Yi Lu

Jialin Yang, Dongfu Jiang, Lipeng He et al. · amazon-science, utoronto

As Large Language Models (LLMs) become integral to software development workflows, their ability to generate structured outputs has become critically important. We introduce StructEval, a comprehensive benchmark for evaluating LLMs' capabilities in producing both non-renderable (JSON, YAML, CSV) and renderable (HTML, React, SVG) structured formats. Unlike prior benchmarks, StructEval systematically evaluates structural fidelity across diverse formats through two paradigms: 1) generation tasks, producing structured output from natural language prompts, and \textbf{2)} conversion tasks, translating between structured formats. Our benchmark encompasses 18 formats and 44 types of task, with novel metrics for format adherence and structural correctness. Results reveal significant performance gaps-even state-of-the-art models like o1-mini achieve only 75.58 average score, with open-source alternatives lagging approximately 10 points behind. We find generation tasks more challenging than conversion tasks, and producing correct visual content more difficult than generating text-only structures.

Hyunji Lee, Jaeyoung Kim, Hoyeon Chang et al. · deepmind

The generative retrieval model depends solely on the information encoded in its model parameters without external memory, its information capacity is limited and fixed. To overcome the limitation, we propose Nonparametric Decoding (Np Decoding) which can be applied to existing generative retrieval models. Np Decoding uses nonparametric contextualized vocab embeddings (external memory) rather than vanilla vocab embeddings as decoder vocab embeddings. By leveraging the contextualized vocab embeddings, the generative retrieval model is able to utilize both the parametric and nonparametric space. Evaluation over 9 datasets (8 single-hop and 1 multi-hop) in the document retrieval task shows that applying Np Decoding to generative retrieval models significantly improves the performance. We also show that Np Decoding is data- and parameter-efficient, and shows high performance in the zero-shot setting.

Zhuofeng Li, Yi Lu, Dongfu Jiang et al. · utoronto

The rapid rise in AI conference submissions has driven increasing exploration of large language models (LLMs) for peer review support. However, LLM-based reviewers often generate superficial, formulaic comments lacking substantive, evidence-grounded feedback. We attribute this to the underutilization of two key components of human reviewing: explicit rubrics and contextual grounding in existing work. To address this, we introduce REVIEWBENCH, a benchmark evaluating review text according to paper-specific rubrics derived from official guidelines, the paper's content, and human-written reviews. We further propose REVIEWGROUNDER, a rubric-guided, tool-integrated multi-agent framework that decomposes reviewing into drafting and grounding stages, enriching shallow drafts via targeted evidence consolidation. Experiments on REVIEWBENCH show that REVIEWGROUNDER, using a Phi-4-14B-based drafter and a GPT-OSS-120B-based grounding stage, consistently outperforms baselines with substantially stronger/larger backbones (e.g., GPT-4.1 and DeepSeek-R1-670B) in both alignment with human judgments and rubric-based review quality across 8 dimensions. The code is available \href{https://github.com/EigenTom/ReviewGrounder}{here}.

Yi Lu, Jing Nathan Yan, Songlin Yang et al.

Broad textual understanding and in-context learning require language models that utilize full document contexts. Due to the implementation challenges associated with directly training long-context models, many methods have been proposed for extending models to handle long contexts. However, owing to differences in data and model classes, it has been challenging to compare these approaches, leading to uncertainty as to how to evaluate long-context performance and whether it differs from standard evaluation. We implement a controlled protocol for extension methods with a standardized evaluation, utilizing consistent base models and extension data. Our study yields several insights into long-context behavior. First, we reaffirm the critical role of perplexity as a general-purpose performance indicator even in longer-context tasks. Second, we find that current approximate attention methods systematically underperform across long-context tasks. Finally, we confirm that exact fine-tuning based methods are generally effective within the range of their extension, whereas extrapolation remains challenging. All codebases, models, and checkpoints will be made available open-source, promoting transparency and facilitating further research in this critical area of AI development.

Xin Zhou, Yi Lu, Ruotian Ma et al.

Large language models (LLMs) have shown great potential as general-purpose AI assistants in various domains. To meet the requirements of different applications, LLMs are often customized by further fine-tuning. However, the powerful learning ability of LLMs not only enables them to acquire new tasks but also makes them susceptible to learning undesired behaviors. For example, even safety-aligned LLMs can be easily fine-tuned into harmful assistants as the fine-tuning data often contains implicit or explicit harmful content. Can we train LLMs on harmful data without learning harmful behaviors? This paper proposes a controllable training framework that makes harmful behaviors unlearnable during the fine-tuning process. Specifically, we introduce ``security vectors'', a few new parameters that can be separated from the LLM, to ensure LLM's responses are consistent with the harmful behavior. Security vectors are activated during fine-tuning, the consistent behavior makes LLM believe that such behavior has already been learned, there is no need to further optimize for harmful data. During inference, we can deactivate security vectors to restore the LLM's normal behavior. The experimental results show that the security vectors generated by 100 harmful samples are enough to prevent LLM from learning 1000 harmful samples, while preserving the ability to learn other useful information.

Shanshan Zhong, Yi Lu, Jingjie Ning et al.

Skills have become the de facto way to enable LLM agents to perform complex real-world tasks with customized instructions, workflows, and tools, but how to learn them automatically and effectively remains unclear. We introduce SkillLearnBench, the first benchmark for evaluating continual skill learning methods, comprising 20 verified, skill-dependent tasks across 15 sub-domains derived from a real-world skill taxonomy , evaluated at three levels: skill quality, execution trajectory, and task outcome. Using this benchmark, we evaluate recent continual learning techniques, those leveraging one-shot, self/teacher feedback, and skill creator to generate skills from agent experiences. We find that all continual learning methods improve over the no-skill baseline, yet consistent gains remain elusive: no method leads across all tasks and LLMs, and scaling to stronger LLMs does not reliably help. Continual learning improves tasks with clear, reusable workflows but struggles on open-ended tasks, and using stronger LLM backbones does not consistently produce better skills. Our analysis also revealed that multiple iterations in continual learning facilitate genuine improvement via external feedback, whereas self-feedback alone induces recursive drift. Our data and code are open-source at https://github.com/cxcscmu/SkillLearnBench to enable further studies of automatic skill generation and continual learning techniques.

Anglin Liu, Ruichao Chen, Yi Lu et al.

Despite recent Multimodal Large Language Models (MLLMs)' linguistic prowess in medical diagnosis, we find even state-of-the-art MLLMs suffer from a critical perceptual deficit: geometric blindness. This failure to ground outputs in objective geometric constraints leads to plausible yet factually incorrect hallucinations, rooted in training paradigms that prioritize linguistic fluency over geometric fidelity. This paper introduces Med-Scout, a novel framework that "cures" this blindness via Reinforcement Learning (RL) that leverages the intrinsic geometric logic latent within unlabeled medical images. Instead of relying on costly expert annotations, Med-Scout derives verifiable supervision signals through three strategic proxy tasks: Hierarchical Scale Localization, Topological Jigsaw Reconstruction, and Anomaly Consistency Detection. To rigorously quantify this deficit, we present Med-Scout-Bench, a new benchmark specifically designed to evaluate geometric perception. Extensive evaluations show that Med-Scout significantly mitigates geometric blindness, outperforming leading proprietary and open-source MLLMs by over 40% on our benchmark. Furthermore, this enhanced geometric perception generalizes to broader medical understanding, achieving superior results on radiological and comprehensive medical VQA tasks.

