Showing 1–50 of 121 results for author: Ban, Y

Policy Improvement Reinforcement Learning

Authors: Huaiyang Wang, Xiaojie Li, Deqing Wang, Haoyi Zhou, Zixuan Huang, Yaodong Yang, Jianxin Li, Yikun Ban

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has become a central post-training paradigm for improving the reasoning capabilities of large language models. Yet existing methods share a common blind spot: they optimize policies based on instantaneous group-level or batch-level statistics without ever verifying whether the resulting update actually improved the model. This open-loop design… ▽ More Reinforcement Learning with Verifiable Rewards (RLVR) has become a central post-training paradigm for improving the reasoning capabilities of large language models. Yet existing methods share a common blind spot: they optimize policies based on instantaneous group-level or batch-level statistics without ever verifying whether the resulting update actually improved the model. This open-loop design -- updating in isolation at each step, guided only by within-group (batch) reward signals -- means optimization can drift or collapse with no mechanism to detect and correct these failures. We argue that the missing ingredient is policy improvement feedback: the ability to measure and optimize inter-iteration progress directly. To this end, we introduce Policy Improvement Reinforcement Learning (PIRL), a framework that replaces surrogate reward maximization with the explicit objective of maximizing cumulative policy improvement across iterations, and prove this temporal objective is perfectly aligned with maximizing final task performance. Building on PIRL, we propose Policy Improvement Policy Optimization (PIPO), which implements closed-loop optimization through retrospective verification. At each iteration, PIPO evaluates whether the previous update yielded genuine improvement against a sliding-window historical baseline, then actively reinforces beneficial updates and suppresses the harmful ones -- transforming an open-loop process into a self-correcting one. We provide theoretical analysis showing that PIPO performs ascent on the PIRL objective in expectation, and experiments on mathematical reasoning benchmarks demonstrate improved stability and performance over GRPO and its variants. △ Less

Submitted 1 April, 2026; originally announced April 2026.

SutureAgent: Learning Surgical Trajectories via Goal-conditioned Offline RL in Pixel Space

Authors: Huanrong Liu, Chunlin Tian, Tongyu Jia, Tailai Zhou, Qin Liu, Yu Gao, Yutong Ban, Yun Gu, Guy Rosman, Xin Ma, Qingbiao Li

Abstract: Predicting surgical needle trajectories from endoscopic video is critical for robot-assisted suturing, enabling anticipatory planning, real-time guidance, and safer motion execution. Existing methods that directly learn motion distributions from visual observations tend to overlook the sequential dependency among adjacent motion steps. Moreover, sparse waypoint annotations often fail to provide su… ▽ More Predicting surgical needle trajectories from endoscopic video is critical for robot-assisted suturing, enabling anticipatory planning, real-time guidance, and safer motion execution. Existing methods that directly learn motion distributions from visual observations tend to overlook the sequential dependency among adjacent motion steps. Moreover, sparse waypoint annotations often fail to provide sufficient supervision, further increasing the difficulty of supervised or imitation learning methods. To address these challenges, we formulate image-based needle trajectory prediction as a sequential decision-making problem, in which the needle tip is treated as an agent that moves step by step in pixel space. This formulation naturally captures the continuity of needle motion and enables the explicit modeling of physically plausible pixel-wise state transitions over time. From this perspective, we propose SutureAgent, a goal-conditioned offline reinforcement learning framework that leverages sparse annotations to dense reward signals via cubic spline interpolation, encouraging the policy to exploit limited expert guidance while exploring plausible future motion paths. SutureAgent encodes variable-length clips using an observation encoder to capture both local spatial cues and long-range temporal dynamics, and autoregressively predicts future waypoints through actions composed of discrete directions and continuous magnitudes. To enable stable offline policy optimization from expert demonstrations, we adopt Conservative Q-Learning with Behavioral Cloning regularization. Experiments on a new kidney wound suturing dataset containing 1,158 trajectories from 50 patients show that SutureAgent reduces Average Displacement Error by 58.6% compared with the strongest baseline, demonstrating the effectiveness of modeling needle trajectory prediction as pixel-level sequential action learning. △ Less

Submitted 18 March, 2026; originally announced March 2026.

GeoHeight-Bench: Towards Height-Aware Multimodal Reasoning in Remote Sensing

Authors: Xuran Hu, Zhitong Xiong, Zhongcheng Hong, Yifang Ban, Xiaoxiang Zhu, Wufan Zhao

Abstract: Current Large Multimodal Models (LMMs) in Earth Observation typically neglect the critical "vertical" dimension, limiting their reasoning capabilities in complex remote sensing geometries and disaster scenarios where physical spatial structures often outweigh planar visual textures. To bridge this gap, we introduce a comprehensive evaluation framework dedicated to height-aware remote sensing under… ▽ More Current Large Multimodal Models (LMMs) in Earth Observation typically neglect the critical "vertical" dimension, limiting their reasoning capabilities in complex remote sensing geometries and disaster scenarios where physical spatial structures often outweigh planar visual textures. To bridge this gap, we introduce a comprehensive evaluation framework dedicated to height-aware remote sensing understanding. First, to overcome the severe scarcity of annotated data, we develop a scalable, VLM-driven data generation pipeline utilizing systematic prompt engineering and metadata extraction. This pipeline constructs two complementary benchmarks: GeoHeight-Bench for relative height analysis, and a more challenging GeoHeight-Bench+ for holistic, terrain-aware reasoning. Furthermore, to validate the necessity of height perception, we propose GeoHeightChat, the first height-aware remote sensing LMM baseline. Serving as a strong proof of concept, our baseline demonstrates that synergizing visual semantics with implicitly injected height geometric features effectively mitigates the "vertical blind spot", successfully unlocking a new paradigm of interactive height reasoning in existing optical models. △ Less

Submitted 26 March, 2026; originally announced March 2026.

Comments: 18 pages, 4 figures

ACM Class: I.2.10

Adaptive Robust Estimator for Multi-Agent Reinforcement Learning

Authors: Zhongyi Li, Wan Tian, Jingyu Chen, Kangyao Huang, Huiming Zhang, Hui Yang, Tao Ren, Jinyang Jiang, Yijie Peng, Yikun Ban, Fuzhen Zhuang

Abstract: Multi-agent collaboration has emerged as a powerful paradigm for enhancing the reasoning capabilities of large language models, yet it suffers from interaction-level ambiguity that blurs generation, critique, and revision, making credit assignment across agents difficult. Moreover, policy optimization in this setting is vulnerable to heavy-tailed and noisy rewards, which can bias advantage estimat… ▽ More Multi-agent collaboration has emerged as a powerful paradigm for enhancing the reasoning capabilities of large language models, yet it suffers from interaction-level ambiguity that blurs generation, critique, and revision, making credit assignment across agents difficult. Moreover, policy optimization in this setting is vulnerable to heavy-tailed and noisy rewards, which can bias advantage estimation and trigger unstable or even divergent training. To address both issues, we propose a robust multi-agent reinforcement learning framework for collaborative reasoning, consisting of two components: Dual-Agent Answer-Critique-Rewrite (DACR) and an Adaptive Robust Estimator (ARE). DACR decomposes reasoning into a structured three-stage pipeline: answer, critique, and rewrite, while enabling explicit attribution of each agent's marginal contribution to its partner's performance. ARE provides robust estimation of batch experience means during multi-agent policy optimization. Across mathematical reasoning and embodied intelligence benchmarks, even under noisy rewards, our method consistently outperforms the baseline in both homogeneous and heterogeneous settings. These results indicate stronger robustness to reward noise and more stable training dynamics, effectively preventing optimization failures caused by noisy reward signals. △ Less

Submitted 23 March, 2026; originally announced March 2026.

Counterfactual Credit Policy Optimization for Multi-Agent Collaboration

Authors: Zhongyi Li, Wan Tian, Yikun Ban, Jinju Chen, Huiming Zhang, Yang Liu, Fuzhen Zhuang

Abstract: Collaborative multi-agent large language models (LLMs) can solve complex reasoning tasks by decomposing roles and aggregating diverse hypotheses. Yet, reinforcement learning (RL) for such systems is often undermined by credit assignment: a shared global reward obscures individual contributions, inflating update variance and encouraging free-riding. We introduce Counterfactual Credit Policy Optimiz… ▽ More Collaborative multi-agent large language models (LLMs) can solve complex reasoning tasks by decomposing roles and aggregating diverse hypotheses. Yet, reinforcement learning (RL) for such systems is often undermined by credit assignment: a shared global reward obscures individual contributions, inflating update variance and encouraging free-riding. We introduce Counterfactual Credit Policy Optimization (CCPO), a framework that assigns agent-specific learning signals by estimating each agent's marginal contribution through counterfactual trajectories. CCPO builds dynamic counterfactual baselines that simulate outcomes with an agent's contribution removed, yielding role-sensitive advantages for policy optimization. To further improve stability under heterogeneous tasks and data distributions, we propose a global-history-aware normalization scheme that calibrates advantages using global rollout statistics. We evaluate CCPO on two collaboration topologies: a sequential Think--Reason dyad and multi-agent voting. Across mathematical and logical reasoning benchmarks, CCPO mitigates free-riding and outperforms strong multi-agent RL baselines, yielding finer-grained and more effective credit assignment for collaborative LLM training. Our code is available at https://github.com/bhai114/ccpo. △ Less

Submitted 23 March, 2026; originally announced March 2026.

SCISSR: Scribble-Conditioned Interactive Surgical Segmentation and Refinement

Authors: Haonan Ping, Jian Jiang, Cheng Yuan, Qizhen Sun, Lv Wu, Yutong Ban

Abstract: Accurate segmentation of tissues and instruments in surgical scenes is annotation-intensive due to irregular shapes, thin structures, specularities, and frequent occlusions. While SAM models support point, box, and mask prompts, points are often too sparse and boxes too coarse to localize such challenging targets. We present SCISSR, a scribble-promptable framework for interactive surgical scene se… ▽ More Accurate segmentation of tissues and instruments in surgical scenes is annotation-intensive due to irregular shapes, thin structures, specularities, and frequent occlusions. While SAM models support point, box, and mask prompts, points are often too sparse and boxes too coarse to localize such challenging targets. We present SCISSR, a scribble-promptable framework for interactive surgical scene segmentation. It introduces a lightweight Scribble Encoder that converts freehand scribbles into dense prompt embeddings compatible with the mask decoder, enabling iterative refinement for a target object by drawing corrective strokes on error regions. Because all added modules (the Scribble Encoder, Spatial Gated Fusion, and LoRA adapters) interact with the backbone only through its standard embedding interfaces, the framework is not tied to a single model: we build on SAM 2 in this work, yet the same components transfer to other prompt-driven segmentation architectures such as SAM 3 without structural modification. To preserve pre-trained capabilities, we train only these lightweight additions while keeping the remaining backbone frozen. Experiments on EndoVis 2018 demonstrate strong in-domain performance, while evaluation on the out-of-distribution CholecSeg8k further confirms robustness across surgical domains. SCISSR achieves 95.41% Dice on EndoVis 2018 with five interaction rounds and 96.30% Dice on CholecSeg8k with three interaction rounds, outperforming iterative point prompting on both benchmarks. △ Less

Submitted 19 March, 2026; originally announced March 2026.

