which leverages Monte Carlo Tree Search (MCTS) to construct numerous and diverse descriptive sentences (\textit{i.e.}, key points) that thoroughly represent video content in an iterative way. This iterative captioning strategy enables the continuous enhancement of video details such as actions, objects' attributes, environment details, etc. We apply AutoCaption to curate MCTS-VCB, a fine-grained video caption benchmark covering video details, thereby enabling a comprehensive evaluation of MLLMs on the video captioning task. We evaluate more than 20 open- and closed-source MLLMs of varying sizes on MCTS-VCB. Results show that MCTS-VCB can effectively and comprehensively evaluate the video captioning capability, with Gemini-1.5-Pro achieving the highest F1 score of 71.2. Interestingly, we fine-tune InternVL2.5-8B with the AutoCaption-generated data, which helps the model achieve an overall improvement of 25.0% on MCTS-VCB and 16.3% on DREAM-1K, further demonstrating the effectiveness of AutoCaption. The code and data are available at https://github.com/tjunlp-lab/MCTS-VCB.

Xingguang Ji, Jiakang Wang, Hongzhi Zhang et al.

With the development of Multimodal Large Language Models (MLLMs), numerous outstanding accomplishments have emerged within the open-source community. Due to the complexity of creating and training multimodal data pairs, it is still a computational and time-consuming process to build powerful MLLMs. In this work, we introduce Capybara-OMNI, an MLLM that trains in a lightweight and efficient manner and supports understanding text, image, video, and audio modalities. We present in detail the framework design, the data construction, and the training recipe, to develop an MLLM step-by-step to obtain competitive performance. We also provide exclusive benchmarks utilized in our experiments to show how to properly verify understanding capabilities across different modalities. Results show that by following our guidance, we can efficiently build an MLLM that achieves competitive performance among models of the same scale on various multimodal benchmarks. Additionally, to enhance the multimodal instruction following and conversational capabilities of the model, we further discuss how to train the chat version upon an MLLM understanding model, which is more in line with user habits for tasks like real-time interaction with humans. We publicly disclose the Capybara-OMNI model, along with its chat-based version. The disclosure includes both the model weights, a portion of the training data, and the inference codes, which are made available on GitHub.

Jiameng Chen, Yida Xiong, Kun Li et al.

Computational antibody design holds immense promise for therapeutic discovery, yet existing generative models are fundamentally limited by two core challenges: (i) a lack of dynamical consistency, which yields physically implausible structures, and (ii) poor generalization due to data scarcity and structural bias. We introduce FP-AbDiff, the first antibody generator to enforce Fokker-Planck Equation (FPE) physics along the entire generative trajectory. Our method minimizes a novel FPE residual loss over the mixed manifold of CDR geometries (R^3 x SO(3)), compelling locally-learned denoising scores to assemble into a globally coherent probability flow. This physics-informed regularizer is synergistically integrated with deep biological priors within a state-of-the-art SE(3)-equivariant diffusion framework. Rigorous evaluation on the RAbD benchmark confirms that FP-AbDiff establishes a new state-of-the-art. In de novo CDR-H3 design, it achieves a mean Root Mean Square Deviation of 0.99 Å when superposing on the variable region, a 25% improvement over the previous state-of-the-art model, AbX, and the highest reported Contact Amino Acid Recovery of 39.91%. This superiority is underscored in the more challenging six-CDR co-design task, where our model delivers consistently superior geometric precision, cutting the average full-chain Root Mean Square Deviation by ~15%, and crucially, achieves the highest full-chain Amino Acid Recovery on the functionally dominant CDR-H3 loop (45.67%). By aligning generative dynamics with physical laws, FP-AbDiff enhances robustness and generalizability, establishing a principled approach for physically faithful and functionally viable antibody design.

Kun Li, Longtao Hu, Yida Xiong et al.

Molecular representation learning aims to learn vector embeddings that capture molecular structure and geometry, thereby enabling property prediction and downstream scientific applications. In many AI for science tasks, labeled data are expensive to obtain and therefore limited in availability. Under the few-shot setting, models trained with scarce supervision often learn brittle structure-property relationships, resulting in substantially higher prediction errors and reduced generalization to unseen molecules. To address this limitation, we propose PCEvo, a path-consistent representation method that learns from virtual paths through dynamic structural evolution. PCEvo enumerates multiple chemically feasible edit paths between retrieved similar molecular pairs under topological dependency constraints. It transforms the labels of the two molecules into stepwise supervision along each virtual evolutionary path. It introduces a path-consistency objective that enforces prediction invariance across alternative paths connecting the same two molecules. Comprehensive experiments on the QM9 and MoleculeNet datasets demonstrate that PCEvo substantially improves the few-shot generalization performance of baseline methods. The code is available at https://anonymous.4open.science/r/PCEvo-4BF2.