Jianing Wang, Jianfei Zhang, Qi Guo et al.

We introduce LongCat-Flash-Prover, a flagship 560-billion-parameter open-source Mixture-of- Experts (MoE) model that advances Native Formal Reasoning in Lean4 through agentic tool-integrated reasoning (TIR). We decompose the native formal reasoning task into three independent formal capabilities, i.e., auto-formalization, sketching, and proving. To facilitate these capabilities, we propose a Hybrid-Experts Iteration Framework to expand high-quality task trajectories, including generating a formal statement based on a given informal problem, producing a whole-proof directly from the statement, or a lemma-style sketch. During agentic RL, we present a Hierarchical Importance Sampling Policy Optimization (HisPO) algorithm, which aims to stabilize the MoE model training on such long-horizon tasks. It employs a gradient masking strategy that accounts for the policy staleness and the inherent train-inference engine discrepancies at both sequence and token levels. Additionally, we also incorporate theorem consistency and legality detection mechanisms to eliminate reward hacking issues. Extensive evaluations show that our LongCat-Flash-Prover sets a new state-of-the-art for open-weights models in both auto-formalization and theorem proving. Demonstrating remarkable sample efficiency, it achieves a 97.1% pass rate on MiniF2F-Test using only 72 inference budget per problem. On more challenging benchmarks, it solves 70.8% of ProverBench and 41.5% of PutnamBench with no more than 220 attempts per problem, significantly outperforming existing open-weights baselines.

Chenglong Wang, Yi Lu, Yongyu Mu et al.

Knowledge distillation addresses the problem of transferring knowledge from a teacher model to a student model. In this process, we typically have multiple types of knowledge extracted from the teacher model. The problem is to make full use of them to train the student model. Our preliminary study shows that: (1) not all of the knowledge is necessary for learning a good student model, and (2) knowledge distillation can benefit from certain knowledge at different training steps. In response to these, we propose an actor-critic approach to selecting appropriate knowledge to transfer during the process of knowledge distillation. In addition, we offer a refinement of the training algorithm to ease the computational burden. Experimental results on the GLUE datasets show that our method outperforms several strong knowledge distillation baselines significantly.

Ruibo Fu, Rui Liu, Chunyu Qiang et al.

The Inspirational and Convincing Audio Generation Challenge 2024 (ICAGC 2024) is part of the ISCSLP 2024 Competitions and Challenges track. While current text-to-speech (TTS) technology can generate high-quality audio, its ability to convey complex emotions and controlled detail content remains limited. This constraint leads to a discrepancy between the generated audio and human subjective perception in practical applications like companion robots for children and marketing bots. The core issue lies in the inconsistency between high-quality audio generation and the ultimate human subjective experience. Therefore, this challenge aims to enhance the persuasiveness and acceptability of synthesized audio, focusing on human alignment convincing and inspirational audio generation. A total of 19 teams have registered for the challenge, and the results of the competition and the competition are described in this paper.

Jiaquan Zhang, Caiyan Qin, Haoyu Bian et al.

Neural operators learn mappings from function-dependent inputs to solutions, providing an effective framework for solving partial differential equations (PDEs). For time-dependent PDEs, existing methods typically perform long-horizon prediction through autoregressive rollout directly in high-dimensional physical field spaces, where each predicted state is recursively fed back as the input for the next step. Although effective for short-term prediction, this autoregressive rollout and the lack of continuous-time modeling lead to progressive error accumulation over long-horizon rollouts. In this work, we propose Autoregression-Free Neural Operators (AFNO), which map the time evolution of PDEs into a latent space and model continuous-time vector fields within it. AFNO uses flow matching to learn the latent vector field, thereby enabling continuous evolution over extended horizons, avoiding autoregressive rollout and capturing dynamics under varying parameter configurations through explicit conditioning on physical parameters. Theoretical analysis and extensive experiments on six PDEs demonstrate that AFNO improves long-horizon prediction stability and consistently reduces rollout errors compared with the baselines.

Yuankun Xie, Chenxu Xiong, Xiaopeng Wang et al.

Currently, Audio Language Models (ALMs) are rapidly advancing due to the developments in large language models and audio neural codecs. These ALMs have significantly lowered the barrier to creating deepfake audio, generating highly realistic and diverse types of deepfake audio, which pose severe threats to society. Consequently, effective audio deepfake detection technologies to detect ALM-based audio have become increasingly critical. This paper investigate the effectiveness of current countermeasure (CM) against ALM-based audio. Specifically, we collect 12 types of the latest ALM-based deepfake audio and utilizing the latest CMs to evaluate. Our findings reveal that the latest codec-trained CM can effectively detect ALM-based audio, achieving 0% equal error rate under most ALM test conditions, which exceeded our expectations. This indicates promising directions for future research in ALM-based deepfake audio detection.

Cheng Ge, Yi Lu, Jia Qu et al.

De novo peptide sequencing from mass spectrometry data is an important method for protein identification. Recently, various deep learning approaches were applied for de novo peptide sequencing and DeepNovoV2 is one of the represetative models. In this study, we proposed an enhanced model, DePS, which can improve the accuracy of de novo peptide sequencing even with missing signal peaks or large number of noisy peaks in tandem mass spectrometry data. It is showed that, for the same test set of DeepNovoV2, the DePS model achieved excellent results of 74.22%, 74.21% and 41.68% for amino acid recall, amino acid precision and peptide recall respectively. Furthermore, the results suggested that DePS outperforms DeepNovoV2 on the cross species dataset.

Yi Lu, Deyang Kong, Jianing Wang et al.

Recent advances in block diffusion language models have demonstrated competitive performance and strong scalability on reasoning tasks. However, existing BDLMs have limited exploration under the test-time scaling setting and face more severe decoding challenges in long Chain-of-Thought reasoning, particularly in balancing the decoding speed and effectiveness. In this work, we propose a unified framework for test-time scaling in BDLMs that introduces adaptivity in both decoding and block-wise generation. At the decoding level, we propose Bounded Adaptive Confidence Decoding (BACD), a difficulty-aware sampling strategy that dynamically adjusts denoising based on model confidence, accelerating inference while controlling error accumulation. Beyond step-wise adaptivity, we introduce Think Coarse, Critic Fine (TCCF), a test-time scaling paradigm that allocates large block sizes to exploratory reasoning and smaller block sizes to refinement, achieving an effective efficiency-effectiveness balance. To enable efficient and effective decoding with a large block size, we adopt Progressive Block Size Extension, which mitigates performance degradation when scaling block sizes. Extensive experiments show that applying BACD and TCCF to TDAR-8B yields significant improvements over strong baselines such as TraDo-8B (2.26x speedup, +11.2 points on AIME24). These results mark an important step toward unlocking the potential of BDLMs for test-time scaling in complex reasoning tasks.

Zhuofeng Li, Haoxiang Zhang, Cong Wei et al.

