arXiv subjects

Enyu Zhou

Publications and source records attributed to Enyu Zhou.

At least 19 recordsLinked to original sources

arXiv2026

ReDraft, Don't Just Distill: Reference-Driven Revision for Continual VLLM Post-Training

Continual post-training of large multimodal models should add new capabilities while preserving those from pre-training, and the two goals pull in opposite directions. SFT gives explicit target supervision that learns a task from near-zero accuracy, but its off-policy targets move the model far enough to cause forgetting; on-policy methods such as RLVR and self-distillation preserve policy proximity yet supply little signal when the policy cannot yet solve the task. We introduce ReDraft (Reference-Driven Revision and Fine-Tuning), which obtains both from the model's own failures: using an expert response only as a reference, it has the model revise its own incorrect rollout, keeps the revision only if a verifier accepts it, and fine-tunes on what survives. Each retained target is therefore explicit, yet still close to the current policy. Across Counting, Clock Reading, and Jigsaw on Qwen2.5-VL-3B/7B, two of them with near zero accuracy, ReDraft gains 56.9 points on the target task against SFT's 52.9 while cutting prior-task loss from 16.6 to 1.5 points (11.3x less forgetting), and improves on OPSD along both axes (19.3 gain, 6.2 loss). Data- and parameter-space analyses match the design: revised targets are more probable under the base model, and the updates they induce stay compact and follow SFT's direction more closely than OPSD's. Repairing the model's own output, rather than replacing it with an expert's, is what lets one objective do both.

cs.AI↗

arXiv2026

Sci-MMR: Benchmarking Multi-Step Evidence-Grounded Scientific Reasoning in Multimodal Agents

Autonomous research agents are increasingly expected to search the literature, analyze experimental evidence, and generate scientific hypotheses. These capabilities require multi-step evidence grounded reasoning that progressively acquires, integrates, and verifies evidence before reaching a conclusion. Existing multimodal benchmarks, however, largely evaluate final-answer accuracy, leaving open whether predictions are actually supported by traceable scientific evidence. We introduce Sci-MMR, a benchmark for multi-step evidence-grounded scientific reasoning built on structured argument graphs linking scientific claims, citation-grounded knowledge, visual evidence, and supporting regions. Sci-MMR comprises 235 multi-hop reasoning tasks spanning four scientific disciplines, with an average of nine figure panels per task. Evaluating eight frontier multimodal models, we find that answer accuracy consistently exceeds complete-evidence recovery rate by more than 20%, revealing a substantial gap that answer-only evaluation is structurally unable to capture. Through controlled interventions, we identify two fundamental bottlenecks. First, evidence acquisition: models struggle to extract complete structured evidence from scientific figures, accounting for 57.2% of failures. While cropping tools yield modest gains (+4.5 points), providing gold evidence improves accuracy by up to 37.0 points, indicating difficulty in assembling complete multi-region evidence. Second, evidence integration: models struggle to translate available evidence into correct conclusions, accounting for 31.8% of failures, while even with gold evidence the strongest model achieves only 69.1% accuracy on the hardest tasks. These findings indicate that current answer-centric benchmarks substantially overestimate the evidence-grounded reasoning capabilities of multimodal research agents

cs.AI↗

arXiv2026

JFTA-Bench: Evaluate LLM's Ability of Tracking and Analyzing Malfunctions Using Fault Trees

In the maintenance of complex systems, fault trees are used to locate problems and provide targeted solutions. To enable fault trees stored as images to be directly processed by large language models, which can assist in tracking and analyzing malfunctions, we propose a novel textual representation of fault trees. Building on it, we construct a benchmark for multi-turn dialogue systems that emphasizes robust interaction in complex environments, evaluating a model's ability to assist in malfunction localization, which contains $3130$ entries and $40.75$ turns per entry on average. We train an end-to-end model to generate vague information to reflect user behavior and introduce long-range rollback and recovery procedures to simulate user error scenarios, enabling assessment of a model's integrated capabilities in task tracking and error recovery, and Gemini 2.5 pro archives the best performance.