Xiaoxue Cheng, Junyi Li, Wayne Xin Zhao et al.

Hallucination detection is a challenging task for large language models (LLMs), and existing studies heavily rely on powerful closed-source LLMs such as GPT-4. In this paper, we propose an autonomous LLM-based agent framework, called HaluAgent, which enables relatively smaller LLMs (e.g. Baichuan2-Chat 7B) to actively select suitable tools for detecting multiple hallucination types such as text, code, and mathematical expression. In HaluAgent, we integrate the LLM, multi-functional toolbox, and design a fine-grained three-stage detection framework along with memory mechanism. To facilitate the effectiveness of HaluAgent, we leverage existing Chinese and English datasets to synthesize detection trajectories for fine-tuning, which endows HaluAgent with the capability for bilingual hallucination detection. Extensive experiments demonstrate that only using 2K samples for tuning LLMs, HaluAgent can perform hallucination detection on various types of tasks and datasets, achieving performance comparable to or even higher than GPT-4 without tool enhancements on both in-domain and out-of-domain datasets. We release our dataset and code at https://github.com/RUCAIBox/HaluAgent.

Chenxi Sun, Hongzhi Zhang, Qi Wang et al.

Instruction-based Image Editing (IIE) models have made significantly improvement due to the progress of multimodal large language models (MLLMs) and diffusion models, which can understand and reason about complex editing instructions. In addition to advancing current IIE models, accurately evaluating their output has become increasingly critical and challenging. Current IIE evaluation methods and their evaluation procedures often fall short of aligning with human judgment and often lack explainability. To address these limitations, we propose JUdgement through Routing of Expertise (JURE). Each expert in JURE is a pre-selected model assumed to be equipped with an atomic expertise that can provide useful feedback to judge output, and the router dynamically routes the evaluation task of a given instruction and its output to appropriate experts, aggregating their feedback into a final judge. JURE is trustworthy in two aspects. First, it can effortlessly provide explanations about its judge by examining the routed experts and their feedback. Second, experimental results demonstrate that JURE is reliable by achieving superior alignment with human judgments, setting a new standard for automated IIE evaluation. Moreover, JURE's flexible design is future-proof - modular experts can be seamlessly replaced or expanded to accommodate advancements in IIE, maintaining consistently high evaluation quality. Our evaluation data and results are available at https://github.com/Cyyyyyrus/JURE.git.

Yuxiang Wei, Zhilong Ji, Jinfeng Bai et al.

Text-to-image (T2I) diffusion models have shown significant success in personalized text-to-image generation, which aims to generate novel images with human identities indicated by the reference images. Despite promising identity fidelity has been achieved by several tuning-free methods, they usually suffer from overfitting issues. The learned identity tends to entangle with irrelevant information, resulting in unsatisfied text controllability, especially on faces. In this work, we present MasterWeaver, a test-time tuning-free method designed to generate personalized images with both faithful identity fidelity and flexible editability. Specifically, MasterWeaver adopts an encoder to extract identity features and steers the image generation through additional introduced cross attention. To improve editability while maintaining identity fidelity, we propose an editing direction loss for training, which aligns the editing directions of our MasterWeaver with those of the original T2I model. Additionally, a face-augmented dataset is constructed to facilitate disentangled identity learning, and further improve the editability. Extensive experiments demonstrate that our MasterWeaver can not only generate personalized images with faithful identity, but also exhibit superiority in text controllability. Our code can be found at https://github.com/csyxwei/MasterWeaver.

Jingbo Lin, Zhilu Zhang, Wenbo Li et al.