Modern retrieval systems, whether lexical or semantic, expose a corpus through a fixed similarity interface that compresses access into a single top-k retrieval step before reasoning. This abstraction is efficient, but for agentic search, it becomes a bottleneck: exact lexical constraints, sparse clue conjunctions, local context checks, and multi-step hypothesis refinement are difficult to implement by calling a conventional off-the-shelf retriever, and evidence filtered out early cannot be recovered by stronger downstream reasoning. Agentic tasks further exacerbate this limitation because they require agents to orchestrate multiple steps, including discovering intermediate entities, combining weak clues, and revising the plan after observing partial evidence. To tackle the limitation, we study direct corpus interaction (DCI), where an agent searches the raw corpus directly with general-purpose terminal tools (e.g., grep, file reads, shell commands, lightweight scripts), without any embedding model, vector index, or retrieval API. This approach requires no offline indexing and adapts naturally to evolving local corpora. Across IR benchmarks and end-to-end agentic search tasks, this simple setup substantially outperforms strong sparse, dense, and reranking baselines on several BRIGHT and BEIR datasets, and attains strong accuracy on BrowseComp-Plus and multi-hop QA without relying on any conventional semantic retriever. Our results indicate that as language agents become stronger, retrieval quality depends not only on reasoning ability but also on the resolution of the interface through which the model interacts with the corpus, with which DCI opens a broader interface-design space for agentic search.

Xin Qi, Ruibo Fu, Zhengqi Wen et al.

In recent years, speech diffusion models have advanced rapidly. Alongside the widely used U-Net architecture, transformer-based models such as the Diffusion Transformer (DiT) have also gained attention. However, current DiT speech models treat Mel spectrograms as general images, which overlooks the specific acoustic properties of speech. To address these limitations, we propose a method called Directional Patch Interaction for Text-to-Speech (DPI-TTS), which builds on DiT and achieves fast training without compromising accuracy. Notably, DPI-TTS employs a low-to-high frequency, frame-by-frame progressive inference approach that aligns more closely with acoustic properties, enhancing the naturalness of the generated speech. Additionally, we introduce a fine-grained style temporal modeling method that further improves speaker style similarity. Experimental results demonstrate that our method increases the training speed by nearly 2 times and significantly outperforms the baseline models.

Qi Guo, Jianing Wang, Deyang Kong et al.

Parallel thinking has emerged as a new paradigm for large reasoning models (LRMs) in tackling complex problems. Recent methods leverage Reinforcement Learning (RL) to enhance parallel thinking, aiming to address the limitations in computational resources and effectiveness encountered with supervised fine-tuning. However, most existing studies primarily focus on optimizing the aggregation phase, with limited attention to the path exploration stage. In this paper, we theoretically analyze the optimization of parallel thinking under the Reinforcement Learning with Verifiable Rewards (RLVR) setting, and identify that the mutual information bottleneck among exploration paths fundamentally restricts overall performance. To address this, we propose Outline-Guided Path Exploration (OPE), which explicitly partitions the solution space by generating diverse reasoning outlines prior to parallel path reasoning, thereby reducing information redundancy and improving the diversity of information captured across exploration paths. We implement OPE with an iterative RL strategy that optimizes outline planning and outline-guided reasoning independently. Extensive experiments across multiple challenging mathematical benchmarks demonstrate that OPE effectively improves reasoning performance in different aggregation strategies, enabling LRMs to more reliably discover correct solutions.

Yuankun Xie, Yi Lu, Ruibo Fu et al.

With the proliferation of Audio Language Model (ALM) based deepfake audio, there is an urgent need for generalized detection methods. ALM-based deepfake audio currently exhibits widespread, high deception, and type versatility, posing a significant challenge to current audio deepfake detection (ADD) models trained solely on vocoded data. To effectively detect ALM-based deepfake audio, we focus on the mechanism of the ALM-based audio generation method, the conversion from neural codec to waveform. We initially constructed the Codecfake dataset, an open-source, large-scale collection comprising over 1 million audio samples in both English and Chinese, focus on ALM-based audio detection. As countermeasure, to achieve universal detection of deepfake audio and tackle domain ascent bias issue of original sharpness aware minimization (SAM), we propose the CSAM strategy to learn a domain balanced and generalized minima. In our experiments, we first demonstrate that ADD model training with the Codecfake dataset can effectively detects ALM-based audio. Furthermore, our proposed generalization countermeasure yields the lowest average equal error rate (EER) of 0.616% across all test conditions compared to baseline models. The dataset and associated code are available online.

Yuxuan Zhang, Yubo Wang, Yipeng Zhu et al.

AI agents may be able to automate your inbox, but can they automate other routine aspects of your life? Everyday online tasks offer a realistic yet unsolved testbed for evaluating the next generation of AI agents. To this end, we introduce ClawBench, an evaluation framework of 153 simple tasks that people need to accomplish regularly in their lives and work, spanning 144 live platforms across 15 categories, from completing purchases and booking appointments to submitting job applications. These tasks require demanding capabilities beyond existing benchmarks, such as obtaining relevant information from user-provided documents, navigating multi-step workflows across diverse platforms, and write-heavy operations like filling in many detailed forms correctly. Unlike existing benchmarks that evaluate agents in offline sandboxes with static pages, ClawBench operates on production websites, preserving the full complexity, dynamic nature, and challenges of real-world web interaction. A lightweight interception layer captures and blocks only the final submission request, ensuring safe evaluation without real-world side effects. Our evaluations of 7 frontier models show that both proprietary and open-source models can complete only a small portion of these tasks. For example, Claude Sonnet 4.6 achieves only 33.3%. Progress on ClawBench brings us closer to AI agents that can function as reliable general-purpose assistants.

Yi Lu, Xin Zhou, Wei He et al.

Large language models (LLMs) have achieved impressive performance in numerous domains but often struggle to process lengthy inputs effectively and efficiently due to limited length generalization and attention's quadratic computational demands. Many sought to mitigate this by restricting the attention window within the pre-trained length. However, these methods introduce new issues such as ignoring the middle context and requiring additional training. To address these problems, we propose LongHeads, a training-free framework that enhances LLM's long context ability by unlocking multi-head attention's untapped potential. Instead of allowing each head to attend to the full sentence, which struggles with generalizing to longer sequences due to out-of-distribution (OOD) issues, we allow each head to process in-distribution length by selecting and attending to important context chunks. To this end, we propose a chunk selection strategy that relies on the inherent correlation between the query and the key representations, efficiently distributing context chunks to different heads. In this way, each head ensures it can effectively process attended tokens within the trained length, while different heads in different layers can collectively process longer contexts. LongHeads works efficiently in linear time, fits seamlessly with many LLMs that use relative positional encoding. LongHeads achieves 100% accuracy at the 128k length on passkey retrieval task, verifying LongHeads's efficacy in extending the usable context window for existing models. We release our code at https://github.com/LuLuLuyi/LongHeads .