Surg$Σ$: A Spectrum of Large-Scale Multimodal Data and Foundation Models for Surgical Intelligence

Authors: Zhitao Zeng, Mengya Xu, Jian Jiang, Pengfei Guo, Yunqiu Xu, Zhu Zhuo, Chang Han Low, Yufan He, Dong Yang, Chenxi Lin, Yiming Gu, Jiaxin Guo, Yutong Ban, Daguang Xu, Qi Dou, Yueming Jin

Abstract: Surgical intelligence has the potential to improve the safety and consistency of surgical care, yet most existing surgical AI frameworks remain task-specific and struggle to generalize across procedures and institutions. Although multimodal foundation models, particularly multimodal large language models, have demonstrated strong cross-task capabilities across various medical domains, their advanc… ▽ More Surgical intelligence has the potential to improve the safety and consistency of surgical care, yet most existing surgical AI frameworks remain task-specific and struggle to generalize across procedures and institutions. Although multimodal foundation models, particularly multimodal large language models, have demonstrated strong cross-task capabilities across various medical domains, their advancement in surgery remains constrained by the lack of large-scale, systematically curated multimodal data. To address this challenge, we introduce Surg$Σ$, a spectrum of large-scale multimodal data and foundation models for surgical intelligence. At the core of this framework lies Surg$Σ$-DB, a large-scale multimodal data foundation designed to support diverse surgical tasks. Surg$Σ$-DB consolidates heterogeneous surgical data sources (including open-source datasets, curated in-house clinical collections and web-source data) into a unified schema, aiming to improve label consistency and data standardization across heterogeneous datasets. Surg$Σ$-DB spans 6 clinical specialties and diverse surgical types, providing rich image- and video-level annotations across 18 practical surgical tasks covering understanding, reasoning, planning, and generation, at an unprecedented scale (over 5.98M conversations). Beyond conventional multimodal conversations, Surg$Σ$-DB incorporates hierarchical reasoning annotations, providing richer semantic cues to support deeper contextual understanding in complex surgical scenarios. We further provide empirical evidence through recently developed surgical foundation models built upon Surg$Σ$-DB, illustrating the practical benefits of large-scale multimodal annotations, unified semantic design, and structured reasoning annotations for improving cross-task generalization and interpretability. △ Less

Submitted 17 March, 2026; originally announced March 2026.

MSC Class: 68T45 ACM Class: I.2.10

Generalized Recognition of Basic Surgical Actions Enables Skill Assessment and Vision-Language-Model-based Surgical Planning

Authors: Mengya Xu, Daiyun Shen, Jie Zhang, Hon Chi Yip, Yujia Gao, Cheng Chen, Dillan Imans, Yonghao Long, Yiru Ye, Yixiao Liu, Rongyun Mai, Kai Chen, Hongliang Ren, Yutong Ban, Guangsuo Wang, Francis Wong, Chi-Fai Ng, Kee Yuan Ngiam, Russell H. Taylor, Daguang Xu, Yueming Jin, Qi Dou

Abstract: Artificial intelligence, imaging, and large language models have the potential to transform surgical practice, training, and automation. Understanding and modeling of basic surgical actions (BSA), the fundamental unit of operation in any surgery, is important to drive the evolution of this field. In this paper, we present a BSA dataset comprising 10 basic actions across 6 surgical specialties with… ▽ More Artificial intelligence, imaging, and large language models have the potential to transform surgical practice, training, and automation. Understanding and modeling of basic surgical actions (BSA), the fundamental unit of operation in any surgery, is important to drive the evolution of this field. In this paper, we present a BSA dataset comprising 10 basic actions across 6 surgical specialties with over 11,000 video clips, which is the largest to date. Based on the BSA dataset, we developed a new foundation model that conducts general-purpose recognition of basic actions. Our approach demonstrates robust cross-specialist performance in experiments validated on datasets from different procedural types and various body parts. Furthermore, we demonstrate downstream applications enabled by the BAS foundation model through surgical skill assessment in prostatectomy using domain-specific knowledge, and action planning in cholecystectomy and nephrectomy using large vision-language models. Multinational surgeons' evaluation of the language model's output of the action planning explainable texts demonstrated clinical relevance. These findings indicate that basic surgical actions can be robustly recognized across scenarios, and an accurate BSA understanding model can essentially facilitate complex applications and speed up the realization of surgical superintelligence. △ Less

Submitted 13 March, 2026; originally announced March 2026.

Comments: 34 pages, 8 figures

Surg-R1: A Hierarchical Reasoning Foundation Model for Scalable and Interpretable Surgical Decision Support with Multi-Center Clinical Validation

Authors: Jian Jiang, Chenxi Lin, Yiming Gu, Zengyi Qin, Zhitao Zeng, Kun Yuan, Yonghao Long, Xiang Xia, Cheng Yuan, Yuqi Wang, Zijie Yue, Kunyi Yang, Yuting Zhang, Zhu Zhuo, Dian Qin, Xin Wang, NG Chi Fai, Brian Anthony, Daguang Xu, Guy Rosman, Ozanan Meireles, Zizhen Zhang, Nicolas Padoy, Hesheng Wang, Qi Dou , et al. (2 additional authors not shown)

Abstract: Surgical scene understanding demands not only accurate predictions but also interpretable reasoning that surgeons can verify against clinical expertise. However, existing surgical vision-language models generate predictions without reasoning chains, and general-purpose reasoning models fail on compositional surgical tasks without domain-specific knowledge. We present Surg-R1, a surgical Vision-Lan… ▽ More Surgical scene understanding demands not only accurate predictions but also interpretable reasoning that surgeons can verify against clinical expertise. However, existing surgical vision-language models generate predictions without reasoning chains, and general-purpose reasoning models fail on compositional surgical tasks without domain-specific knowledge. We present Surg-R1, a surgical Vision-Language Model that addresses this gap through hierarchical reasoning trained via a four-stage pipeline. Our approach introduces three key contributions: (1) a three-level reasoning hierarchy decomposing surgical interpretation into perceptual grounding, relational understanding, and contextual reasoning; (2) the largest surgical chain-of-thought dataset with 320,000 reasoning pairs; and (3) a four-stage training pipeline progressing from supervised fine-tuning to group relative policy optimization and iterative self-improvement. Evaluation on SurgBench, comprising six public benchmarks and six multi-center external validation datasets from five institutions, demonstrates that Surg-R1 achieves the highest Arena Score (64.9%) on public benchmarks versus Gemini 3.0 Pro (46.1%) and GPT-5.1 (37.9%), outperforming both proprietary reasoning models and specialized surgical VLMs on the majority of tasks spanning instrument localization, triplet recognition, phase recognition, action recognition, and critical view of safety assessment, with a 15.2 percentage point improvement over the strongest surgical baseline on external validation. △ Less

Submitted 12 March, 2026; originally announced March 2026.

OnFly: Onboard Zero-Shot Aerial Vision-Language Navigation toward Safety and Efficiency

Authors: Guiyong Zheng, Yueting Ban, Mingjie Zhang, Juepeng Zheng, Boyu Zhou

Abstract: Aerial vision-language navigation (AVLN) enables UAVs to follow natural-language instructions in complex 3D environments. However, existing zero-shot AVLN methods often suffer from unstable single-stream Vision-Language Model decision-making, unreliable long-horizon progress monitoring, and a trade-off between safety and efficiency. We propose OnFly, a fully onboard, real-time framework for zero-s… ▽ More Aerial vision-language navigation (AVLN) enables UAVs to follow natural-language instructions in complex 3D environments. However, existing zero-shot AVLN methods often suffer from unstable single-stream Vision-Language Model decision-making, unreliable long-horizon progress monitoring, and a trade-off between safety and efficiency. We propose OnFly, a fully onboard, real-time framework for zero-shot AVLN. OnFly adopts a shared-perception dual-agent architecture that decouples high-frequency target generation from low-frequency progress monitoring, thereby stabilizing decision-making. It further employs a hybrid keyframe-recent-frame memory to preserve global trajectory context while maintaining KV-cache prefix stability, enabling reliable long-horizon monitoring with termination and recovery signals. In addition, a semantic-geometric verifier refines VLM-predicted targets for instruction consistency and geometric safety using VLM features and depth cues, while a receding-horizon planner generates optimized collision-free trajectories under geometric safety constraints, improving both safety and efficiency. In simulation, OnFly improves task success from 26.4% to 67.8%, compared with the strongest state-of-the-art baseline, while fully onboard real-world flights validate its feasibility for real-time deployment. The code will be released at https://github.com/Robotics-STAR-Lab/OnFly △ Less

Submitted 11 March, 2026; originally announced March 2026.

Does the Question Really Matter? Training-Free Data Selection for Vision-Language SFT

Authors: Peng Sun, Huawen Shen, Yi Ban, Tianfan Fu, Yanbo Wang, Yuqiang Li

Abstract: Visual instruction tuning is crucial for improving vision-language large models (VLLMs). However, many samples can be solved via linguistic patterns or common-sense shortcuts, without genuine cross-modal reasoning, limiting the effectiveness of multimodal learning. Prior data selection methods often rely on costly proxy model training and focus on difficulty or diversity, failing to capture a samp… ▽ More Visual instruction tuning is crucial for improving vision-language large models (VLLMs). However, many samples can be solved via linguistic patterns or common-sense shortcuts, without genuine cross-modal reasoning, limiting the effectiveness of multimodal learning. Prior data selection methods often rely on costly proxy model training and focus on difficulty or diversity, failing to capture a sample's true contribution to vision-language joint reasoning. In this paper, we propose CVS, a training-free data selection method based on the insight that, for high-quality multimodal samples, introducing the question should substantially alter the model's assessment of answer validity given an image. CVS leverages a frozen VLLM as an evaluator and measures the discrepancy in answer validity with and without conditioning on the question, enabling the identification of samples that require vision-language joint reasoning while filtering semantic-conflict noise. Experiments on Vision-Flan and The Cauldron show that CVS achieves solid performance across datasets. On Vision-Flan, CVS outperforms full-data training by 3.5% and 4.8% using only 10% and 15% of the data, respectively, and remains robust on the highly heterogeneous Cauldron dataset. Moreover, CVS reduces computational cost by 17.3% and 44.4% compared to COINCIDE and XMAS. △ Less

Submitted 10 March, 2026; originally announced March 2026.

CDRRM: Contrast-Driven Rubric Generation for Reliable and Interpretable Reward Modeling

Authors: Dengcan Liu, Fengkai Yang, Xiaohan Wang, Shurui Yan, Jiajun Chai, Jiahao Li, Yikun Ban, Zhendong Mao, Wei Lin, Guojun Yin

Abstract: Reward modeling is essential for aligning Large Language Models(LLMs) with human preferences, yet conventional reward models suffer from poor interpretability and heavy reliance on costly expert annotations. While recent rubric-based approaches enhance evaluation transparency, they lack systematic quality control, yielding noisy and redundant criteria, failing to mitigate persistent biases (e.g.,… ▽ More Reward modeling is essential for aligning Large Language Models(LLMs) with human preferences, yet conventional reward models suffer from poor interpretability and heavy reliance on costly expert annotations. While recent rubric-based approaches enhance evaluation transparency, they lack systematic quality control, yielding noisy and redundant criteria, failing to mitigate persistent biases (e.g., verbosity, position) in LLM evaluators, and creating a scalability-reliability trade-off. To address these limitations, we propose CDRRM (Contrast-Driven Rubric Reward Model), a framework built on a novel Contrast-then-Synthesis paradigm for high-quality rubric generation and guided preference judgment. CDRRM first conducts multi-dimensional contrastive profiling on preference pairs to identify causal discriminative factors, then synthesizes these insights into compact, context-aware rubrics to guide preference judg- ments. Extensive experiments on three authoritative benchmarks (RewardBench, RMBench, RMB) demonstrate that CDRRM achieves state-of-the-art performance across diverse domains and effectively mitigates aforementioned evaluation biases. Notably, our approach delivers exceptional data efficiency: training the rubric generator on only 3k high-quality samples empowers a frozen pre-trained judge model to outperform fully fine-tuned baselines. This work offers a scalable, interpretable, and data-efficient path for reward modeling. △ Less

Submitted 9 March, 2026; originally announced March 2026.