Recently, considerable progress has been made in all-in-one image restoration. Generally, existing methods can be degradation-agnostic or degradation-aware. However, the former are limited in leveraging degradation-specific restoration, and the latter suffer from the inevitable error in degradation estimation. Consequently, the performance of existing methods has a large gap compared to specific single-task models. In this work, we make a step forward in this topic, and present our UniRestorer with improved restoration performance. Specifically, we perform hierarchical clustering on degradation space, and train a multi-granularity mixture-of-experts (MoE) restoration model. Then, UniRestorer adopts both degradation and granularity estimation to adaptively select an appropriate expert for image restoration. In contrast to existing degradation-agnostic and -aware methods, UniRestorer can leverage degradation estimation to benefit degradation specific restoration, and use granularity estimation to make the model robust to degradation estimation error. Experimental results show that our UniRestorer outperforms state-of-the-art all-in-one methods by a large margin, and is promising in closing the performance gap to specific single task models.

Hongzhi Zhang, Jingyuan Zhang, Xingguang Ji et al.

Typical video modeling methods, such as LLava, represent videos as sequences of visual tokens, which are then processed by the LLM backbone for effective video understanding. However, this approach leads to a massive number of visual tokens, especially for long videos. A practical solution is to first extract relevant visual information from the large visual context before feeding it into the LLM backbone, thereby reducing computational overhead. In this work, we introduce DynTok, a novel \textbf{Dyn}amic video \textbf{Tok}en compression strategy. DynTok adaptively splits visual tokens into groups and merges them within each group, achieving high compression in regions with low information density while preserving essential content. Our method reduces the number of tokens to 44.4% of the original size while maintaining comparable performance. It further benefits from increasing the number of video frames and achieves 65.3% on Video-MME and 72.5% on MLVU. By applying this simple yet effective compression method, we expose the redundancy in video token representations and offer insights for designing more efficient video modeling techniques.

Haoze Zhang, Tianyu Huang, Zichen Wan et al.

While recent video generation models have achieved significant visual fidelity, they often suffer from the lack of explicit physical controllability and plausibility. To address this, some recent studies attempted to guide the video generation with physics-based rendering. However, these methods face inherent challenges in accurately modeling complex physical properties and effectively control ling the resulting physical behavior over extended temporal sequences. In this work, we introduce PhysChoreo, a novel framework that can generate videos with diverse controllability and physical realism from a single image. Our method consists of two stages: first, it estimates the static initial physical properties of all objects in the image through part-aware physical property reconstruction. Then, through temporally instructed and physically editable simulation, it synthesizes high-quality videos with rich dynamic behaviors and physical realism. Experimental results show that PhysChoreo can generate videos with rich behaviors and physical realism, outperforming state-of-the-art methods on multiple evaluation metrics.

Yida Xiong, Jiameng Chen, Kun Li et al.

Molecular graph generation is essentially a classification generation problem, aimed at predicting categories of atoms and bonds. Currently, prevailing paradigms such as continuous diffusion models are trained to predict continuous numerical values, treating the training process as a regression task. However, the final generation necessitates a rounding step to convert these predictions back into discrete classification categories, which is intrinsically a classification operation. Given that the rounding operation is not incorporated during training, there exists a significant discrepancy between the model's training objective and its inference procedure. As a consequence, an excessive emphasis on point-wise precision can lead to overfitting and inefficient learning. This occurs because considerable efforts are devoted to capturing intra-bin variations that are ultimately irrelevant to the discrete nature of the task at hand. Such a flaw results in diminished molecular diversity and constrains the model's generalization capabilities. To address this fundamental limitation, we propose GraphBFN, a novel hierarchical coarse-to-fine framework based on Bayesian Flow Networks that operates on the parameters of distributions. By innovatively introducing Cumulative Distribution Function, GraphBFN is capable of calculating the probability of selecting the correct category, thereby unifying the training objective with the sampling rounding operation. We demonstrate that our method achieves superior performance and faster generation, setting new state-of-the-art results on the QM9 and ZINC250k molecular graph generation benchmarks.

Kun Li, Zhennan Wu, Yida Xiong et al.

Drug discovery is of great social significance in safeguarding human health, prolonging life, and addressing the challenges of major diseases. In recent years, artificial intelligence has demonstrated remarkable advantages in key tasks across bioinformatics and pharmacology, owing to its efficient data processing and data representation capabilities. However, most existing computational platforms cover only a subset of core tasks, leading to fragmented workflows and low efficiency. In addition, they often lack algorithmic innovation and show poor generalization to out-of-distribution (OOD) data, which greatly hinders the progress of drug discovery. To address these limitations, we propose Baishenglai (BSL), a deep learning-enhanced, open-access platform designed for virtual drug discovery. BSL integrates seven core tasks within a unified and modular framework, incorporating advanced technologies such as generative models and graph neural networks. In addition to achieving state-of-the-art (SOTA) performance on multiple benchmark datasets, the platform emphasizes evaluation mechanisms that focus on generalization to OOD molecular structures. Comparative experiments with existing platforms and baseline methods demonstrate that BSL provides a comprehensive, scalable, and effective solution for virtual drug discovery, offering both algorithmic innovation and high-precision prediction for real-world pharmaceutical research. In addition, BSL demonstrated its practical utility by discovering novel modulators of the GluN1/GluN3A NMDA receptor, successfully identifying three compounds with clear bioactivity in in-vitro electrophysiological assays. These results highlight BSL as a promising and comprehensive platform for accelerating biomedical research and drug discovery. The platform is accessible at https://www.baishenglai.net.