cs.AI↗

arXiv2026

ContextWeave: A Real-World Workflow Benchmark

Memory is essential as language agents move from isolated tasks to long-horizon, stateful workflows, yet existing evaluations often reduce it to retrieval or question answering. We introduce ContextWeave, a longitudinal benchmark that evaluates whether recalled experience improves downstream agent performance in realistic office-work streams. ContextWeave reconstructs privacy-preserved, multi-month workflows of 14 participants into 1,005 executable tasks, including 568 core evaluation tasks, with instructions, containerized environments, trajectories, and task-specific rubrics. It measures workspace quality and alignment with participant-specific preferences, complemented by diagnostics of relevance, continuity, solvability, and robustness to misleading recall. Across six memory components under a fixed model, the strongest configuration raises Workspace Score from 68.08 to 78.20 and Preference Score from 41.50 to 70.60. With a fixed memory component, recall improves both outcomes for all five tested base models, although gains vary substantially. Our analysis shows that actionable, experience-rich memory supports workflow continuation and reduces redundant exploration more effectively than compact summaries, while it can also be more susceptible to misleading recall. These findings motivate memory systems that optimize not only retrieval relevance but also reliable use during execution.

cs.AI↗

arXiv2026

VRPO: Rethinking Value Modeling for Robust RL under Noisy Supervision in LLM Post-Training

Reinforcement Learning (RL) in real-world environments often suffers from ambiguous or incomplete reward supervision, which undermines policy stability and generalization. Such noise may cause models to ignore key information or even collapse in advantage estimation. We find that a strong value model is essential for absorbing unstable signals and producing reliable advantages, offering denser and more robust supervision than the reward model. To better optimize noisy supervision, we propose VRPO, a framework that enhances value modeling for robust RL in LLM post-training. VRPO integrates (1) auxiliary losses guided by entropy and perplexity from a frozen language model, and (2) a variational information bottleneck, enabling the value model to filter noise and capture key words. This design allows the value model to correct noise rewards and generate more reliable advantage estimates, transforming it from a passive predictor into an active noise regulator. Experiments on multi-turn dialogue, math reasoning, and science QA with both rule-based and model-based rewards show that VRPO consistently outperforms baselines such as PPO and GRPO. Our work highlight the central role of the value model in Robust RL and provide a principled and practical approach to policy optimization under noisy supervision.

cs.LG↗

arXiv2026

MM-Doc-R1: Training Agents for Long Document Visual Question Answering through Multi-turn Reinforcement Learning

Conventional Retrieval-Augmented Generation (RAG) systems often struggle with complex multi-hop queries over long documents due to their single-pass retrieval. We introduce MM-Doc-R1, a novel framework that employs an agentic, vision-aware workflow to address long document visual question answering through iterative information discovery and synthesis. To incentivize the information seeking capabilities of our agents, we propose Similarity-based Policy Optimization (SPO), addressing baseline estimation bias in existing multi-turn reinforcement learning (RL) algorithms like GRPO. Our core insight is that in multi-turn RL, the more semantically similar two trajectories are, the more accurate their shared baseline estimation becomes. Leveraging this, SPO calculates a more precise baseline by similarity-weighted averaging of rewards across multiple trajectories, unlike GRPO which inappropriately applies the initial state's baseline to all intermediate states. This provides a more stable and accurate learning signal for our agents, leading to superior training performance that surpasses GRPO. Our experiments on the MMLongbench-Doc benchmark show that MM-Doc-R1 outperforms previous baselines by 10.4%. Furthermore, SPO demonstrates superior performance over GRPO, boosting results by 5.0% with Qwen3-8B and 6.1% with Qwen3-4B. These results highlight the effectiveness of our integrated framework and novel training algorithm in advancing the state-of-the-art for complex, long-document visual question answering.

cs.CL↗

arXiv2026

Why Reinforcement Fine-Tuning Enables MLLMs Preserve Prior Knowledge Better: A Data Perspective