SSP: Safety-guaranteed Surgical Policy via Joint Optimization of Behavioral and Spatial Constraints

Authors: Jianshu Hu, ZhiYuan Guan, Lei Song, Kantaphat Leelakunwet, Hesheng Wang, Wei Xiao, Qi Dou, Yutong Ban

Abstract: The paradigm of robot-assisted surgery is shifting toward data-driven autonomy, where policies learned via Reinforcement Learning (RL) or Imitation Learning (IL) enable the execution of complex tasks. However, these ``black-box" policies often lack formal safety guarantees, a critical requirement for clinical deployment. In this paper, we propose the Safety-guaranteed Surgical Policy (SSP) framewo… ▽ More The paradigm of robot-assisted surgery is shifting toward data-driven autonomy, where policies learned via Reinforcement Learning (RL) or Imitation Learning (IL) enable the execution of complex tasks. However, these ``black-box" policies often lack formal safety guarantees, a critical requirement for clinical deployment. In this paper, we propose the Safety-guaranteed Surgical Policy (SSP) framework to bridge the gap between data-driven generality and formal safety. We utilize Neural Ordinary Differential Equations (Neural ODEs) to learn an uncertainty-aware dynamics model from demonstration data. This learned model underpins a robust Control Barrier Function (CBF) safety controller, which minimally alters the actions of a surgical policy to ensure strict safety under uncertainty. Our controller enforces two constraint categories: behavioral constraints (restricting the task space of the agent) and spatial constraints (defining surgical no-go zones). We instantiate the SSP framework with surgical policies derived from RL, IL and Control Lyapunov Functions (CLF). Validation on in both the SurRoL simulation and da Vinci Research Kit (dVRK) demonstrates that our method achieves a near-zero constraint violation rate while maintaining high task success rates compared to unconstrained baselines. △ Less

Submitted 6 March, 2026; originally announced March 2026.

Heterogeneous Agent Collaborative Reinforcement Learning

Authors: Zhixia Zhang, Zixuan Huang, Xin Xia, Deqing Wang, Fuzhen Zhuang, Shuai Ma, Ning Ding, Yaodong Yang, Jianxin Li, Yikun Ban

Abstract: We introduce Heterogeneous Agent Collaborative Reinforcement Learning (HACRL), a new learning paradigm that addresses the inefficiencies of isolated on-policy optimization. HACRL enables collaborative optimization with independent execution: heterogeneous agents share verified rollouts during training to mutually improve, while operating independently at inference time. Unlike LLM-based multi-agen… ▽ More We introduce Heterogeneous Agent Collaborative Reinforcement Learning (HACRL), a new learning paradigm that addresses the inefficiencies of isolated on-policy optimization. HACRL enables collaborative optimization with independent execution: heterogeneous agents share verified rollouts during training to mutually improve, while operating independently at inference time. Unlike LLM-based multi-agent reinforcement learning (MARL), HACRL does not require coordinated deployment, and unlike on-/off-policy distillation, it enables bidirectional mutual learning among heterogeneous agents rather than one-directional teacher-to-student transfer. Building on this paradigm, we propose HACPO, a collaborative RL algorithm that enables principled rollout sharing to maximize sample utilization and cross-agent knowledge transfer. To mitigate capability discrepancies and policy distribution shifts, HACPO introduces four tailored mechanisms with theoretical guarantees on unbiased advantage estimation and optimization correctness. Extensive experiments across diverse heterogeneous model combinations and reasoning benchmarks show that HACPO consistently improves all participating agents, outperforming GSPO by an average of 3.3\% while using only half the rollout cost. △ Less

Submitted 3 March, 2026; originally announced March 2026.

UniFAR: A Unified Facet-Aware Retrieval Framework for Scientific Documents

Authors: Zheng Dou, Zhao Zhang, Deqing Wang, Yikun Ban, Fuzhen Zhuang

Abstract: Existing scientific document retrieval (SDR) methods primarily rely on document-centric representations learned from inter-document relationships for document-document (doc-doc) retrieval. However, the rise of LLMs and RAG has shifted SDR toward question-driven retrieval, where documents are retrieved in response to natural-language questions (q-doc). This change has led to systematic mismatches b… ▽ More Existing scientific document retrieval (SDR) methods primarily rely on document-centric representations learned from inter-document relationships for document-document (doc-doc) retrieval. However, the rise of LLMs and RAG has shifted SDR toward question-driven retrieval, where documents are retrieved in response to natural-language questions (q-doc). This change has led to systematic mismatches between document-centric models and question-driven retrieval, including (1) input granularity (long documents vs. short questions), (2) semantic focus (scientific discourse structure vs. specific question intent), and (3) training signals (citation-based similarity vs. question-oriented relevance). To this end, we propose UniFAR, a Unified Facet-Aware Retrieval framework to jointly support doc-doc and q-doc SDR within a single architecture. UniFAR reconciles granularity differences through adaptive multi-granularity aggregation, aligns document structure with question intent via learnable facet anchors, and unifies doc-doc and q-doc supervision through joint training. Experimental results show that UniFAR consistently outperforms prior methods across multiple retrieval tasks and base models, confirming its effectiveness and generality. △ Less

Submitted 27 February, 2026; originally announced February 2026.

Federated Reasoning Distillation Framework with Model Learnability-Aware Data Allocation

Authors: Wei Guo, Siyuan Lu, Xiangdong Ran, Yiqi Tong, Yikun Ban, Zelong Xu, Jing Fan, Zixuan Huang, Xiao Zhang, Zhaojun Hu, Fuzhen Zhuang

Abstract: Data allocation plays a critical role in federated large language model (LLM) and small language models (SLMs) reasoning collaboration. Nevertheless, existing data allocation methods fail to address an under-explored challenge in collaboration: bidirectional model learnability gap, where client-side SLMs cannot identify high-reward samples matching their learnability constraints for effective know… ▽ More Data allocation plays a critical role in federated large language model (LLM) and small language models (SLMs) reasoning collaboration. Nevertheless, existing data allocation methods fail to address an under-explored challenge in collaboration: bidirectional model learnability gap, where client-side SLMs cannot identify high-reward samples matching their learnability constraints for effective knowledge transfer from LLMs, while LLMs struggle to select samples contributing novel knowledge beyond their existing data. Furthermore, these collaboration frameworks face another key challenge: domain-agnostic reasoning transfer, where existing reasoning transfer methods fail to flexibly adapt to the local domain data, preventing SLMs from effectively acquiring step-by-step reasoning abilities within from general LLM. To address these challenges, we propose LaDa, a federated reasoning distillation framework with model learnability-aware data allocation. It introduces a model learnability-aware data filter that adaptively allocates high-reward samples based on the learnability gap between each SLM and LLM pair, effectively facilitating bidirectional knowledge transfer. We further design a domain adaptive reasoning distillation method that aligns joint probabilities of reasoning paths on filtered high-reward samples through contrastive distillation learning between SLM and LLM, enabling SLM to capture underlying reasoning patterns under local data distribution. LaDa operates as a plug-in module for existing collaboration frameworks, adapting knowledge transfer based on model learnability gaps. △ Less

Submitted 21 February, 2026; originally announced February 2026.

UniARM: Towards a Unified Autoregressive Reward Model for Multi-Objective Test-Time Alignment

Authors: Hongyan Xie, Yikun Ban, Ruiyu Fang, Zixuan Huang, Deqing Wang, Jianxin Li, Yitong Yao, Chao Wang, Shuangyong Song

Abstract: Multi-objective alignment aims to align LLM responses with multiple human preference objectives. Among existing methods, guiding the generation of frozen LLMs through autoregressive reward models (ARMs) to accomplish multi-objective test-time alignment is a low-cost solution. However, these methods typically rely on independent parameters for each preference objective, either by training ARMs inde… ▽ More Multi-objective alignment aims to align LLM responses with multiple human preference objectives. Among existing methods, guiding the generation of frozen LLMs through autoregressive reward models (ARMs) to accomplish multi-objective test-time alignment is a low-cost solution. However, these methods typically rely on independent parameters for each preference objective, either by training ARMs independently across preference dimensions, which neglects interactions among preference features, or by training a single ARM with separate feature extraction modules for each preference, which can cause feature entanglement. Both strategies can result in misalignment between generated outputs and user preferences. To address this limitation, we propose Preference-Modulated \& Shared Low-Rank Adaptation (MoSLoRA) for ARM training, which first extracts shared features via a preference-agnostic module and then applies affine transformations to shared features via a preference modulation module conditioned on mixed preference vectors. This design mitigates feature entanglement and enables precise control over preference trade-offs during inference. Building on this, we introduce the Unified Autoregressive Reward Model (UniARM), a novel framework for multi-objective test-time alignment. UniARM jointly models all preference dimensions in a single parameter space, eliminating the need for independent parameters for each preference objective. es on larger-scale LLMs, enhancing its practical usability. △ Less

Submitted 10 February, 2026; originally announced February 2026.

Comments: Under Review

Camel: Frame-Level Bandwidth Estimation for Low-Latency Live Streaming under Video Bitrate Undershooting

Authors: Liming Liu, Zhidong Jia, Li Jiang, Wei Zhang, Lan Xie, Feng Qian, Leju Yan, Bing Yan, Qiang Ma, Zhou Sha, Wei Yang, Yixuan Ban, Xinggong Zhang

Abstract: Low-latency live streaming (LLS) has emerged as a popular web application, with many platforms adopting real-time protocols such as WebRTC to minimize end-to-end latency. However, we observe a counter-intuitive phenomenon: even when the actual encoded bitrate does not fully utilize the available bandwidth, stalling events remain frequent. This insufficient bandwidth utilization arises from the int… ▽ More Low-latency live streaming (LLS) has emerged as a popular web application, with many platforms adopting real-time protocols such as WebRTC to minimize end-to-end latency. However, we observe a counter-intuitive phenomenon: even when the actual encoded bitrate does not fully utilize the available bandwidth, stalling events remain frequent. This insufficient bandwidth utilization arises from the intrinsic temporal variations of real-time video encoding, which cause conventional packet-level congestion control algorithms to misestimate available bandwidth. When a high-bitrate frame is suddenly produced, sending at the wrong rate can either trigger packet loss or increase queueing delay, resulting in playback stalls. To address these issues, we present Camel, a novel frame-level congestion control algorithm (CCA) tailored for LLS. Our insight is to use frame-level network feedback to capture the true network capacity, immune to the irregular sending pattern caused by encoding. Camel comprises three key modules: the Bandwidth and Delay Estimator and the Congestion Detector, which jointly determine the average sending rate, and the Bursting Length Controller, which governs the emission pattern to prevent packet loss. We evaluate Camel on both large-scale real-world deployments and controlled simulations. In the real-world platform with 250M users and 2B sessions across 150+ countries, Camel achieves up to a 70.8% increase in 1080P resolution ratio, a 14.4% increase in media bitrate, and up to a 14.1% reduction in stalling ratio. In simulations under undershooting, shallow buffers, and network jitter, Camel outperforms existing congestion control algorithms, with up to 19.8% higher bitrate, 93.0% lower stalling ratio, and 23.9% improvement in bandwidth estimation accuracy. △ Less

Submitted 10 February, 2026; originally announced February 2026.