Lingyu Kong, Hongzhi Zhang, Jingyuan Zhang et al.

Current vision-language models (VLMs) have demonstrated remarkable capabilities across diverse video understanding applications. Designing VLMs for video inputs requires effectively modeling the temporal dimension (i.e. capturing dependencies across frames) and balancing the processing of short and long videos. Specifically, short videos demand preservation of fine-grained details, whereas long videos require strategic compression of visual information to handle extensive temporal contexts efficiently. However, our empirical analysis reveals a critical limitation: most existing VLMs suffer severe performance degradation in long video understanding tasks when compressing visual tokens below a quarter of their original visual tokens. To enable more effective modeling of both short and long video inputs, we propose Clapper, a method that utilizes a slow-fast strategy for video representation and introduces a novel module named TimePerceiver for efficient temporal-spatial encoding within existing VLM backbones. By using our method, we achieves 13x compression of visual tokens per frame (averaging 61 tokens/frame) without compromising QA accuracy. In our experiments, Clapper achieves 62.0% on VideoMME, 69.8% on MLVU, and 67.4% on TempCompass, all with fewer than 6,000 visual tokens per video. The code will be publicly available on the homepage.

Fanheng Kong, Jingyuan Zhang, Yahui Liu et al.

Multimodal information retrieval (MIR) faces inherent challenges due to the heterogeneity of data sources and the complexity of cross-modal alignment. While previous studies have identified modal gaps in feature spaces, a systematic approach to address these challenges remains unexplored. In this work, we introduce UNITE, a universal framework that tackles these challenges through two critical yet underexplored aspects: data curation and modality-aware training configurations. Our work provides the first comprehensive analysis of how modality-specific data properties influence downstream task performance across diverse scenarios. Moreover, we propose Modal-Aware Masked Contrastive Learning (MAMCL) to mitigate the competitive relationships among the instances of different modalities. Our framework achieves state-of-the-art results on multiple multimodal retrieval benchmarks, outperforming existing methods by notable margins. Through extensive experiments, we demonstrate that strategic modality curation and tailored training protocols are pivotal for robust cross-modal representation learning. This work not only advances MIR performance but also provides a foundational blueprint for future research in multimodal systems. Our project is available at https://friedrichor.github.io/projects/UNITE.

Heming Xu, Xiaohui Liu, Zhilu Zhang et al.

Real-world image de-weathering aims at removingvarious undesirable weather-related artifacts, e.g., rain, snow,and fog. To this end, acquiring ideal training pairs is crucial.Existing real-world datasets are typically constructed paired databy extracting clean and degraded images from live streamsof landscape scene on the Internet. Despite the use of strictfiltering mechanisms during collection, training pairs inevitablyencounter inconsistency in terms of lighting, object position, scenedetails, etc, making de-weathering models possibly suffer fromdeformation artifacts under non-ideal supervision. In this work,we propose a unified solution for real-world image de-weatheringwith non-ideal supervision, i.e., a pseudo-label guided learningframework, to address various inconsistencies within the realworld paired dataset. Generally, it consists of a de-weatheringmodel (De-W) and a Consistent Label Constructor (CLC), bywhich restoration result can be adaptively supervised by originalground-truth image to recover sharp textures while maintainingconsistency with the degraded inputs in non-weather contentthrough the supervision of pseudo-labels. Particularly, a Crossframe Similarity Aggregation (CSA) module is deployed withinCLC to enhance the quality of pseudo-labels by exploring thepotential complementary information of multi-frames throughgraph model. Moreover, we introduce an Information AllocationStrategy (IAS) to integrate the original ground-truth imagesand pseudo-labels, thereby facilitating the joint supervision forthe training of de-weathering model. Extensive experimentsdemonstrate that our method exhibits significant advantageswhen trained on imperfectly aligned de-weathering datasets incomparison with other approaches.