Dongfu Jiang, Yi Lu, Zhuofeng Li et al. · utoronto

Reinforcement Learning with Verifiable Rewards (RLVR) has demonstrated success in enhancing LLM reasoning capabilities, but remains limited to single-turn interactions without tool integration. While recent Agentic Reinforcement Learning with Tool use (ARLT) approaches have emerged to address multi-turn tool interactions, existing works develop task-specific codebases that suffer from fragmentation, synchronous execution bottlenecks, and limited extensibility across domains. These inefficiencies hinder broader community adoption and algorithmic innovation. We introduce VerlTool, a unified and modular framework that addresses these limitations through systematic design principles. VerlTool provides four key contributions: (1) upstream alignment with VeRL ensuring compatibility and simplified maintenance, (2) unified tool management via standardized APIs supporting diverse modalities including code execution, search, SQL databases, and vision processing, (3) asynchronous rollout execution achieving near 2$\times$ speedup by eliminating synchronization bottlenecks, and (4) comprehensive evaluation demonstrating competitive performance across 6 ARLT domains. Our framework formalizes ARLT as multi-turn trajectories with multi-modal observation tokens (text/image/video), extending beyond single-turn RLVR paradigms. We train and evaluate models on mathematical reasoning, knowledge QA, SQL generation, visual reasoning, web search, and software engineering tasks, achieving results comparable to specialized systems while providing unified training infrastructure. The modular plugin architecture enables rapid tool integration requiring only lightweight Python definitions, significantly reducing development overhead and providing a scalable foundation for tool-augmented RL research. Our code is open-sourced at https://github.com/TIGER-AI-Lab/verl-tool.

Yi Lu, Jianhao Ding, Zhaofei Yu

Spiking neural networks (SNNs) offer a biologically inspired computing paradigm with significant potential for energy-efficient neural processing. Among neural coding schemes of SNNs, Time-To-First-Spike (TTFS) coding, which encodes information through the precise timing of a neuron's first spike, provides exceptional energy efficiency and biological plausibility. Despite its theoretical advantages, existing TTFS models lack efficient training methods, suffering from high inference latency and limited performance. In this work, we present a comprehensive framework, which enables the efficient training of deep TTFS-coded SNNs by employing backpropagation throuh time (BPTT) algorithm. We name the generalized TTFS coding method in our framework as latency coding. The framework includes: (1) a latency encoding (LE) module with feature extraction and straight-through estimators to address severe information loss in direct intensity-to-latency mapping and ensure smooth gradient flow; (2) relaxation of the strict single-spike constraint of traditional TTFS, allowing neurons of intermediate layers to fire multiple times to mitigating gradient vanishing in deep networks; (3) a temporal adaptive decision (TAD) loss function that dynamically weights supervision signals based on sample-dependent confidence, resolving the incompatibility between latency coding and standard cross-entropy loss. Experimental results demonstrate that our method achieves state-of-the-art accuracy in comparison to existing TTFS-coded SNNs with ultra-low inference latency and superior energy efficiency. The framework also demonstrates improved robustness against input corruptions. Our study investigates the characteristics and potential of latency coding in scenarios demanding rapid response, providing valuable insights for further exploiting the temporal learning capabilities of SNNs.

Dengjia Zhang, Charles Weng, Katherine Guerrerio et al.

The HLTCOE Evaluation team participated in TREC VQA's Answer Generation (AG) task, for which we developed a listwise learning framework that aims to improve semantic precision and ranking consistency in answer generation. Given a video-question pair, a base multimodal model first generates multiple candidate answers, which are then reranked using a model trained with a novel Masked Pointer Cross-Entropy Loss with Rank Weights. This objective integrates pointer-based candidate selection, rank-dependent weighting, and masked cross-entropy under vocabulary restriction, enabling stable and interpretable listwise optimization. By bridging generative modeling with discriminative ranking, our method produces coherent, fine-grained answer lists. Experiments reveal consistent gains in accuracy and ranking stability, especially for questions requiring temporal reasoning and semantic disambiguation.

Songcheng Cai, Zhiheng Lyu, Yuansheng Ni et al.

Agentic repository-level code understanding is essential for automating complex software engineering tasks, yet the field lacks reliable benchmarks. Existing evaluations often overlook the long tail topics and rely on popular repositories where Large Language Models (LLMs) can cheat via memorized knowledge. To address this, we introduce SWE-QA-Pro, a benchmark constructed from diverse, long-tail repositories with executable environments. We enforce topical balance via issue-driven clustering to cover under-represented task types and apply a rigorous difficulty calibration process: questions solvable by direct-answer baselines are filtered out. This results in a dataset where agentic workflows significantly outperform direct answering (e.g., a ~13-point gap for Claude Sonnet 4.5), confirming the necessity of agentic codebase exploration. Furthermore, to tackle the scarcity of training data for such complex behaviors, we propose a scalable synthetic data pipeline that powers a two-stage training recipe: Supervised Fine-Tuning (SFT) followed by Reinforcement Learning from AI Feedback (RLAIF). This approach allows small open models to learn efficient tool usage and reasoning. Empirically, a Qwen3-8B model trained with our recipe surpasses GPT-4o by 2.3 points on SWE-QA-Pro and substantially narrows the gap to state-of-the-art proprietary models, demonstrating both the validity of our evaluation and the effectiveness of our agentic training workflow.

Qingrui Zhao, Kaiyue Yang, Xiyu Wang et al.

Humanoid robots require diverse motor skills to integrate into complex environments, but bridging the kinematic and dynamic embodiment gap from human data remains a major bottleneck. We demonstrate through Hessian analysis that traditional optimization-based retargeting is inherently non-convex and prone to local optima, leading to physical artifacts like joint jumps and self-penetration. To address this, we reformulate the targeting problem as learning data distribution rather than optimizing optimal solutions, where we propose NMR, a Neural Motion Retargeting framework that transforms static geometric mapping into a dynamics-aware learned process. We first propose Clustered-Expert Physics Refinement (CEPR), a hierarchical data pipeline that leverages VAE-based motion clustering to group heterogeneous movements into latent motifs. This strategy significantly reduces the computational overhead of massively parallel reinforcement learning experts, which project and repair noisy human demonstrations onto the robot's feasible motion manifold. The resulting high-fidelity data supervises a non-autoregressive CNN-Transformer architecture that reasons over global temporal context to suppress reconstruction noise and bypass geometric traps. Experiments on the Unitree G1 humanoid across diverse dynamic tasks (e.g., martial arts, dancing) show that NMR eliminates joint jumps and significantly reduces self-collisions compared to state-of-the-art baselines. Furthermore, NMR-generated references accelerate the convergence of downstream whole-body control policies, establishing a scalable path for bridging the human-robot embodiment gap.

Xin Jin, Qianqian Qiao, Yi Lu et al.

Datasets play a pivotal role in training visual models, facilitating the development of abstract understandings of visual features through diverse image samples and multidimensional attributes. However, in the realm of aesthetic evaluation of artistic images, datasets remain relatively scarce. Existing painting datasets are often characterized by limited scoring dimensions and insufficient annotations, thereby constraining the advancement and application of automatic aesthetic evaluation methods in the domain of painting. To bridge this gap, we introduce the Aesthetics Paintings and Drawings Dataset (APDD), the first comprehensive collection of paintings encompassing 24 distinct artistic categories and 10 aesthetic attributes. Building upon the initial release of APDDv1, our ongoing research has identified opportunities for enhancement in data scale and annotation precision. Consequently, APDDv2 boasts an expanded image corpus and improved annotation quality, featuring detailed language comments to better cater to the needs of both researchers and practitioners seeking high-quality painting datasets. Furthermore, we present an updated version of the Art Assessment Network for Specific Painting Styles, denoted as ArtCLIP. Experimental validation demonstrates the superior performance of this revised model in the realm of aesthetic evaluation, surpassing its predecessor in accuracy and efficacy. The dataset and model are available at https://github.com/BestiVictory/APDDv2.git.

Matias Quintana, Fangqi Liu, Jussi Torkko et al.