Comments: 8 pages, 20 figures, to appear in WWW 2026

Journal ref: Proceedings of the ACM Web Conference 2026 (WWW '26)

Contextual Rollout Bandits for Reinforcement Learning with Verifiable Rewards

Authors: Xiaodong Lu, Xiaohan Wang, Jiajun Chai, Guojun Yin, Wei Lin, Zhijun Chen, Yu Luo, Fuzhen Zhuang, Yikun Ban, Deqing Wang

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) is an effective paradigm for improving the reasoning capabilities of large language models. However, existing RLVR methods utilize rollouts in an indiscriminate and short-horizon manner: responses of heterogeneous quality within each prompt are treated uniformly, and historical rollouts are discarded after a single use. This leads to noisy supe… ▽ More Reinforcement Learning with Verifiable Rewards (RLVR) is an effective paradigm for improving the reasoning capabilities of large language models. However, existing RLVR methods utilize rollouts in an indiscriminate and short-horizon manner: responses of heterogeneous quality within each prompt are treated uniformly, and historical rollouts are discarded after a single use. This leads to noisy supervision, poor sample efficiency, and suboptimal policy updates. We address these issues by formulating rollout scheduling in RLVR as a contextual bandit problem and proposing a unified neural scheduling framework that adaptively selects high-value rollouts throughout training. Each rollout is treated as an arm whose reward is defined by the induced performance gain between consecutive optimization steps. The resulting scheduler supports both noise-aware intra-group selection and adaptive global reuse of historical rollouts within a single principled framework. We provide theoretical justification by deriving sublinear regret bounds and showing that enlarging the rollout buffer improves the achievable performance upper bound. Experiments on six mathematical reasoning benchmarks demonstrate consistent gains in performance and training efficiency across multiple RLVR optimization methods. △ Less

Submitted 9 February, 2026; originally announced February 2026.

Does Your Reasoning Model Implicitly Know When to Stop Thinking?

Authors: Zixuan Huang, Xin Xia, Yuxi Ren, Jianbin Zheng, Xuanda Wang, Zhixia Zhang, Hongyan Xie, Songshi Liang, Zehao Chen, Xuefeng Xiao, Fuzhen Zhuang, Jianxin Li, Yikun Ban, Deqing Wang

Abstract: Recent advancements in large reasoning models (LRMs) have greatly improved their capabilities on complex reasoning tasks through Long Chains of Thought (CoTs). However, this approach often results in substantial redundancy, impairing computational efficiency and causing significant delays in real-time applications. Recent studies show that longer reasoning chains are frequently uncorrelated with c… ▽ More Recent advancements in large reasoning models (LRMs) have greatly improved their capabilities on complex reasoning tasks through Long Chains of Thought (CoTs). However, this approach often results in substantial redundancy, impairing computational efficiency and causing significant delays in real-time applications. Recent studies show that longer reasoning chains are frequently uncorrelated with correctness and can even be detrimental to accuracy. In a further in-depth analysis of this phenomenon, we surprisingly uncover and empirically verify that LRMs implicitly know the appropriate time to stop thinking, while this capability is obscured by current sampling paradigms. Motivated by this, we introduce SAGE (Self-Aware Guided Efficient Reasoning), a novel sampling paradigm that unleashes this efficient reasoning potential. Furthermore, integrating SAGE as mixed sampling into group-based reinforcement learning (SAGE-RL) enables SAGE-RL to effectively incorporate SAGE-discovered efficient reasoning patterns into standard pass@1 inference, markedly enhancing both the reasoning accuracy and efficiency of LRMs across multiple challenging mathematical benchmarks. △ Less

Submitted 27 February, 2026; v1 submitted 9 February, 2026; originally announced February 2026.

Weak-Driven Learning: How Weak Agents make Strong Agents Stronger

Authors: Zehao Chen, Gongxun Li, Tianxiang Ai, Yifei Li, Zixuan Huang, Wang Zhou, Fuzhen Zhuang, Xianglong Liu, Jianxin Li, Deqing Wang, Yikun Ban

Abstract: As post-training optimization becomes central to improving large language models, we observe a persistent saturation bottleneck: once models grow highly confident, further training yields diminishing returns. While existing methods continue to reinforce target predictions, we find that informative supervision signals remain latent in models' own historical weak states. Motivated by this observatio… ▽ More As post-training optimization becomes central to improving large language models, we observe a persistent saturation bottleneck: once models grow highly confident, further training yields diminishing returns. While existing methods continue to reinforce target predictions, we find that informative supervision signals remain latent in models' own historical weak states. Motivated by this observation, we propose WMSS (Weak Agents Can Make Strong Agents Stronger), a post-training paradigm that leverages weak checkpoints to guide continued optimization. By identifying recoverable learning gaps via entropy dynamics and reinforcing them through compensatory learning, WMSS enables strong agents to improve beyond conventional post-training saturation. Experiments on mathematical reasoning and code generation datasets show that agents trained with our approach achieve effective performance improvements, while incurring zero additional inference cost. △ Less

Submitted 8 February, 2026; originally announced February 2026.

Real-Time Aligned Reward Model beyond Semantics

Authors: Zixuan Huang, Xin Xia, Yuxi Ren, Jianbin Zheng, Xuefeng Xiao, Hongyan Xie, Li Huaqiu, Songshi Liang, Zhongxiang Dai, Fuzhen Zhuang, Jianxin Li, Yikun Ban, Deqing Wang

Abstract: Reinforcement Learning from Human Feedback (RLHF) is a pivotal technique for aligning large language models (LLMs) with human preferences, yet it is susceptible to reward overoptimization, in which policy models overfit to the reward model, exploit spurious reward patterns instead of faithfully capturing human intent. Prior mitigations primarily relies on surface semantic information and fails to… ▽ More Reinforcement Learning from Human Feedback (RLHF) is a pivotal technique for aligning large language models (LLMs) with human preferences, yet it is susceptible to reward overoptimization, in which policy models overfit to the reward model, exploit spurious reward patterns instead of faithfully capturing human intent. Prior mitigations primarily relies on surface semantic information and fails to efficiently address the misalignment between the reward model (RM) and the policy model caused by continuous policy distribution shifts. This inevitably leads to an increasing reward discrepancy, exacerbating reward overoptimization. To address these limitations, we introduce R2M (Real-Time Aligned Reward Model), a novel lightweight RLHF framework. R2M goes beyond vanilla reward models that solely depend on the semantic representations of a pretrained LLM. Instead, it leverages the evolving hidden states of the policy (namely policy feedback) to align with the real-time distribution shift of the policy during the RL process. This work points to a promising new direction for improving the performance of reward models through real-time utilization of feedback from policy models. △ Less

Submitted 9 March, 2026; v1 submitted 30 January, 2026; originally announced January 2026.

Reward-Forcing: Autoregressive Video Generation with Reward Feedback

Authors: Jingran Zhang, Ning Li, Yuanhao Ban, Andrew Bai, Justin Cui

Abstract: While most prior work in video generation relies on bidirectional architectures, recent efforts have sought to adapt these models into autoregressive variants to support near real-time generation. However, such adaptations often depend heavily on teacher models, which can limit performance, particularly in the absence of a strong autoregressive teacher, resulting in output quality that typically l… ▽ More While most prior work in video generation relies on bidirectional architectures, recent efforts have sought to adapt these models into autoregressive variants to support near real-time generation. However, such adaptations often depend heavily on teacher models, which can limit performance, particularly in the absence of a strong autoregressive teacher, resulting in output quality that typically lags behind their bidirectional counterparts. In this paper, we explore an alternative approach that uses reward signals to guide the generation process, enabling more efficient and scalable autoregressive generation. By using reward signals to guide the model, our method simplifies training while preserving high visual fidelity and temporal consistency. Through extensive experiments on standard benchmarks, we find that our approach performs comparably to existing autoregressive models and, in some cases, surpasses similarly sized bidirectional models by avoiding constraints imposed by teacher architectures. For example, on VBench, our method achieves a total score of 84.92, closely matching state-of-the-art autoregressive methods that score 84.31 but require significant heterogeneous distillation. △ Less

Submitted 2 April, 2026; v1 submitted 23 January, 2026; originally announced January 2026.

Comments: https://openreview.net/forum?id=K8Qjsxxl7y&noteId=K8Qjsxxl7y

LoL: Longer than Longer, Scaling Video Generation to Hour

Authors: Justin Cui, Jie Wu, Ming Li, Tao Yang, Xiaojie Li, Rui Wang, Andrew Bai, Yuanhao Ban, Cho-Jui Hsieh

Abstract: Recent research in long-form video generation has shifted from bidirectional to autoregressive models, yet these methods commonly suffer from error accumulation and a loss of long-term coherence. While attention sink frames have been introduced to mitigate this performance decay, they often induce a critical failure mode we term sink-collapse: the generated content repeatedly reverts to the sink f… ▽ More Recent research in long-form video generation has shifted from bidirectional to autoregressive models, yet these methods commonly suffer from error accumulation and a loss of long-term coherence. While attention sink frames have been introduced to mitigate this performance decay, they often induce a critical failure mode we term sink-collapse: the generated content repeatedly reverts to the sink frame, resulting in abrupt scene resets and cyclic motion patterns. Our analysis reveals that sink-collapse originates from an inherent conflict between the periodic structure of Rotary Position Embedding (RoPE) and the multi-head attention mechanisms prevalent in current generative models. To address it, we propose a lightweight, training-free approach that effectively suppresses this behavior by introducing multi-head RoPE jitter that breaks inter-head attention homogenization and mitigates long-horizon collapse. Extensive experiments show that our method successfully alleviates sink-collapse while preserving generation quality. To the best of our knowledge, this work achieves the first demonstration of real-time, streaming, and infinite-length video generation with little quality decay. As an illustration of this robustness, we generate continuous videos up to 12 hours in length, which, to our knowledge, is among the longest publicly demonstrated results in streaming video generation. △ Less

Submitted 23 January, 2026; originally announced January 2026.

Comments: preprint

Your Group-Relative Advantage Is Biased

Authors: Fengkai Yang, Zherui Chen, Xiaohan Wang, Xiaodong Lu, Jiajun Chai, Guojun Yin, Wei Lin, Shuai Ma, Fuzhen Zhuang, Deqing Wang, Yaodong Yang, Jianxin Li, Yikun Ban

Abstract: Reinforcement Learning from Verifier Rewards (RLVR) has emerged as a widely used approach for post-training large language models on reasoning tasks, with group-based methods such as GRPO and its variants gaining broad adoption. These methods rely on group-relative advantage estimation to avoid learned critics, yet its theoretical properties remain poorly understood. In this work, we uncover a f… ▽ More Reinforcement Learning from Verifier Rewards (RLVR) has emerged as a widely used approach for post-training large language models on reasoning tasks, with group-based methods such as GRPO and its variants gaining broad adoption. These methods rely on group-relative advantage estimation to avoid learned critics, yet its theoretical properties remain poorly understood. In this work, we uncover a fundamental issue of group-based RL: the group-relative advantage estimator is inherently biased relative to the true (expected) advantage. We provide the first theoretical analysis showing that it systematically underestimates advantages for hard prompts and overestimates them for easy prompts, leading to imbalanced exploration and exploitation. To address this issue, we propose History-Aware Adaptive Difficulty Weighting (HA-DW), an adaptive reweighting scheme that adjusts advantage estimates based on an evolving difficulty anchor and training dynamics. Both theoretical analysis and experiments on five mathematical reasoning benchmarks demonstrate that HA-DW consistently improves performance when integrated into GRPO and its variants. Our results suggest that correcting biased advantage estimation is critical for robust and efficient RLVR training. △ Less

Submitted 21 January, 2026; v1 submitted 13 January, 2026; originally announced January 2026.