Kun Li, Yida Xiong, Hongzhi Zhang et al.

Due to their excellent drug-like and pharmacokinetic properties, small molecule drugs are widely used to treat various diseases, making them a critical component of drug discovery. In recent years, with the rapid development of deep learning (DL) techniques, DL-based small molecule drug discovery methods have achieved excellent performance in prediction accuracy, speed, and complex molecular relationship modeling compared to traditional machine learning approaches. These advancements enhance drug screening efficiency and optimization and provide more precise and effective solutions for various drug discovery tasks. Contributing to this field's development, this paper aims to systematically summarize and generalize the recent key tasks and representative techniques in graph-structured small molecule drug discovery in recent years. Specifically, we provide an overview of the major tasks in small molecule drug discovery and their interrelationships. Next, we analyze the six core tasks, summarizing the related methods, commonly used datasets, and technological development trends. Finally, we discuss key challenges, such as interpretability and out-of-distribution generalization, and offer our insights into future research directions for small molecule drug discovery.

Kaiwen Wei, Jingyuan Zhang, Hongzhi Zhang et al.

Large Language Models (LLMs) exhibit remarkable generative capabilities, enabling the generation of valuable information. Despite these advancements, previous research found that LLMs sometimes struggle with adhering to specific constraints (e.g., in specific place or at specific time), at times even overlooking them, which leads to responses that are either too generic or not fully satisfactory. Existing approaches attempted to address this issue by decomposing or rewriting input instructions, yet they fall short in adequately emphasizing specific constraints and in unlocking the underlying knowledge (e.g., programming within the context of software development). In response, this paper proposes a simple yet effective method named Chain-of-Specificity (CoS). Specifically, CoS iteratively emphasizes the specific constraints in the input instructions, unlocks knowledge within LLMs, and refines responses. Experiments conducted on publicly available and self-build complex datasets demonstrate that CoS outperforms existing methods in enhancing generated content especially for the specificity. Besides, as the number of specific constraints increase, other baselines falter, while CoS still performs well. Moreover, we show that distilling responses generated by CoS effectively enhances the ability of smaller models to follow the constrained instructions. Resources of this paper will be released for further research.

Zhaoyi Yan, Ruimao Zhang, Hongzhi Zhang et al.

Crowd counting is critical for numerous video surveillance scenarios. One of the main issues in this task is how to handle the dramatic scale variations of pedestrians caused by the perspective effect. To address this issue, this paper proposes a novel convolution neural network-based crowd counting method, termed Perspective-guided Fractional-Dilation Network (PFDNet). By modeling the continuous scale variations, the proposed PFDNet is able to select the proper fractional dilation kernels for adapting to different spatial locations. It significantly improves the flexibility of the state-of-the-arts that only consider the discrete representative scales. In addition, by avoiding the multi-scale or multi-column architecture that used in other methods, it is computationally more efficient. In practice, the proposed PFDNet is constructed by stacking multiple Perspective-guided Fractional-Dilation Convolutions (PFC) on a VGG16-BN backbone. By introducing a novel generalized dilation convolution operation, the PFC can handle fractional dilation ratios in the spatial domain under the guidance of perspective annotations, achieving continuous scales modeling of pedestrians. To deal with the problem of unavailable perspective information in some cases, we further introduce an effective perspective estimation branch to the proposed PFDNet, which can be trained in either supervised or weakly-supervised setting once the branch has been pre-trained. Extensive experiments show that the proposed PFDNet outperforms state-of-the-art methods on ShanghaiTech A, ShanghaiTech B, WorldExpo'10, UCF-QNRF, UCF_CC_50 and TRANCOS dataset, achieving MAE 53.8, 6.5, 6.8, 84.3, 205.8, and 3.06 respectively.

Jiazhi Du, Xin Qiao, Zifei Yan et al.

For flexible non-blind image denoising, existing deep networks usually take both noisy image and noise level map as the input to handle various noise levels with a single model. However, in this kind of solution, the noise variance (i.e., noise level) is only deployed to modulate the first layer of convolution feature with channel-wise shifting, which is limited in balancing noise removal and detail preservation. In this paper, we present a novel flexible image enoising network (CFMNet) by equipping an U-Net backbone with multi-layer conditional feature modulation (CFM) modules. In comparison to channel-wise shifting only in the first layer, CFMNet can make better use of noise level information by deploying multiple layers of CFM. Moreover, each CFM module takes onvolutional features from both noisy image and noise level map as input for better trade-off between noise removal and detail preservation. Experimental results show that our CFMNet is effective in exploiting noise level information for flexible non-blind denoising, and performs favorably against the existing deep image denoising methods in terms of both quantitative metrics and visual quality.