Quantifying and assessing urban greenery is consequential for planning and development, reflecting the everlasting importance of green spaces for multiple climate and well-being dimensions of cities. Evaluation can be broadly grouped into objective (e.g., measuring the amount of greenery) and subjective (e.g., polling the perception of people) approaches, which may differ -- what people see and feel about how green a place is might not match the measurements of the actual amount of vegetation. In this work, we advance the state of the art by measuring such differences and explaining them through human, geographic, and spatial dimensions. The experiments rely on contextual information extracted from street view imagery and a comprehensive urban visual perception survey collected from 1,000 people across five countries with their extensive demographic and personality information. We analyze the discrepancies between objective measures (e.g., Green View Index (GVI)) and subjective scores (e.g., pairwise ratings), examining whether they can be explained by a variety of human and visual factors such as age group and spatial variation of greenery in the scene. The findings reveal that such discrepancies are comparable around the world and that demographics and personality do not play a significant role in perception. Further, while perceived and measured greenery correlate consistently across geographies (both where people and where imagery are from), where people live plays a significant role in explaining perceptual differences, with these two, as the top among seven, features that influences perceived greenery the most. This location influence suggests that cultural, environmental, and experiential factors substantially shape how individuals observe greenery in cities.

Chenghao Li, Chaoning Zhang, Yi Lu et al.

Chain-of-Thought (CoT) prompting enhances the reasoning of large language models (LLMs) by decomposing problems into sequential steps, mimicking human logic and reducing errors. However, complex tasks with vast solution spaces and vague constraints often exceed the capacity of a single reasoning chain. Inspired by Minimal Free Resolution (MFR) in commutative algebra and algebraic geometry, we propose Syzygy of Thoughts (SoT)-a novel framework that extends CoT by introducing auxiliary, interrelated reasoning paths. SoT captures deeper logical dependencies, enabling more robust and structured problem-solving. MFR decomposes a module into a sequence of free modules with minimal rank, providing a structured analytical approach to complex systems. This method introduces the concepts of "Module", "Betti numbers","Freeness", "Mapping", "Exactness" and "Minimality", enabling the systematic decomposition of the original complex problem into logically complete minimal subproblems while preserving key problem features and reducing reasoning length. We tested SoT across diverse datasets (e.g., GSM8K, MATH) and models (e.g., GPT-4o-mini, Qwen2.5), achieving inference accuracy that matches or surpasses mainstream CoTs standards. Additionally, by aligning the sampling process with algebraic constraints, our approach enhances the scalability of inference time in LLMs, ensuring both transparent reasoning and high performance. Our code will be publicly available at https://github.com/dlMARiA/Syzygy-of-thoughts.

Chenghao Li, Chaoning Zhang, Yi Lu et al.

With the development of large language models (LLMs), particularly with the introduction of the long reasoning chain technique, the reasoning ability of LLMs in complex problem-solving has been significantly enhanced. While acknowledging the power of long reasoning chains, we cannot help but wonder: Why do different reasoning chains perform differently in reasoning? What components of the reasoning chains play a key role? Existing studies mainly focus on evaluating reasoning chains from a functional perspective, with little attention paid to their structural mechanisms. To address this gap, this work is the first to analyze and evaluate the quality of the reasoning chain from a structural perspective. We apply persistent homology from Topological Data Analysis (TDA) to map reasoning steps into semantic space, extract topological features, and analyze structural changes. These changes reveal semantic coherence, logical redundancy, and identify logical breaks and gaps. By calculating homology groups, we assess connectivity and redundancy at various scales, using barcode and persistence diagrams to quantify stability and consistency. Our results show that the topological structural complexity of reasoning chains correlates positively with accuracy. More complex chains identify correct answers sooner, while successful reasoning exhibits simpler topologies, reducing redundancy and cycles, enhancing efficiency and interpretability. This work provides a new perspective on reasoning chain quality assessment and offers guidance for future optimization.

Cherag Aroraa, Tracy Holloway King, Jayant Kumar et al.

As user content and queries become increasingly multi-modal, the need for effective multi-modal search systems has grown. Traditional search systems often rely on textual and metadata annotations for indexed images, while multi-modal embeddings like CLIP enable direct search using text and image embeddings. However, embedding-based approaches face challenges in integrating contextual features such as user locale and recency. Building a scalable multi-modal search system requires fine-tuning several components. This paper presents a multi-modal search architecture and a series of AB tests that optimize embeddings and multi-modal technologies in Adobe Express template search. We address considerations such as embedding model selection, the roles of embeddings in matching and ranking, and the balance between dense and sparse embeddings. Our iterative approach demonstrates how utilizing sparse, dense, and contextual features enhances short and long query search, significantly reduces null rates (over 70\%), and increases click-through rates (CTR). Our findings provide insights into developing robust multi-modal search systems, thereby enhancing relevance for complex queries.

Ziqiao Kong, Wanxu Xia, Chong Wang et al.

Smart contracts govern billions of dollars in decentralized finance (DeFi), yet automated vulnerability detection remains challenging because many vulnerabilities are tightly coupled with project-specific business logic. We observe that recurring vulnerabilities across diverse DeFi business models often share the same underlying economic mechanisms, which we term DeFi semantics, and that capturing these shared abstractions can enable more systematic auditing. Building on this insight, we propose Knowdit, a knowledge-driven, agentic framework for smart contract vulnerability detection. Knowdit first constructs an auditing knowledge graph from historical human audit reports, linking fine-grained DeFi semantics with recurring vulnerability patterns. Given a new project, a multi-agent framework leverages this knowledge through an iterative loop of specification generation, harness synthesis, fuzz execution, and finding reflection, driven by a shared working memory for continuous refinement. We evaluate Knowdit on 12 recent Code4rena projects with 75 ground-truth vulnerabilities. Knowdit detects all 14 high-severity and 77\% of medium-severity vulnerabilities with only 2 false positives, significantly outperforming all baselines. Applied to six real-world projects, Knowdit further discovers 12 high- and 10 medium-severity previously unknown vulnerabilities, proving its outstanding performance.

Bozheng Li, Miao Yang, Zhenhan Chen et al. · utoronto

Dynamic Chart Generation (DCG) involves producing code-rendered animated visualizations as charts. While recent advances in multi-modal large language models (MLLMs) have significantly improved their capability on static chart generation and comprehension, MLLMs' potential for handling dynamic chart generation and understanding remains underexplored. To bridge this research gap, we introduce DCG-Bench (Dynamic Chart Generation Benchmark), the first benchmark evaluating MLLM's capability on dynamic chart generation tasks from three dimensions: Simple Text-to-Chart, Detailed Text-to-Chart, and Video-to-Chart tasks. We construct DCG-8K, a high-quality DCG dataset with annotations covering instruction-code-video triplets and QA pairs for both code and video evaluation. Based on DCG-8K, we explored a two-stage training recipe, proposing Joint-Code-Visual Reward for group relative policy optimization to construct expert MLLM Qwen2.5-VL-DCG-3B for the DCG task. Our benchmarking result reveals shortcomings of existing MLLMs in the visual-to-chart task, and our model beats the best open-sourced MLLM with an average 8.31% performance gain across three tasks, and shows on par performance against proprietary models with only 3B parameters, proving the effectiveness of our training recipe. Our code and dataset will be publicly available.