Scoring, Reasoning, and Selecting the Best! Ensembling Large Language Models via a Peer-Review Process

Authors: Zhijun Chen, Zeyu Ji, Qianren Mao, Hao Wu, Junhang Cheng, Bangjie Qin, Zhuoran Li, Jingzheng Li, Kai Sun, Zizhe Wang, Yikun Ban, Zhu Sun, Xiangyang Ji, Hailong Sun

Abstract: We propose LLM-PeerReview, an unsupervised LLM Ensemble method that selects the most ideal response from multiple LLM-generated candidates for each query, harnessing the collective wisdom of multiple models with diverse strengths. LLM-PeerReview is built on a novel, peer-review-inspired framework that offers a transparent and interpretable mechanism, while remaining fully unsupervised for flexible… ▽ More We propose LLM-PeerReview, an unsupervised LLM Ensemble method that selects the most ideal response from multiple LLM-generated candidates for each query, harnessing the collective wisdom of multiple models with diverse strengths. LLM-PeerReview is built on a novel, peer-review-inspired framework that offers a transparent and interpretable mechanism, while remaining fully unsupervised for flexible adaptability and generalization. Specifically, it operates in three stages: For scoring, we use the emerging LLM-as-a-Judge technique to evaluate each response by reusing multiple LLMs at hand; For reasoning, we can apply a straightforward averaging strategy or a principled graphical model-based truth inference algorithm to aggregate multiple scores to produce a final score for each response; Finally, the highest-scoring response is selected as the best ensemble output. LLM-PeerReview is conceptually simple and empirically powerful. Our results across four datasets show that the two variants of the proposed approach outperform the advanced model Smoothie-Global by 6.9% and 7.3% points, cross diverse task types including factual recall QA, math reasoning, and instruction following. △ Less

Submitted 6 February, 2026; v1 submitted 29 December, 2025; originally announced December 2025.

LLMBoost: Make Large Language Models Stronger with Boosting

Authors: Zehao Chen, Tianxiang Ai, Yifei Li, Gongxun Li, Yuyang Wei, Wang Zhou, Guanghui Li, Bin Yu, Zhijun Chen, Hailong Sun, Fuzhen Zhuang, Jianxin Li, Deqing Wang, Yikun Ban

Abstract: Ensemble learning of LLMs has emerged as a promising alternative to enhance performance, but existing approaches typically treat models as black boxes, combining the inputs or final outputs while overlooking the rich internal representations and interactions across models.In this work, we introduce LLMBoost, a novel ensemble fine-tuning framework that breaks this barrier by explicitly leveraging i… ▽ More Ensemble learning of LLMs has emerged as a promising alternative to enhance performance, but existing approaches typically treat models as black boxes, combining the inputs or final outputs while overlooking the rich internal representations and interactions across models.In this work, we introduce LLMBoost, a novel ensemble fine-tuning framework that breaks this barrier by explicitly leveraging intermediate states of LLMs. Inspired by the boosting paradigm, LLMBoost incorporates three key innovations. First, a cross-model attention mechanism enables successor models to access and fuse hidden states from predecessors, facilitating hierarchical error correction and knowledge transfer. Second, a chain training paradigm progressively fine-tunes connected models with an error-suppression objective, ensuring that each model rectifies the mispredictions of its predecessor with minimal additional computation. Third, a near-parallel inference paradigm design pipelines hidden states across models layer by layer, achieving inference efficiency approaching single-model decoding. We further establish the theoretical foundations of LLMBoost, proving that sequential integration guarantees monotonic improvements under bounded correction assumptions. Extensive experiments on commonsense reasoning and arithmetic reasoning tasks demonstrate that LLMBoost consistently boosts accuracy while reducing inference latency. △ Less

Submitted 26 December, 2025; originally announced December 2025.

OMP: One-step Meanflow Policy with Directional Alignment

Authors: Han Fang, Yize Huang, Yuheng Zhao, Paul Weng, Xiao Li, Yutong Ban

Abstract: Robot manipulation has increasingly adopted data-driven generative policy frameworks, yet the field faces a persistent trade-off: diffusion models suffer from high inference latency, while flow-based methods often require complex architectural constraints. Although in image generation domain, the MeanFlow paradigm offers a path to single-step inference, its direct application to robotics is impede… ▽ More Robot manipulation has increasingly adopted data-driven generative policy frameworks, yet the field faces a persistent trade-off: diffusion models suffer from high inference latency, while flow-based methods often require complex architectural constraints. Although in image generation domain, the MeanFlow paradigm offers a path to single-step inference, its direct application to robotics is impeded by critical theoretical pathologies, specifically spectral bias and gradient starvation in low-velocity regimes. To overcome these limitations, we propose the One-step MeanFlow Policy (OMP), a novel framework designed for high-fidelity, real-time manipulation. We introduce a lightweight directional alignment mechanism to explicitly synchronize predicted velocities with true mean velocities. Furthermore, we implement a Differential Derivation Equation (DDE) to approximate the Jacobian-Vector Product (JVP) operator, which decouples forward and backward passes to significantly reduce memory complexity. Extensive experiments on the Adroit and Meta-World benchmarks demonstrate that OMP outperforms state-of-the-art methods in success rate and trajectory accuracy, particularly in high-precision tasks, while retaining the efficiency of single-step generation. △ Less

Submitted 29 January, 2026; v1 submitted 22 December, 2025; originally announced December 2025.

When Distance Distracts: Representation Distance Bias in BT-Loss for Reward Models

Authors: Tong Xie, Andrew Bai, Yuanhao Ban, Yunqi Hong, Haoyu Li, Cho-jui Hsieh

Abstract: Reward models are central to Large Language Model (LLM) alignment within the framework of RLHF. The standard objective used in reward modeling is the Bradley-Terry (BT) loss, which learns from pairwise data consisting of chosen and rejected responses. In this work, we analyze the per-sample gradient of BT-loss and show spurious learning signals due to representation distance. In particular, BT gra… ▽ More Reward models are central to Large Language Model (LLM) alignment within the framework of RLHF. The standard objective used in reward modeling is the Bradley-Terry (BT) loss, which learns from pairwise data consisting of chosen and rejected responses. In this work, we analyze the per-sample gradient of BT-loss and show spurious learning signals due to representation distance. In particular, BT gradient norm scales with two distinct components: (1) prediction error, reflected by the difference in predicted rewards between chosen and rejected responses, and critically, (2) representation distance between the pair measured in the output space of the final layer. While the first term captures the intended training signal, the second term can significantly impact the update magnitude and misalign learning. Specifically, pairs with small representation distance often receive vanishingly weak updates, even when misranked, while pairs with large distance receive disproportionately strong updates. This leads to gradients from large-distance pairs to overshadow those from small-distance pairs, where fine-grained distinctions are especially important. To overcome this limitation, we propose NormBT, an adaptive pair-wise normalization scheme that rescales updates to balance representation-driven effects and focuses learning signals on prediction error. NormBT is a lightweight, drop-in modification to BT loss with negligible overhead. Across various LLM backbones and datasets, NormBT improves reward model performance consistently, with notable gains of over 5% on the Reasoning category of RewardBench, which contains numerous fine-grained pairs. △ Less

Submitted 31 January, 2026; v1 submitted 6 December, 2025; originally announced December 2025.

See in Depth: Training-Free Surgical Scene Segmentation with Monocular Depth Priors

Authors: Kunyi Yang, Qingyu Wang, Cheng Yuan, Yutong Ban

Abstract: Pixel-wise segmentation of laparoscopic scenes is essential for computer-assisted surgery but difficult to scale due to the high cost of dense annotations. We propose depth-guided surgical scene segmentation (DepSeg), a training-free framework that utilizes monocular depth as a geometric prior together with pretrained vision foundation models. DepSeg first estimates a relative depth map with a pre… ▽ More Pixel-wise segmentation of laparoscopic scenes is essential for computer-assisted surgery but difficult to scale due to the high cost of dense annotations. We propose depth-guided surgical scene segmentation (DepSeg), a training-free framework that utilizes monocular depth as a geometric prior together with pretrained vision foundation models. DepSeg first estimates a relative depth map with a pretrained monocular depth estimation network and proposes depth-guided point prompts, which SAM2 converts into class-agnostic masks. Each mask is then described by a pooled pretrained visual feature and classified via template matching against a template bank built from annotated frames. On the CholecSeg8k dataset, DepSeg improves over a direct SAM2 auto segmentation baseline (35.9% vs. 14.7% mIoU) and maintains competitive performance even when using only 10--20% of the object templates. These results show that depth-guided prompting and template-based classification offer an annotation-efficient segmentation approach. △ Less

Submitted 5 December, 2025; originally announced December 2025.

Comments: The first two authors contributed equally

See, Hear, and Understand: Benchmarking Audiovisual Human Speech Understanding in Multimodal Large Language Models

Authors: Le Thien Phuc Nguyen, Zhuoran Yu, Samuel Low Yu Hang, Subin An, Jeongik Lee, Yohan Ban, SeungEun Chung, Thanh-Huy Nguyen, JuWan Maeng, Soochahn Lee, Yong Jae Lee

Abstract: Multimodal large language models (MLLMs) are expected to jointly interpret vision, audio, and language, yet existing video benchmarks rarely assess fine-grained reasoning about human speech. Many tasks remain visually solvable or only coarsely evaluate speech, offering limited insight into whether models can align who speaks, what is said, and when it occurs. We introduce AV-SpeakerBench, a curate… ▽ More Multimodal large language models (MLLMs) are expected to jointly interpret vision, audio, and language, yet existing video benchmarks rarely assess fine-grained reasoning about human speech. Many tasks remain visually solvable or only coarsely evaluate speech, offering limited insight into whether models can align who speaks, what is said, and when it occurs. We introduce AV-SpeakerBench, a curated benchmark of 3,212 multiple-choice questions focused on speaker-centric audiovisual reasoning in real-world videos. It features: (1) a speaker-centered formulation that treats speakers-not scenes-as the core reasoning unit; (2) fusion-grounded question design embedding audiovisual dependencies into question semantics; and (3) expert-curated annotations ensuring temporal precision and cross-modal validity. Comprehensive evaluations show that the Gemini family consistently outperforms open-source systems, with Gemini 2.5 Pro achieving the best results. Among open models, Qwen3-Omni-30B approaches Gemini 2.0 Flash but remains far behind Gemini 2.5 Pro, primarily due to weaker audiovisual fusion rather than visual perception. We believe AV-SpeakerBench establishes a rigorous foundation for advancing fine-grained audiovisual reasoning in future multimodal systems. △ Less

Submitted 1 December, 2025; originally announced December 2025.