Pengju Liu, Hongzhi Zhang, Wei Lian et al.

In computer vision, convolutional networks (CNNs) often adopts pooling to enlarge receptive field which has the advantage of low computational complexity. However, pooling can cause information loss and thus is detrimental to further operations such as features extraction and analysis. Recently, dilated filter has been proposed to trade off between receptive field size and efficiency. But the accompanying gridding effect can cause a sparse sampling of input images with checkerboard patterns. To address this problem, in this paper, we propose a novel multi-level wavelet CNN (MWCNN) model to achieve better trade-off between receptive field size and computational efficiency. The core idea is to embed wavelet transform into CNN architecture to reduce the resolution of feature maps while at the same time, increasing receptive field. Specifically, MWCNN for image restoration is based on U-Net architecture, and inverse wavelet transform (IWT) is deployed to reconstruct the high resolution (HR) feature maps. The proposed MWCNN can also be viewed as an improvement of dilated filter and a generalization of average pooling, and can be applied to not only image restoration tasks, but also any CNNs requiring a pooling operation. The experimental results demonstrate effectiveness of the proposed MWCNN for tasks such as image denoising, single image super-resolution, JPEG image artifacts removal and object classification.

Pengju Liu, Hongzhi Zhang, Kai Zhang et al.

The tradeoff between receptive field size and efficiency is a crucial issue in low level vision. Plain convolutional networks (CNNs) generally enlarge the receptive field at the expense of computational cost. Recently, dilated filtering has been adopted to address this issue. But it suffers from gridding effect, and the resulting receptive field is only a sparse sampling of input image with checkerboard patterns. In this paper, we present a novel multi-level wavelet CNN (MWCNN) model for better tradeoff between receptive field size and computational efficiency. With the modified U-Net architecture, wavelet transform is introduced to reduce the size of feature maps in the contracting subnetwork. Furthermore, another convolutional layer is further used to decrease the channels of feature maps. In the expanding subnetwork, inverse wavelet transform is then deployed to reconstruct the high resolution feature maps. Our MWCNN can also be explained as the generalization of dilated filtering and subsampling, and can be applied to many image restoration tasks. The experimental results clearly show the effectiveness of MWCNN for image denoising, single image super-resolution, and JPEG image artifacts removal.

Ibrahim Omara, Hongzhi Zhang, Faqiang Wang et al.

Ear recognition task is known as predicting whether two ear images belong to the same person or not. In this paper, we present a novel metric learning method for ear recognition. This method is formulated as a pairwise constrained optimization problem. In each training cycle, this method selects the nearest similar and dissimilar neighbors of each sample to construct the pairwise constraints, and then solve the optimization problem by the iterated Bregman projections. Experiments are conducted on AMI, USTB II and WPUT databases. The results show that the proposed approach can achieve promising recognition rates in ear recognition, and its training process is much more efficient than the other competing metric learning methods.

Hongzhi Zhang, Guandong Xu, Xiao Liang et al.

Relation detection plays a crucial role in Knowledge Base Question Answering (KBQA) because of the high variance of relation expression in the question. Traditional deep learning methods follow an encoding-comparing paradigm, where the question and the candidate relation are represented as vectors to compare their semantic similarity. Max- or average- pooling operation, which compresses the sequence of words into fixed-dimensional vectors, becomes the bottleneck of information. In this paper, we propose to learn attention-based word-level interactions between questions and relations to alleviate the bottleneck issue. Similar to the traditional models, the question and relation are firstly represented as sequences of vectors. Then, instead of merging the sequence into a single vector with pooling operation, soft alignments between words from the question and the relation are learned. The aligned words are subsequently compared with the convolutional neural network (CNN) and the comparison results are merged finally. Through performing the comparison on low-level representations, the attention-based word-level interaction model (ABWIM) relieves the information loss issue caused by merging the sequence into a fixed-dimensional vector before the comparison. The experimental results of relation detection on both SimpleQuestions and WebQuestions datasets show that ABWIM achieves state-of-the-art accuracy, demonstrating its effectiveness.