Post-training algorithms such as Supervised Fine-Tuning (SFT) and Reinforcement Fine-Tuning (RFT) are widely used to adapt (multimodal) large language models to downstream tasks. While effective at task adaptation, their impact on retaining prior knowledge remains unclear. In this paper, we introduce jigsaw puzzles as a novel task absent from existing pretraining corpora and systematically study the behavior of SFT and RFT on the open-source Qwen2.5-VL series. Our experiments reveal a sharp trade-off: SFT enables rapid task acquisition but leads to catastrophic forgetting, whereas RFT learns more slowly but better maintains prior knowledge. We study this phenomenon through learning dynamics by examining both the magnitude and direction of how training data influence prior knowledge. Our analysis shows that RFT mainly reinforces correct samples naturally aligned with the base model's probability landscape, leading to weaker interference with prior knowledge. Moreover, training on RFT-simulated rollouts, which exert a smaller magnitude of influence and are better aligned in direction to prior knowledge, allows SFT to preserve prior knowledge better while rapidly learning new tasks. We further validate our framework on Qwen2.5 post-training in math and scientific QA, observing consistent forgetting and learning-dynamics trends. These findings suggest that the distribution of post-training data, rather than algorithmic differences alone, plays a central role in forgetting, and highlight RFT as a promising ingredient for stable continual post-training.

cs.CL↗

arXiv2026

Steering LLMs via Scalable Interactive Oversight

As Large Language Models increasingly automate complex, long-horizon tasks such as \emph{vibe coding}, a supervision gap has emerged. While models excel at execution, users often struggle to guide them effectively due to insufficient domain expertise, the difficulty of articulating precise intent, and the inability to reliably validate complex outputs. It presents a critical challenge in scalable oversight: enabling humans to responsibly steer AI systems on tasks that surpass their own ability to specify or verify. To tackle this, we propose Scalable Interactive Oversight, a framework that decomposes complex intent into a recursive tree of manageable decisions to amplify human supervision. Rather than relying on open-ended prompting, our system elicits low-burden feedback at each node and recursively aggregates these signals into precise global guidance. Validated in web development task, our framework enables non-experts to produce expert-level Product Requirement Documents, achieving a 54\% improvement in alignment. Crucially, we demonstrate that this framework can be optimized via Reinforcement Learning using only online user feedback, offering a practical pathway for maintaining human control as AI scales.

cs.AI↗

arXiv2026

Pre-Trained Policy Discriminators are General Reward Models

We offer a novel perspective on reward modeling by formulating it as a policy discriminator, which quantifies the difference between two policies to generate a reward signal, guiding the training policy towards a target policy with desired behaviors. Based on this conceptual insight, we propose a scalable pre-training method named Policy Discriminative Learning (POLAR), which trains a reward model (RM) to discern identical policies and discriminate different ones. Unlike traditional reward modeling methods relying on absolute preferences, POLAR captures the relative difference between one policy and an arbitrary target policy, which is a scalable, high-level optimization objective suitable for modeling generic ranking relationships. Leveraging the POLAR pre-training paradigm, we present a series of RMs with parameter scales from 1.8B to 7B. Empirical results show that POLAR substantially outperforms traditional non-pre-trained methods, significantly enhancing RM performance. For instance, POLAR-7B could improve preference accuracy from 54.8% to 81.0% on STEM tasks and from 57.9% to 85.5% on creative writing tasks compared to SOTA baselines. POLAR also shows robust generalization capabilities in RLHF using Reinforcement Fine-tuning (RFT), providing reliable reward signals and markedly enhancing policy performance--improving LLaMa3.1-8B from an average of 47.36% to 56.33% and Qwen2.5-32B from 64.49% to 70.47% on 20 benchmarks. Moreover, scaling experiments reveal a clear power-law relationship between computation and performance, supported by linear correlation coefficients approaching 0.99. The impressive performance, strong generalization, and scaling properties suggest that POLAR is a promising direction for developing general and strong reward models.