Comments: preprint

GCL-OT: Graph Contrastive Learning with Optimal Transport for Heterophilic Text-Attributed Graphs

Authors: Yating Ren, Yikun Ban, Huobin Tan

Abstract: Recently, structure-text contrastive learning has shown promising performance on text-attributed graphs by leveraging the complementary strengths of graph neural networks and language models. However, existing methods typically rely on homophily assumptions in similarity estimation and hard optimization objectives, which limit their applicability to heterophilic graphs. Although existing methods c… ▽ More Recently, structure-text contrastive learning has shown promising performance on text-attributed graphs by leveraging the complementary strengths of graph neural networks and language models. However, existing methods typically rely on homophily assumptions in similarity estimation and hard optimization objectives, which limit their applicability to heterophilic graphs. Although existing methods can mitigate heterophily through structural adjustments or neighbor aggregation, they usually treat textual embeddings as static targets, leading to suboptimal alignment. In this work, we identify multi-granular heterophily in text-attributed graphs, including complete heterophily, partial heterophily, and latent homophily, which makes structure-text alignment particularly challenging due to mixed, noisy, and missing semantic correlations. To achieve flexible and bidirectional alignment, we propose GCL-OT, a novel graph contrastive learning framework with optimal transport, equipped with tailored mechanisms for each type of heterophily. Specifically, for partial heterophily, we design a RealSoftMax-based similarity estimator to emphasize key neighbor-word interactions while easing background noise. For complete heterophily, we introduce a prompt-based filter that adaptively excludes irrelevant noise during optimal transport alignment. Furthermore, we incorporate OT-guided soft supervision to uncover potential neighbors with similar semantics, enhancing the learning of latent homophily. Theoretical analysis shows that GCL-OT can improve the mutual information bound and Bayes error guarantees. Extensive experiments on nine benchmarks show that GCL-OT outperforms state-of-the-art methods, demonstrating its effectiveness and robustness. △ Less

Submitted 28 January, 2026; v1 submitted 20 November, 2025; originally announced November 2025.

Comments: AAAI 2026

OSGym: Scalable OS Infra for Computer Use Agents

Authors: Zengyi Qin, Jinyuan Chen, Yunze Man, Shengcao Cao, Ziqi Pang, Zhuoyuan Wang, Han Fang, Ling Zhu, Zixin Xie, Zibu Wei, Tianshu Ran, Haoran Geng, Ray Pan, Qizhen Sun, Zachary Bright, Yuyang Cai, Chongye Yang, Jiace Zhao, Tianrui Liu, Han Cao, Yeyang Zhou, Rui Wang, Song Wang, Xiang Ren, Bo Zhang , et al. (3 additional authors not shown)

Abstract: Training computer use agents requires full-featured OS sandboxes with GUI environments, which consume substantial hardware resources as the number of sandboxes scales. Stochastic errors arising from diverse software execution within these sandboxes further demand robust infrastructure design and reliable error recovery. We present OSGym, a scalable OS environment infrastructure for computer use ag… ▽ More Training computer use agents requires full-featured OS sandboxes with GUI environments, which consume substantial hardware resources as the number of sandboxes scales. Stochastic errors arising from diverse software execution within these sandboxes further demand robust infrastructure design and reliable error recovery. We present OSGym, a scalable OS environment infrastructure for computer use agents, built around these key optimization strategies: (1) Decentralized OS state management, which isolates failures to individual replicas and significantly enhances overall system reliability; (2) Hardware-aware OS replica orchestration, which addresses CPU-bounded scaling bottlenecks and substantially reduces compute overhead; (3) KVM virtualization with copy-on-write disk management, which shares a common bootable disk across VM instances and provisions only instance-specific modifications, reducing physical disk consumption by 88% and increasing disk provisioning speed by 37 times; and (4) Robust container pool with multi-layer fault recovery. Together, these optimizations yield strong scalability and resource efficiency: OSGym manages over a thousand OS replicas under constrained resources, supports parallel trajectory generation at 1420 multi-turn trajectories per minute, and reduces per-replica cost to 0.2-0.3 USD per day, a 90% reduction over standard deployment. Our experiments validate OSGym across end-to-end pipelines for data collection and training for computer use agents. We believe OSGym establishes a new foundation for scalable, general-purpose computer use agent research. △ Less

Submitted 1 April, 2026; v1 submitted 11 November, 2025; originally announced November 2025.

Relaxed Sequence Sampling for Diverse Protein Design

Authors: Joohwan Ko, Aristofanis Rontogiannis, Yih-En Andrew Ban, Axel Elaldi, Nicholas Franklin

Abstract: Protein design using structure prediction models such as AlphaFold2 has shown remarkable success, but existing approaches like relaxed sequence optimization (RSO) rely on single-path gradient descent and ignore sequence-space constraints, limiting diversity and designability. We introduce Relaxed Sequence Sampling (RSS), a Markov chain Monte Carlo (MCMC) framework that integrates structural and ev… ▽ More Protein design using structure prediction models such as AlphaFold2 has shown remarkable success, but existing approaches like relaxed sequence optimization (RSO) rely on single-path gradient descent and ignore sequence-space constraints, limiting diversity and designability. We introduce Relaxed Sequence Sampling (RSS), a Markov chain Monte Carlo (MCMC) framework that integrates structural and evolutionary information for protein design. RSS operates in continuous logit space, combining gradient-guided exploration with protein language model-informed jumps. Its energy function couples AlphaFold2-derived structural objectives with ESM2-derived sequence priors, balancing accuracy and biological plausibility. In an in silico protein binder design task, RSS produces 5$\times$ more designable structures and 2-3$\times$ greater structural diversity than RSO baselines, at equal computational cost. These results highlight RSS as a principled approach for efficiently exploring the protein design landscape. △ Less

Submitted 27 October, 2025; originally announced October 2025.

Self-Forcing++: Towards Minute-Scale High-Quality Video Generation

Authors: Justin Cui, Jie Wu, Ming Li, Tao Yang, Xiaojie Li, Rui Wang, Andrew Bai, Yuanhao Ban, Cho-Jui Hsieh

Abstract: Diffusion models have revolutionized image and video generation, achieving unprecedented visual quality. However, their reliance on transformer architectures incurs prohibitively high computational costs, particularly when extending generation to long videos. Recent work has explored autoregressive formulations for long video generation, typically by distilling from short-horizon bidirectional tea… ▽ More Diffusion models have revolutionized image and video generation, achieving unprecedented visual quality. However, their reliance on transformer architectures incurs prohibitively high computational costs, particularly when extending generation to long videos. Recent work has explored autoregressive formulations for long video generation, typically by distilling from short-horizon bidirectional teachers. Nevertheless, given that teacher models cannot synthesize long videos, the extrapolation of student models beyond their training horizon often leads to pronounced quality degradation, arising from the compounding of errors within the continuous latent space. In this paper, we propose a simple yet effective approach to mitigate quality degradation in long-horizon video generation without requiring supervision from long-video teachers or retraining on long video datasets. Our approach centers on exploiting the rich knowledge of teacher models to provide guidance for the student model through sampled segments drawn from self-generated long videos. Our method maintains temporal consistency while scaling video length by up to 20x beyond teacher's capability, avoiding common issues such as over-exposure and error-accumulation without recomputing overlapping frames like previous methods. When scaling up the computation, our method shows the capability of generating videos up to 4 minutes and 15 seconds, equivalent to 99.9% of the maximum span supported by our base model's position embedding and more than 50x longer than that of our baseline model. Experiments on standard benchmarks and our proposed improved benchmark demonstrate that our approach substantially outperforms baseline methods in both fidelity and consistency. Our long-horizon videos demo can be found at https://self-forcing-plus-plus.github.io/ △ Less

Submitted 2 October, 2025; originally announced October 2025.

Comments: preprint

IRIS: Intrinsic Reward Image Synthesis

Authors: Yihang Chen, Yuanhao Ban, Yunqi Hong, Cho-Jui Hsieh

Abstract: Despite the success of Reinforcement Learning from Human Feedback (RLHF) in language reasoning, its application to autoregressive Text-to-Image (T2I) generation is often constrained by the limited availability of human preference data. This paper explores how an autoregressive T2I model can learn from internal signals without relying on external rewards or labeled data. Contrary to recent findings… ▽ More Despite the success of Reinforcement Learning from Human Feedback (RLHF) in language reasoning, its application to autoregressive Text-to-Image (T2I) generation is often constrained by the limited availability of human preference data. This paper explores how an autoregressive T2I model can learn from internal signals without relying on external rewards or labeled data. Contrary to recent findings in math and code reasoning, we show that minimizing self-certainty, rather than maximizing it, improves image generation. We observe that autoregressive T2I models with higher certainty are likely to generate simple and uniform images, which are less aligned with human preferences, and models with lower certainty are likely to generate vivid images rich in detail. Based on this observation, we propose IRIS(Intrinsic Reward Image Synthesis), the first framework to improve autoregressive T2I models with reinforcement learning using only an intrinsic reward. Empirical results demonstrate that applying IRIS to autoregressive T2I models achieves performance superior to those trained by individual external rewards, and matching those trained by ensemble external rewards. IRIS also incentivizes the emergence of nuanced CoT reasoning for high-quality image generation. △ Less

Submitted 29 January, 2026; v1 submitted 29 September, 2025; originally announced September 2025.

FedPOB: Sample-Efficient Federated Prompt Optimization via Bandits

Authors: Pingchen Lu, Zhi Hong, Zhiwei Shang, Zhiyong Wang, Yikun Ban, Yao Shu, Min Zhang, Shuang Qiu, Zhongxiang Dai

Abstract: The performance of large language models (LLMs) is highly sensitive to the input prompt, making prompt optimization a critical task. However, real-world application is hindered by three major challenges: (1) the black-box nature of powerful proprietary LLMs, (2) the need for high sample efficiency due to query costs, and (3) the desire for privacy-preserving collaboration among multiple users. To… ▽ More The performance of large language models (LLMs) is highly sensitive to the input prompt, making prompt optimization a critical task. However, real-world application is hindered by three major challenges: (1) the black-box nature of powerful proprietary LLMs, (2) the need for high sample efficiency due to query costs, and (3) the desire for privacy-preserving collaboration among multiple users. To address these challenges simultaneously, we introduce a novel framework for sample-efficient federated prompt optimization based on multi-armed bandits (MABs). The MAB framework is uniquely suited for this problem as it is (1) inherently a black-box optimization method, (2) practically sample-efficient, and (3) enables collaborative learning with theoretically guaranteed benefit from more participating agents. We first propose the Federated Prompt Optimization via Bandits (FedPOB) algorithm, a federated variant of the Linear UCB algorithm, where agents collaborate by sharing model parameters instead of raw data. We then extend our approach to the practical setting of comparative user feedback by introducing FedPOB with Preference Feedback (FedPOB-Pref), an efficient algorithm based on federated dueling bandits. Extensive experiments demonstrate that both FedPOB and FedPOB-Pref significantly outperform existing baselines and that their performance consistently improves as more agents participate in the collaboration, validating the effectiveness of our federated approach. △ Less

Submitted 29 September, 2025; originally announced September 2025.