Bozheng Li, Yongliang Wu, Yi Lu et al. · utoronto

Widely shared videos on the internet are often edited. Recently, although Video Large Language Models (Vid-LLMs) have made great progress in general video understanding tasks, their capabilities in video editing understanding (VEU) tasks remain unexplored. To address this gap, in this paper, we introduce VEU-Bench (Video Editing Understanding Benchmark), a comprehensive benchmark that categorizes video editing components across various dimensions, from intra-frame features like shot size to inter-shot attributes such as cut types and transitions. Unlike previous video editing understanding benchmarks that focus mainly on editing element classification, VEU-Bench encompasses 19 fine-grained tasks across three stages: recognition, reasoning, and judging. To enhance the annotation of VEU automatically, we built an annotation pipeline integrated with an ontology-based knowledge base. Through extensive experiments with 11 state-of-the-art Vid-LLMs, our findings reveal that current Vid-LLMs face significant challenges in VEU tasks, with some performing worse than random choice. To alleviate this issue, we develop Oscars, a VEU expert model fine-tuned on the curated VEU-Bench dataset. It outperforms existing open-source Vid-LLMs on VEU-Bench by over 28.3% in accuracy and achieves performance comparable to commercial models like GPT-4o. We also demonstrate that incorporating VEU data significantly enhances the performance of Vid-LLMs on general video understanding benchmarks, with an average improvement of 8.3% across nine reasoning tasks.

Anglin Liu, Rundong Xue, Xu R. Cao et al.

Medical image segmentation is fundamental for biomedical discovery. Existing methods lack generalizability and demand extensive, time-consuming manual annotation for new clinical application. Here, we propose MedSAM-3, a text promptable medical segmentation model for medical image and video segmentation. By fine-tuning the Segment Anything Model (SAM) 3 architecture on medical images paired with semantic conceptual labels, our MedSAM-3 enables medical Promptable Concept Segmentation (PCS), allowing precise targeting of anatomical structures via open-vocabulary text descriptions rather than solely geometric prompts. We further introduce the MedSAM-3 Agent, a framework that integrates Multimodal Large Language Models (MLLMs) to perform complex reasoning and iterative refinement in an agent-in-the-loop workflow. Comprehensive experiments across diverse medical imaging modalities, including X-ray, MRI, Ultrasound, CT, and video, demonstrate that our approach significantly outperforms existing specialist and foundation models. We will release our code and model at https://github.com/Joey-S-Liu/MedSAM3.

Yanusha Mehendran, Maolin Tang, Yi Lu

Software vulnerabilities continue to undermine the reliability and security of modern systems, particularly as software complexity outpaces the capabilities of traditional detection methods. This study introduces a genetic algorithm-based method for test input generation that innovatively integrates genetic operators and adaptive learning to enhance software vulnerability detection. A key contribution is the application of the crossover operator, which facilitates exploration by searching across a broader space of potential test inputs. Complementing this, an adaptive feedback mechanism continuously learns from the system's execution behavior and dynamically guides input generation toward promising areas of the input space. Rather than relying on fixed or randomly selected inputs, the approach evolves a population of structurally valid test cases using feedback-driven selection, enabling deeper and more effective code traversal. This strategic integration of exploration and exploitation ensures that both diverse and targeted test inputs are developed over time. Evaluation was conducted across nine open-source JSON-processing libraries. The proposed method achieved substantial improvements in coverage compared to a benchmark evolutionary fuzzing method, with average gains of 39.8% in class coverage, 62.4% in method coverage, 105.0% in line coverage, 114.0% in instruction coverage, and 166.0% in branch coverage. These results highlight the method's capacity to detect deeper and more complex vulnerabilities, offering a scalable and adaptive solution to software security testing.

Ruibo Fu, Shuchen Shi, Hongming Guo et al.

Foley audio, critical for enhancing the immersive experience in multimedia content, faces significant challenges in the AI-generated content (AIGC) landscape. Despite advancements in AIGC technologies for text and image generation, the foley audio dubbing remains rudimentary due to difficulties in cross-modal scene matching and content correlation. Current text-to-audio technology, which relies on detailed and acoustically relevant textual descriptions, falls short in practical video dubbing applications. Existing datasets like AudioSet, AudioCaps, Clotho, Sound-of-Story, and WavCaps do not fully meet the requirements for real-world foley audio dubbing task. To address this, we introduce the Multi-modal Image and Narrative Text Dubbing Dataset (MINT), designed to enhance mainstream dubbing tasks such as literary story audiobooks dubbing, image/silent video dubbing. Besides, to address the limitations of existing TTA technology in understanding and planning complex prompts, a Foley Audio Content Planning, Generation, and Alignment (CPGA) framework is proposed, which includes a content planning module leveraging large language models for complex multi-modal prompts comprehension. Additionally, the training process is optimized using Proximal Policy Optimization based reinforcement learning, significantly improving the alignment and auditory realism of generated foley audio. Experimental results demonstrate that our approach significantly advances the field of foley audio dubbing, providing robust solutions for the challenges of multi-modal dubbing. Even when utilizing the relatively lightweight GPT-2 model, our framework outperforms open-source multimodal large models such as LLaVA, DeepSeek-VL, and Moondream2. The dataset is available at https://github.com/borisfrb/MINT .

Jie Lian, Jingyu Liu, Yizhou Yu et al.

The detection of thoracic abnormalities challenge is organized by the Deepwise AI Lab. The challenge is divided into two rounds. In this paper, we present the results of 6 teams which reach the second round. The challenge adopts the ChestX-Det10 dateset proposed by the Deepwise AI Lab. ChestX-Det10 is the first chest X-Ray dataset with instance-level annotations, including 10 categories of disease/abnormality of 3,543 images. The annotations are located at https://github.com/Deepwise-AILab/ChestX-Det10-Dataset. In the challenge, we randomly split all data into 3001 images for training and 542 images for testing.

Xiyin Zeng, Yi Lu, Hao Wang

Visual Question Answering (VQA) requires models to identify the correct answer options based on both visual and textual evidence. Recent Mixture-of-Experts (MoE) methods improve option reasoning by grouping similar concepts or routing based on examples. However, unstable routing can lead to inconsistent expert selection in the same question type, while overly stable routing may reduce flexibility. To address this, we propose Concept-Guided Routing framework (CoGR-MoE), which incorporates semantics of the answer options to guide expert selection in the training phase. Next, option features are used to reweight the selected experts, producing discriminative representations for each candidate option. These option-level representations are further used for option comparison and optimized via contrastive learning. The experimental results indicate that CoGR-MoE delivers strong performance across multiple VQA tasks, demonstrating the effectiveness of our approach.

Jun Zhao, Can Zu, Hao Xu et al.

Large language models (LLMs) have demonstrated impressive performance in understanding language and executing complex reasoning tasks. However, LLMs with long context windows have been notorious for their expensive training costs and high inference latency. Even the most advanced models such as GPT-4 and Claude2 often make mistakes when processing inputs of over $100k$ tokens, a phenomenon also known as \textit{lost in the middle}. In this paper, we propose \textsc{LongAgent}, a method based on multi-agent collaboration, which scales LLMs (e.g., LLaMA) to a context of 128K and demonstrates potential superiority in long-text processing compared to GPT-4. In \textsc{LongAgent}, a leader is responsible for understanding user intent and directing team members to acquire information from documents. Due to members' hallucinations, it is non-trivial for a leader to obtain accurate information from the responses of dozens to hundreds of members. To address this, we develop an \textit{inter-member communication} mechanism to resolve response conflicts caused by hallucinations through information sharing. Our experimental results indicate that \textsc{LongAgent} offers a promising alternative for long-text processing. The agent team instantiated with LLaMA-7B achieves significant improvements in tasks such as 128k-long text retrieval, multi-hop question answering, compared to GPT-4.