cs.CL↗

arXiv2026

FRoM-W1: Towards General Humanoid Whole-Body Control with Language Instructions

Humanoid robots are capable of performing various actions such as greeting, dancing and even backflipping. However, these motions are often hard-coded or specifically trained, which limits their versatility. In this work, we present FRoM-W1, an open-source framework designed to achieve general humanoid whole-body motion control using natural language. To universally understand natural language and generate corresponding motions, as well as enable various humanoid robots to stably execute these motions in the physical world under gravity, FRoM-W1 operates in two stages: (a) H-GPT: utilizing massive human data, a large-scale language-driven human whole-body motion generation model is trained to generate diverse natural behaviors. We further leverage the Chain-of-Thought technique to improve the model's generalization in instruction understanding. (b) H-ACT: After retargeting generated human whole-body motions into robot-specific actions, a motion controller that is pretrained and further fine-tuned through reinforcement learning in physical simulation enables humanoid robots to accurately and stably perform corresponding actions. It is then deployed on real robots via a modular simulation-to-reality module. We extensively evaluate FRoM-W1 on Unitree H1 and G1 robots. Results demonstrate superior performance on the HumanML3D-X benchmark for human whole-body motion generation, and our introduced reinforcement learning fine-tuning consistently improves both motion tracking accuracy and task success rates of these humanoid robots. We open-source the entire FRoM-W1 framework and hope it will advance the development of humanoid intelligence.

cs.RO↗

arXiv2025

Nex-N1: Agentic Models Trained via a Unified Ecosystem for Large-Scale Environment Construction

The evolution of Large Language Models (LLMs) from passive responders to autonomous agents necessitates a fundamental shift in learning paradigms -- from static imitation to incentive-driven decision making. However, this transition is significantly impeded by the lack of scalable infrastructure capable of constructing high-quality interaction signals for effective policy learning. To address this, we introduce a comprehensive method designed to systematically scale the diversity and complexity of interactive environments. Our method realizes this scaling by addressing three orthogonal dimensions: (1) Complexity: NexAU, a flexible agent framework that supports building complex agent hierarchies via simple configurations; (2) Diversity: NexA4A automatically generates diverse agent hierarchies from natural language to cover infinite domains; and (3) Fidelity: NexGAP bridges the simulation-reality gap by integrating dynamic real-world environment for grounded trajectories synthesis. We train Nex-N1 upon the diverse and complex interactive environments established by our infrastructure. Empirical results on benchmarks such as SWE-bench and tau2 demonstrate that Nex-N1 consistently outperforms SOTA open-source models and achieves competitive performance against frontier proprietary models on complex agentic tasks. We open-source the Nex ecosystem and model weights to facilitate further research.

cs.CL↗

arXiv2025

BAPO: Stabilizing Off-Policy Reinforcement Learning for LLMs via Balanced Policy Optimization with Adaptive Clipping

Reinforcement learning (RL) has recently become the core paradigm for aligning and strengthening large language models (LLMs). Yet, applying RL in off-policy settings--where stale data from past policies are used for training--improves sample efficiency, but remains challenging: policy entropy declines sharply, optimization often becomes unstable and may even collapse. Through theoretical and empirical analysis, we identify two key insights: (i) an imbalance in optimization, where negative-advantage samples dominate the policy gradient, suppressing useful behaviors and risking gradient explosions; and (ii) the derived Entropy-Clip Rule, which reveals that the fixed clipping mechanism in PPO-like objectives systematically blocks entropy-increasing updates, thereby driving the policy toward over-exploitation at the expense of exploration. Building on these insights, we propose BAlanced Policy Optimization with Adaptive Clipping (BAPO), a simple yet effective method that dynamically adjusts clipping bounds to adaptively re-balance positive and negative contributions, preserve entropy, and stabilize RL optimization. Across diverse off-policy scenarios--including sample replay and partial rollout--BAPO achieves fast, stable, and data-efficient training. On AIME 2024 and AIME 2025 benchmarks, our 7B BAPO model surpasses open-source counterparts such as SkyWork-OR1-7B, while our 32B BAPO model not only achieves state-of-the-art results among models of the same scale but also outperforms leading proprietary systems like o3-mini and Gemini-2.5-Flash-Thinking.