Comments: Preprint

T-POP: Test-Time Personalization with Online Preference Feedback

Authors: Zikun Qu, Min Zhang, Mingze Kong, Xiang Li, Zhiwei Shang, Zhiyong Wang, Yikun Ban, Shuang Qiu, Yao Shu, Zhongxiang Dai

Abstract: Personalizing large language models (LLMs) to individual user preferences is a critical step beyond generating generically helpful responses. However, current personalization methods are ill-suited for new users, as they typically require either slow, resource-intensive fine-tuning or a substantial amount of pre-existing user data, creating a significant cold-start problem. To address this challen… ▽ More Personalizing large language models (LLMs) to individual user preferences is a critical step beyond generating generically helpful responses. However, current personalization methods are ill-suited for new users, as they typically require either slow, resource-intensive fine-tuning or a substantial amount of pre-existing user data, creating a significant cold-start problem. To address this challenge, we introduce a new paradigm for real-time personalization by learning from online pairwise preference feedback collected during text generation. We propose T-POP (Test-Time Personalization with Online Preference Feedback}), a novel algorithm that synergistically combines test-time alignment with dueling bandits. Without updating the LLM parameters, T-POP steers the decoding process of a frozen LLM by learning a reward function online that captures user preferences. By leveraging dueling bandits, T-POP intelligently queries the user to efficiently balance between exploring their preferences and exploiting the learned knowledge to generate personalized text. Extensive experiments demonstrate that T-POP achieves rapid and data-efficient personalization, significantly outperforming existing baselines and showing consistent improvement with more user interactions. △ Less

Submitted 29 September, 2025; originally announced September 2025.

Comments: Preprint

Generalizable Coarse-to-Fine Robot Manipulation via Language-Aligned 3D Keypoints

Authors: Jianshu Hu, Lidi Wang, Shujia Li, Yunpeng Jiang, Xiao Li, Paul Weng, Yutong Ban

Abstract: Hierarchical coarse-to-fine policy, where a coarse branch predicts a region of interest to guide a fine-grained action predictor, has demonstrated significant potential in robotic 3D manipulation tasks by especially enhancing sample efficiency and enabling more precise manipulation. However, even augmented with pre-trained models, these hierarchical policies still suffer from generalization issues… ▽ More Hierarchical coarse-to-fine policy, where a coarse branch predicts a region of interest to guide a fine-grained action predictor, has demonstrated significant potential in robotic 3D manipulation tasks by especially enhancing sample efficiency and enabling more precise manipulation. However, even augmented with pre-trained models, these hierarchical policies still suffer from generalization issues. To enhance generalization to novel instructions and environment variations, we propose Coarse-to-fine Language-Aligned manipulation Policy (CLAP), a framework that integrates three key components: 1) task decomposition, 2) VLM fine-tuning for 3D keypoint prediction, and 3) 3D-aware representation. Through comprehensive experiments in simulation and on a real robot, we demonstrate its superior generalization capability. Specifically, on GemBench, a benchmark designed for evaluating generalization, our approach achieves a 12\% higher average success rate than the SOTA method while using only 1/5 of the training trajectories. In real-world experiments, our policy, trained on only 10 demonstrations, successfully generalizes to novel instructions and environments. △ Less

Submitted 20 February, 2026; v1 submitted 27 September, 2025; originally announced September 2025.

Comments: Published in ICLR 2026

The SAGES Critical View of Safety Challenge: A Global Benchmark for AI-Assisted Surgical Quality Assessment

Authors: Deepak Alapatt, Jennifer Eckhoff, Zhiliang Lyu, Yutong Ban, Jean-Paul Mazellier, Sarah Choksi, Kunyi Yang, Po-Hsing Chiang, Noemi Zorzetti, Samuele Cannas, Daniel Neimark, Omri Bar, Amine Yamlahi, Jakob Hennighausen, Xiaohan Wang, Rui Li, Long Liang, Yuxian Wang, Saurabh Koju, Binod Bhattarai, Tim Jaspers, Zhehua Mao, Anjana Wijekoon, Jun Ma, Yinan Xu , et al. (16 additional authors not shown)

Abstract: Advances in artificial intelligence (AI) for surgical quality assessment promise to democratize access to expertise, with applications in training, guidance, and accreditation. This study presents the SAGES Critical View of Safety (CVS) Challenge, the first AI competition organized by a surgical society, using the CVS in laparoscopic cholecystectomy, a universally recommended yet inconsistently pe… ▽ More Advances in artificial intelligence (AI) for surgical quality assessment promise to democratize access to expertise, with applications in training, guidance, and accreditation. This study presents the SAGES Critical View of Safety (CVS) Challenge, the first AI competition organized by a surgical society, using the CVS in laparoscopic cholecystectomy, a universally recommended yet inconsistently performed safety step, as an exemplar of surgical quality assessment. A global collaboration across 54 institutions in 24 countries engaged hundreds of clinicians and engineers to curate 1,000 videos annotated by 20 surgical experts according to a consensus-validated protocol. The challenge addressed key barriers to real-world deployment in surgery, including achieving high performance, capturing uncertainty in subjective assessment, and ensuring robustness to clinical variability. To enable this scale of effort, we developed EndoGlacier, a framework for managing large, heterogeneous surgical video and multi-annotator workflows. Thirteen international teams participated, achieving up to a 17% relative gain in assessment performance, over 80% reduction in calibration error, and a 17% relative improvement in robustness over the state-of-the-art. Analysis of results highlighted methodological trends linked to model performance, providing guidance for future research toward robust, clinically deployable AI for surgical quality assessment. △ Less

Submitted 28 January, 2026; v1 submitted 21 September, 2025; originally announced September 2025.

Comments: 21 pages, 10 figures

MSC Class: 68T07 ACM Class: I.2.10; J.3

Mitigating Spurious Correlations Between Question and Answer via Chain-of-Thought Correctness Perception Distillation

Authors: Hongyan Xie, Yitong Yao, Yikun Ban, Zixuan Huang, Deqing Wang, Zhenhe Wu, Haoxiang Su, Chao Wang, Shuangyong Song

Abstract: Large language models (LLMs) excel at reasoning tasks but are expensive to deploy. Thus small language models (SLMs) are fine-tuned on CoT data generated by LLMs to copy LLMs' abilities. However, these CoT data may include noisy rationales that either fail to substantiate the answers or contribute no additional information to support answer prediction, which leads SLMs to capture spurious correlat… ▽ More Large language models (LLMs) excel at reasoning tasks but are expensive to deploy. Thus small language models (SLMs) are fine-tuned on CoT data generated by LLMs to copy LLMs' abilities. However, these CoT data may include noisy rationales that either fail to substantiate the answers or contribute no additional information to support answer prediction, which leads SLMs to capture spurious correlations between questions and answers and compromise the quality of reasoning. In this work, we propose Chain-of-Thought Correctness Perception Distillation (CoPeD), which aims to improve the reasoning quality of the student model from the perspectives of task setting and data utilization. Firstly, we introduce a correctness-aware task setting that encourages the student model to predict answers based on correct rationales and revise them when they are incorrect. This setting improves the faithfulness of reasoning and allows the model to learn from its mistakes. Then, we propose a Correctness-Aware Weighted loss, which dynamically adjusts the contribution of each training instance based on the combined loss of the rationale and the answer. This strategy encourages the model to focus more on samples where the rationale offers stronger support for the correct answer. Experiments have shown that CoPeD is effective on both in-distribution (IND) and out-of-distribution (OOD) benchmark reasoning datasets. △ Less

Submitted 9 September, 2025; v1 submitted 6 September, 2025; originally announced September 2025.

Comments: PrePrint

Holistic Surgical Phase Recognition with Hierarchical Input Dependent State Space Models

Authors: Haoyang Wu, Tsun-Hsuan Wang, Mathias Lechner, Ramin Hasani, Jennifer A. Eckhoff, Paul Pak, Ozanan R. Meireles, Guy Rosman, Yutong Ban, Daniela Rus

Abstract: Surgical workflow analysis is essential in robot-assisted surgeries, yet the long duration of such procedures poses significant challenges for comprehensive video analysis. Recent approaches have predominantly relied on transformer models; however, their quadratic attention mechanism restricts efficient processing of lengthy surgical videos. In this paper, we propose a novel hierarchical input-dep… ▽ More Surgical workflow analysis is essential in robot-assisted surgeries, yet the long duration of such procedures poses significant challenges for comprehensive video analysis. Recent approaches have predominantly relied on transformer models; however, their quadratic attention mechanism restricts efficient processing of lengthy surgical videos. In this paper, we propose a novel hierarchical input-dependent state space model that leverages the linear scaling property of state space models to enable decision making on full-length videos while capturing both local and global dynamics. Our framework incorporates a temporally consistent visual feature extractor, which appends a state space model head to a visual feature extractor to propagate temporal information. The proposed model consists of two key modules: a local-aggregation state space model block that effectively captures intricate local dynamics, and a global-relation state space model block that models temporal dependencies across the entire video. The model is trained using a hybrid discrete-continuous supervision strategy, where both signals of discrete phase labels and continuous phase progresses are propagated through the network. Experiments have shown that our method outperforms the current state-of-the-art methods by a large margin (+2.8% on Cholec80, +4.3% on MICCAI2016, and +12.9% on Heichole datasets). Code will be publicly available after paper acceptance. △ Less

Submitted 26 June, 2025; originally announced June 2025.

Adaptive Batch-Wise Sample Scheduling for Direct Preference Optimization

Authors: Zixuan Huang, Yikun Ban, Lean Fu, Xiaojie Li, Zhongxiang Dai, Jianxin Li, Deqing Wang

Abstract: Direct Preference Optimization (DPO) has emerged as an effective approach for aligning large language models (LLMs) with human preferences. However, its performance is highly dependent on the quality of the underlying human preference data. To address this bottleneck, prior work has explored various data selection strategies, but these methods often overlook the impact of the evolving states of th… ▽ More Direct Preference Optimization (DPO) has emerged as an effective approach for aligning large language models (LLMs) with human preferences. However, its performance is highly dependent on the quality of the underlying human preference data. To address this bottleneck, prior work has explored various data selection strategies, but these methods often overlook the impact of the evolving states of the language model during the optimization process. In this paper, we introduce a novel problem: Sample Scheduling for DPO, which aims to dynamically and adaptively schedule training samples based on the model's evolving batch-wise states throughout preference optimization. To solve this problem, we propose SamS, an efficient and effective algorithm that adaptively selects samples in each training batch based on the LLM's learning feedback to maximize the potential generalization performance. Notably, without modifying the core DPO algorithm, simply integrating SamS significantly improves performance across tasks, with minimal additional computational overhead. This work points to a promising new direction for improving LLM alignment through batch-wise sample selection, with potential generalization to RLHF and broader supervised learning paradigms. △ Less

Submitted 9 March, 2026; v1 submitted 8 June, 2025; originally announced June 2025.

GenHOI: Generalizing Text-driven 4D Human-Object Interaction Synthesis for Unseen Objects

Authors: Shujia Li, Haiyu Zhang, Xinyuan Chen, Yaohui Wang, Yutong Ban

Abstract: While diffusion models and large-scale motion datasets have advanced text-driven human motion synthesis, extending these advances to 4D human-object interaction (HOI) remains challenging, mainly due to the limited availability of large-scale 4D HOI datasets. In our study, we introduce GenHOI, a novel two-stage framework aimed at achieving two key objectives: 1) generalization to unseen objects and… ▽ More While diffusion models and large-scale motion datasets have advanced text-driven human motion synthesis, extending these advances to 4D human-object interaction (HOI) remains challenging, mainly due to the limited availability of large-scale 4D HOI datasets. In our study, we introduce GenHOI, a novel two-stage framework aimed at achieving two key objectives: 1) generalization to unseen objects and 2) the synthesis of high-fidelity 4D HOI sequences. In the initial stage of our framework, we employ an Object-AnchorNet to reconstruct sparse 3D HOI keyframes for unseen objects, learning solely from 3D HOI datasets, thereby mitigating the dependence on large-scale 4D HOI datasets. Subsequently, we introduce a Contact-Aware Diffusion Model (ContactDM) in the second stage to seamlessly interpolate sparse 3D HOI keyframes into densely temporally coherent 4D HOI sequences. To enhance the quality of generated 4D HOI sequences, we propose a novel Contact-Aware Encoder within ContactDM to extract human-object contact patterns and a novel Contact-Aware HOI Attention to effectively integrate the contact signals into diffusion models. Experimental results show that we achieve state-of-the-art results on the publicly available OMOMO and 3D-FUTURE datasets, demonstrating strong generalization abilities to unseen objects, while enabling high-fidelity 4D HOI generation. △ Less

Submitted 18 June, 2025; originally announced June 2025.