Yi Lu, Jiawang Cao, Yongliang Wu et al. · utoronto

Multi-modal Large Language Models (MLLMs) have demonstrated remarkable reasoning capability while lack explicit mechanisms for visual grounding and segmentation, creating a gap between cognitive reasoning and visual perception. To bridge this gap, we introduce Reasoning Segmentation via Visual Prompting (RSVP), a novel framework that unifies multi-step multimodal reasoning with grounded visual understanding. RSVP is a two-stage structuralized framework that integrates reasoning-driven localization with segmentation refinement. In the reasoning stage, RSVP employs multimodal chain-of-thought visual prompts to help MLLMs understand queries and infer targets, generating interpretable region proposals that enhance visual grounding. In segmentation stage, RSVP refines these proposals with a Vision-Language Segmentation Module (VLSM), seamlessly integrates textual and visual cues to produce precise segmentation masks. By explicitly modelling the interaction between multimodal reasoning and segmentation, RSVP introduces a new paradigm for interpretable reasoning segmentation. It exploits MLLMs' inherent localization capabilities, enabling the models to not only reason about objects but also generate structured visual representations. Our extensive experiments demonstrate that RSVP achieves state-of-the-art performance, surpasses state-of-the-art methods by up to +6.5 gIoU and +9.2 cIoU on ReasonSeg, and achieves 49.7 mAP on SegInW under zero-shot settings. These results validate RSVP as an effective and scalable framework for integrating cognitive reasoning with structured visual understanding.

Xin Jin, Qianqian Qiao, Yi Lu et al.

Image aesthetic evaluation is a highly prominent research domain in the field of computer vision. In recent years, there has been a proliferation of datasets and corresponding evaluation methodologies for assessing the aesthetic quality of photographic works, leading to the establishment of a relatively mature research environment. However, in contrast to the extensive research in photographic aesthetics, the field of aesthetic evaluation for paintings and Drawings has seen limited attention until the introduction of the BAID dataset in March 2023. This dataset solely comprises overall scores for high-quality artistic images. Our research marks the pioneering introduction of a multi-attribute, multi-category dataset specifically tailored to the field of painting: Aesthetics of Paintings and Drawings Dataset (APDD). The construction of APDD received active participation from 28 professional artists worldwide, along with dozens of students specializing in the field of art. This dataset encompasses 24 distinct artistic categories and 10 different aesthetic attributes. Each image in APDD has been evaluated by six professionally trained experts in the field of art, including assessments for both total aesthetic scores and aesthetic attribute scores. The final APDD dataset comprises a total of 4985 images, with an annotation count exceeding 31100 entries. Concurrently, we propose an innovative approach: Art Assessment Network for Specific Painting Styles (AANSPS), designed for the assessment of aesthetic attributes in mixed-attribute art datasets. Through this research, our goal is to catalyze advancements in the field of aesthetic evaluation for paintings and drawings, while enriching the available resources and methodologies for its further development and application.

Yubo Wang, Xueguang Ma, Ping Nie et al. · utoronto

Academic writing requires both coherent text generation and precise citation of relevant literature. Although recent Retrieval-Augmented Generation (RAG) systems have significantly improved factual accuracy in general-purpose text generation, their ability to support professional academic writing remains limited. In this work, we introduce ScholarCopilot, a unified framework designed to enhance existing large language models for generating professional academic articles with accurate and contextually relevant citations. ScholarCopilot dynamically determines when to retrieve scholarly references by generating a retrieval token [RET], which is then used to query a citation database. The retrieved references are fed into the model to augment the generation process. We jointly optimize both the generation and citation tasks within a single framework to improve efficiency. Our model is built upon Qwen-2.5-7B and trained on 500K papers from arXiv. It achieves a top-1 retrieval accuracy of 40.1% on our evaluation dataset, outperforming baselines such as E5-Mistral-7B-Instruct (15.0%) and BM25 (9.8%). On a dataset of 1,000 academic writing samples, ScholarCopilot scores 16.2/25 in generation quality -- measured across relevance, coherence, academic rigor, completeness, and innovation -- significantly surpassing all existing models, including much larger ones like the Retrieval-Augmented Qwen2.5-72B-Instruct. Human studies further demonstrate that ScholarCopilot, despite being a 7B model, significantly outperforms ChatGPT, achieving 100% preference in citation quality and over 70% in overall usefulness.

Shuo Li, Jiajun Sun, Guodong Zheng et al.

Recently, multimodal large language models (MLLMs) have demonstrated remarkable performance in visual-language tasks. However, the authenticity of the responses generated by MLLMs is often compromised by object hallucinations. We identify that a key cause of these hallucinations is the model's over-susceptibility to specific image frequency features in detecting objects. In this paper, we introduce Multi-Frequency Perturbations (MFP), a simple, cost-effective, and pluggable method that leverages both low-frequency and high-frequency features of images to perturb visual feature representations and explicitly suppress redundant frequency-domain features during inference, thereby mitigating hallucinations. Experimental results demonstrate that our method significantly mitigates object hallucinations across various model architectures. Furthermore, as a training-time method, MFP can be combined with inference-time methods to achieve state-of-the-art performance on the CHAIR benchmark.

Yongliang Wu, Wenbo Zhu, Jiawang Cao et al. · utoronto

The demand for producing short-form videos for sharing on social media platforms has experienced significant growth in recent times. Despite notable advancements in the fields of video summarization and highlight detection, which can create partially usable short films from raw videos, these approaches are often domain-specific and require an in-depth understanding of real-world video content. To tackle this predicament, we propose Repurpose-10K, an extensive dataset comprising over 10,000 videos with more than 120,000 annotated clips aimed at resolving the video long-to-short task. Recognizing the inherent constraints posed by untrained human annotators, which can result in inaccurate annotations for repurposed videos, we propose a two-stage solution to obtain annotations from real-world user-generated content. Furthermore, we offer a baseline model to address this challenging task by integrating audio, visual, and caption aspects through a cross-modal fusion and alignment framework. We aspire for our work to ignite groundbreaking research in the lesser-explored realms of video repurposing.

Wei He, Shichun Liu, Jun Zhao et al.

Large language models (LLMs) have shown promising abilities of in-context learning (ICL), adapting swiftly to new tasks with only few-shot demonstrations. However, current few-shot methods heavily depend on high-quality, query-specific demos, which are often lacking. When faced with out-of-demonstration (OOD) queries, methods that rely on hand-crafted demos or external retrievers might fail. To bridge the gap between limited demos and OOD queries, we propose Self-Demos, a novel prompting method that elicits the inherent generalizability in LLMs by query-aware demo generation. The generated demos strategically interpolate between existing demos and the given query, transforming the query from OOD to ID. To evaluate the effectiveness of our approach, we manually constructed OOD-Toolset, a dataset in the tool-using scenario with over 300 real-world APIs and 1000 instances, each consisting of three tool-use cases as demos and an OOD query. Thorough experiments on our dataset and two public math benchmarks have shown that our method can outperform state-of-the-art baselines in the OOD setting. Moreover, we conduct a range of analyses to validate Self-Demos's generalization and provide more insights.

Shenghao Ren, Yi Lu, Jiayi Huang et al.

Existing human Motion Capture (MoCap) methods mostly focus on the visual similarity while neglecting the physical plausibility. As a result, downstream tasks such as driving virtual human in 3D scene or humanoid robots in real world suffer from issues such as timing drift and jitter, spatial problems like sliding and penetration, and poor global trajectory accuracy. In this paper, we revisit human MoCap from the perspective of interaction between human body and physical world by exploring the role of pressure. Firstly, we construct a large-scale human Motion capture dataset with Pressure, RGB and Optical sensors (named MotionPRO), which comprises 70 volunteers performing 400 types of motion, encompassing a total of 12.4M pose frames. Secondly, we examine both the necessity and effectiveness of the pressure signal through two challenging tasks: (1) pose and trajectory estimation based solely on pressure: We propose a network that incorporates a small kernel decoder and a long-short-term attention module, and proof that pressure could provide accurate global trajectory and plausible lower body pose. (2) pose and trajectory estimation by fusing pressure and RGB: We impose constraints on orthographic similarity along the camera axis and whole-body contact along the vertical axis to enhance the cross-attention strategy to fuse pressure and RGB feature maps. Experiments demonstrate that fusing pressure with RGB features not only significantly improves performance in terms of objective metrics, but also plausibly drives virtual humans (SMPL) in 3D scene. Furthermore, we demonstrate that incorporating physical perception enables humanoid robots to perform more precise and stable actions, which is highly beneficial for the development of embodied artificial intelligence. Project page is available at: https://nju-cite-mocaphumanoid.github.io/MotionPRO/

Qingyu Fan, Zhaoxiang Li, Yi Lu et al.

Bimanual manipulation in cluttered scenes requires policies that remain stable under occlusions, viewpoint and scene variations. Existing vision-language-action models often fail to generalize because (i) multi-view features are fused via view-agnostic token concatenation, yielding weak 3D-consistent spatial understanding, and (ii) language is injected as global conditioning, resulting in coarse instruction grounding. In this paper, we introduce PEAfowl, a perception-enhanced multi-view VLA policy for bimanual manipulation. For spatial reasoning, PEAfowl predicts per-token depth distributions, performs differentiable 3D lifting, and aggregates local cross-view neighbors to form geometrically grounded, cross-view consistent representations. For instruction grounding, we propose to replace global conditioning with a Perceiver-style text-aware readout over frozen CLIP visual features, enabling iterative evidence accumulation. To overcome noisy and incomplete commodity depth without adding inference overhead, we apply training-only depth distillation from a pretrained depth teacher to supervise the depth-distribution head, providing perception front-end with geometry-aware priors. On RoboTwin 2.0 under domain-randomized setting, PEAfowl improves the strongest baseline by 23.0 pp in success rate, and real-robot experiments further demonstrate reliable sim-to-real transfer and consistent improvements from depth distillation. Project website: https://peafowlvla.github.io/.

Chenghao Li, Chaoning Zhang, Boheng Zeng et al.

Attention mechanisms have significantly advanced visual models by capturing global context effectively. However, their reliance on large-scale datasets and substantial computational resources poses challenges in data-scarce and resource-constrained scenarios. Moreover, traditional self-attention mechanisms lack inherent spatial inductive biases, making them suboptimal for modeling local features critical to tasks involving smaller datasets. In this work, we introduce Large Kernel Convolutional Attention (LKCA), a novel formulation that reinterprets attention operations as a single large-kernel convolution. This design unifies the strengths of convolutional architectures locality and translation invariance with the global context modeling capabilities of self-attention. By embedding these properties into a computationally efficient framework, LKCA addresses key limitations of traditional attention mechanisms. The proposed LKCA achieves competitive performance across various visual tasks, particularly in data-constrained settings. Experimental results on CIFAR-10, CIFAR-100, SVHN, and Tiny-ImageNet demonstrate its ability to excel in image classification, outperforming conventional attention mechanisms and vision transformers in compact model settings. These findings highlight the effectiveness of LKCA in bridging local and global feature modeling, offering a practical and robust solution for real-world applications with limited data and resources.

Haidong Kang, Ketong Qian, Yi Lu

Efforts to overcome catastrophic forgetting in Few-Shot Class-Incremental Learning (FSCIL) have primarily focused on developing more effective gradient-based optimization strategies. In contrast, little attention has been paid to the training cost explosion that inevitably arises as the number of novel classes increases, a consequence of relying on gradient learning even under extreme data scarcity. More critically, since FSCIL typically provides only a few samples for each new class, gradient-based updates not only induce severe catastrophic forgetting on base classes but also hinder adaptation to novel ones. This paper seeks to break this long-standing limitation by asking: Can we design a training-free FSCIL paradigm that entirely removes gradient optimization? We provide an affirmative answer by uncovering an intriguing connection between gradient-based optimization and the Conditional Diffusion process. Building on this observation, we propose a Conditional Diffusion-driven FSCIL (CD-FSCIL) framework that substitutes the conventional gradient update process with a diffusion-based generative transition, enabling training-free incremental adaptation while effectively mitigating forgetting. Furthermore, to enhance representation under few-shot constraints, we introduce a multimodal learning strategy that integrates visual features with natural language descriptions automatically generated by Large Language Models (LLMs). This synergy substantially alleviates the sample scarcity issue and improves generalization across novel classes. Extensive experiments on mainstream FSCIL benchmarks demonstrate that our method not only achieves state-of-the-art performance but also drastically reduces computational and memory overhead, marking a paradigm shift toward training-free continual adaptation.

Yi Lu, Jianing Wang, Linsen Guo et al.

Recent trends in test-time scaling for reasoning models (e.g., OpenAI o1, DeepSeek-R1) have led to remarkable improvements through long Chain-of-Thought (CoT). However, existing benchmarks mainly focus on immediate, single-horizon tasks, failing to adequately evaluate models' ability to understand and respond to complex, long-horizon scenarios. To address this incomplete evaluation of Large Reasoning Models (LRMs), we propose R-HORIZON, a method designed to stimulate long-horizon reasoning behaviors in LRMs through query composition. Based on R-HORIZON, we construct a long-horizon reasoning benchmark, comprising complex multi-step reasoning tasks with interdependent problems that span long reasoning horizons. Through comprehensive evaluation of LRMs using the R-HORIZON benchmark, we find that even the most advanced LRMs suffer significant performance degradation. Our analysis reveals that LRMs exhibit limited effective reasoning length and struggle to allocate thinking budget across multiple problems appropriately. Recognizing these limitations, we use R-HORIZON to construct long-horizon reasoning data for reinforcement learning with verified rewards (RLVR). Compared to training with single-horizon data, RLVR with R-HORIZON not only substantially improves performance on the multi-horizon reasoning tasks, but also promotes accuracy on standard reasoning tasks, with an increase of 7.5 on AIME2024. These results position R-HORIZON as a scalable, controllable, and low-cost paradigm for enhancing and evaluating the long-horizon reasoning capabilities of LRMs.