SurgVLM: A Large Vision-Language Model and Systematic Evaluation Benchmark for Surgical Intelligence

Authors: Zhitao Zeng, Zhu Zhuo, Xiaojun Jia, Erli Zhang, Junde Wu, Jiaan Zhang, Yuxuan Wang, Chang Han Low, Jian Jiang, Zilong Zheng, Xiaochun Cao, Yutong Ban, Qi Dou, Yang Liu, Yueming Jin

Abstract: Foundation models have achieved transformative success across biomedical domains by enabling holistic understanding of multimodal data. However, their application in surgery remains underexplored. Surgical intelligence presents unique challenges - requiring surgical visual perception, temporal analysis, and reasoning. Existing general-purpose vision-language models fail to address these needs due… ▽ More Foundation models have achieved transformative success across biomedical domains by enabling holistic understanding of multimodal data. However, their application in surgery remains underexplored. Surgical intelligence presents unique challenges - requiring surgical visual perception, temporal analysis, and reasoning. Existing general-purpose vision-language models fail to address these needs due to insufficient domain-specific supervision and the lack of a large-scale high-quality surgical database. To bridge this gap, we propose SurgVLM, one of the first large vision-language foundation models for surgical intelligence, where this single universal model can tackle versatile surgical tasks. To enable this, we construct a large-scale multimodal surgical database, SurgVLM-DB, comprising over 1.81 million frames with 7.79 million conversations, spanning more than 16 surgical types and 18 anatomical structures. We unify and reorganize 23 public datasets across 10 surgical tasks, followed by standardizing labels and doing hierarchical vision-language alignment to facilitate comprehensive coverage of gradually finer-grained surgical tasks, from visual perception, temporal analysis, to high-level reasoning. Building upon this comprehensive dataset, we propose SurgVLM, which is built upon Qwen2.5-VL, and undergoes instruction tuning to 10+ surgical tasks. We further construct a surgical multimodal benchmark, SurgVLM-Bench, for method evaluation. SurgVLM-Bench consists of 6 popular and widely-used datasets in surgical domain, covering several crucial downstream tasks. Based on SurgVLM-Bench, we evaluate the performance of our SurgVLM (3 SurgVLM variants: SurgVLM-7B, SurgVLM-32B, and SurgVLM-72B), and conduct comprehensive comparisons with 14 mainstream commercial VLMs (e.g., GPT-4o, Gemini 2.0 Flash, Qwen2.5-Max). △ Less

Submitted 3 June, 2025; originally announced June 2025.

Comments: 29 pages, 5 figures

MSC Class: 68T45 ACM Class: I.2.10

Transformer Copilot: Learning from The Mistake Log in LLM Fine-tuning

Authors: Jiaru Zou, Yikun Ban, Zihao Li, Yunzhe Qi, Ruizhong Qiu, Ling Yang, Jingrui He

Abstract: Large language models are typically adapted to downstream tasks through supervised fine-tuning on domain-specific data. While standard fine-tuning focuses on minimizing generation loss to optimize model parameters, we take a deeper step by retaining and leveraging the model's own learning signals, analogous to how human learners reflect on past mistakes to improve future performance. We first intr… ▽ More Large language models are typically adapted to downstream tasks through supervised fine-tuning on domain-specific data. While standard fine-tuning focuses on minimizing generation loss to optimize model parameters, we take a deeper step by retaining and leveraging the model's own learning signals, analogous to how human learners reflect on past mistakes to improve future performance. We first introduce the concept of Mistake Log to systematically track the model's learning behavior and recurring errors throughout fine-tuning. Treating the original transformer-based model as the Pilot, we correspondingly design a Copilot model to refine the Pilot's inference performance via logits rectification. We name the overall Pilot-Copilot framework the Transformer Copilot, which introduces (i) a novel Copilot model design, (ii) a joint training paradigm where the Copilot continuously learns from the evolving Mistake Log alongside the Pilot, and (iii) a fused inference paradigm where the Copilot rectifies the Pilot's logits for enhanced generation. We provide both theoretical and empirical analyses on our new learning framework. Experiments on 12 benchmarks spanning commonsense, arithmetic, and recommendation tasks demonstrate that Transformer Copilot consistently improves performance by up to 34.5%, while introducing marginal computational overhead to Pilot models and exhibiting strong scalability and transferability. Our code is released at https://github.com/jiaruzouu/TransformerCopilot. △ Less

Submitted 13 November, 2025; v1 submitted 22 May, 2025; originally announced May 2025.

Comments: NeurIPS 2025 Spotlight

Time Reversal Symmetry for Efficient Robotic Manipulations in Deep Reinforcement Learning

Authors: Yunpeng Jiang, Jianshu Hu, Paul Weng, Yutong Ban

Abstract: Symmetry is pervasive in robotics and has been widely exploited to improve sample efficiency in deep reinforcement learning (DRL). However, existing approaches primarily focus on spatial symmetries, such as reflection, rotation, and translation, while largely neglecting temporal symmetries. To address this gap, we explore time reversal symmetry, a form of temporal symmetry commonly found in roboti… ▽ More Symmetry is pervasive in robotics and has been widely exploited to improve sample efficiency in deep reinforcement learning (DRL). However, existing approaches primarily focus on spatial symmetries, such as reflection, rotation, and translation, while largely neglecting temporal symmetries. To address this gap, we explore time reversal symmetry, a form of temporal symmetry commonly found in robotics tasks such as door opening and closing. We propose Time Reversal symmetry enhanced Deep Reinforcement Learning (TR-DRL), a framework that combines trajectory reversal augmentation and time reversal guided reward shaping to efficiently solve temporally symmetric tasks. Our method generates reversed transitions from fully reversible transitions, identified by a proposed dynamics-consistent filter, to augment the training data. For partially reversible transitions, we apply reward shaping to guide learning, according to successful trajectories from the reversed task. Extensive experiments on the Robosuite and MetaWorld benchmarks demonstrate that TR-DRL is effective in both single-task and multi-task settings, achieving higher sample efficiency and stronger final performance compared to baseline methods. △ Less

Submitted 21 October, 2025; v1 submitted 20 May, 2025; originally announced May 2025.

Comments: Accepted in NeurIPS 2025

Flash Invariant Point Attention

Authors: Andrew Liu, Axel Elaldi, Nicholas T Franklin, Nathan Russell, Gurinder S Atwal, Yih-En A Ban, Olivia Viessmann

Abstract: Invariant Point Attention (IPA) is a key algorithm for geometry-aware modeling in structural biology, central to many protein and RNA models. However, its quadratic complexity limits the input sequence length. We introduce FlashIPA, a factorized reformulation of IPA that leverages hardware-efficient FlashAttention to achieve linear scaling in GPU memory and wall-clock time with sequence length. Fl… ▽ More Invariant Point Attention (IPA) is a key algorithm for geometry-aware modeling in structural biology, central to many protein and RNA models. However, its quadratic complexity limits the input sequence length. We introduce FlashIPA, a factorized reformulation of IPA that leverages hardware-efficient FlashAttention to achieve linear scaling in GPU memory and wall-clock time with sequence length. FlashIPA matches or exceeds standard IPA performance while substantially reducing computational costs. FlashIPA extends training to previously unattainable lengths, and we demonstrate this by re-training generative models without length restrictions and generating structures of thousands of residues. FlashIPA is available at https://github.com/flagshippioneering/flash_ipa. △ Less

Submitted 16 May, 2025; originally announced May 2025.

Temporal Propagation of Asymmetric Feature Pyramid for Surgical Scene Segmentation

Authors: Cheng Yuan, Yutong Ban

Abstract: Surgical scene segmentation is crucial for robot-assisted laparoscopic surgery understanding. Current approaches face two challenges: (i) static image limitations including ambiguous local feature similarities and fine-grained structural details, and (ii) dynamic video complexities arising from rapid instrument motion and persistent visual occlusions. While existing methods mainly focus on spatial… ▽ More Surgical scene segmentation is crucial for robot-assisted laparoscopic surgery understanding. Current approaches face two challenges: (i) static image limitations including ambiguous local feature similarities and fine-grained structural details, and (ii) dynamic video complexities arising from rapid instrument motion and persistent visual occlusions. While existing methods mainly focus on spatial feature extraction, they fundamentally overlook temporal dependencies in surgical video streams. To address this, we present temporal asymmetric feature propagation network, a bidirectional attention architecture enabling cross-frame feature propagation. The proposed method contains a temporal query propagator that integrates multi-directional consistency constraints to enhance frame-specific feature representation, and an aggregated asymmetric feature pyramid module that preserves discriminative features for anatomical structures and surgical instruments. Our framework uniquely enables both temporal guidance and contextual reasoning for surgical scene understanding. Comprehensive evaluations on two public benchmarks show the proposed method outperforms the current SOTA methods by a large margin, with +16.4\% mIoU on EndoVis2018 and +3.3\% mAP on Endoscapes2023. The code will be publicly available after paper acceptance. △ Less

Submitted 17 April, 2025; originally announced April 2025.

Tracking-Aware Deformation Field Estimation for Non-rigid 3D Reconstruction in Robotic Surgeries

Authors: Zeqing Wang, Han Fang, Yihong Xu, Yutong Ban

Abstract: Minimally invasive procedures have been advanced rapidly by the robotic laparoscopic surgery. The latter greatly assists surgeons in sophisticated and precise operations with reduced invasiveness. Nevertheless, it is still safety critical to be aware of even the least tissue deformation during instrument-tissue interactions, especially in 3D space. To address this, recent works rely on NeRF to ren… ▽ More Minimally invasive procedures have been advanced rapidly by the robotic laparoscopic surgery. The latter greatly assists surgeons in sophisticated and precise operations with reduced invasiveness. Nevertheless, it is still safety critical to be aware of even the least tissue deformation during instrument-tissue interactions, especially in 3D space. To address this, recent works rely on NeRF to render 2D videos from different perspectives and eliminate occlusions. However, most of the methods fail to predict the accurate 3D shapes and associated deformation estimates robustly. Differently, we propose Tracking-Aware Deformation Field (TADF), a novel framework which reconstructs the 3D mesh along with the 3D tissue deformation simultaneously. It first tracks the key points of soft tissue by a foundation vision model, providing an accurate 2D deformation field. Then, the 2D deformation field is smoothly incorporated with a neural implicit reconstruction network to obtain tissue deformation in the 3D space. Finally, we experimentally demonstrate that the proposed method provides more accurate deformation estimation compared with other 3D neural reconstruction methods in two public datasets. △ Less

Submitted 4 March, 2025; originally announced March 2025.