cs.LG↗

arXiv2025

CARD: A Cache-Assisted Parallel Speculative Decoding Framework via Query-and-Correct Paradigm for Accelerating LLM Inference

Speculative decoding (SD), where a draft model provides multiple candidate tokens for the target model to verify in parallel, has demonstrated significant potential for accelerating LLM inference. Yet, existing SD approaches adhere to a strict draft-then-verify paradigm, enforcing a sequential process that hampers performance and constrains the draft model's capacity. Moreover, rejecting a token in the candidate sequence invalidates all subsequent tokens, leading to wasted computation during drafting. To overcome these limitations, we propose a cache-assisted parallel speculative decoding framework called CARD, which employs a novel query-and-correct paradigm. Our approach decouples drafting from verification: the draft model populates a shared cache with candidate tokens, while the target model concurrently refines the draft's trajectory. This enables inference at near-draft-speed, effectively leveraging the draft model's efficiency without additional fine-tuning. Experimental results show that CARD significantly outperforms existing state-of-the-art methods, achieving up to a 4.83x acceleration over vanilla autoregressive decoding, with no fine-tuning required for either models.

cs.LG↗

arXiv2025

RMB: Comprehensively Benchmarking Reward Models in LLM Alignment

Reward models (RMs) guide the alignment of large language models (LLMs), steering them toward behaviors preferred by humans. Evaluating RMs is the key to better aligning LLMs. However, the current evaluation of RMs may not directly correspond to their alignment performance due to the limited distribution of evaluation data and evaluation methods that are not closely related to alignment objectives. To address these limitations, we propose RMB, a comprehensive RM benchmark that covers over 49 real-world scenarios and includes both pairwise and Best-of-N (BoN) evaluations to better reflect the effectiveness of RMs in guiding alignment optimization. We demonstrate a positive correlation between our benchmark and the downstream alignment task performance. Based on our benchmark, we conduct extensive analysis on the state-of-the-art RMs, revealing their generalization defects that were not discovered by previous benchmarks, and highlighting the potential of generative RMs. Furthermore, we delve into open questions in reward models, specifically examining the effectiveness of majority voting for the evaluation of reward models and analyzing the impact factors of generative RMs, including the influence of evaluation criteria and instructing methods. Our evaluation code and datasets are available at https://github.com/Zhou-Zoey/RMB-Reward-Model-Benchmark.

cs.CL↗

arXiv2024

SafeAligner: Safety Alignment against Jailbreak Attacks via Response Disparity Guidance

As the development of large language models (LLMs) rapidly advances, securing these models effectively without compromising their utility has become a pivotal area of research. However, current defense strategies against jailbreak attacks (i.e., efforts to bypass security protocols) often suffer from limited adaptability, restricted general capability, and high cost. To address these challenges, we introduce SafeAligner, a methodology implemented at the decoding stage to fortify defenses against jailbreak attacks. We begin by developing two specialized models: the Sentinel Model, which is trained to foster safety, and the Intruder Model, designed to generate riskier responses. SafeAligner leverages the disparity in security levels between the responses from these models to differentiate between harmful and beneficial tokens, effectively guiding the safety alignment by altering the output token distribution of the target model. Extensive experiments show that SafeAligner can increase the likelihood of beneficial tokens, while reducing the occurrence of harmful ones, thereby ensuring secure alignment with minimal loss to generality.

cs.CR↗

arXiv2024

Aligning Large Language Models from Self-Reference AI Feedback with one General Principle

In aligning large language models (LLMs), utilizing feedback from existing advanced AI rather than humans is an important method to scale supervisory signals. However, it is highly challenging for AI to understand human intentions and societal values, and provide accurate preference feedback based on these. Current AI feedback methods rely on powerful LLMs, carefully designed specific principles to describe human intentions, and are easily influenced by position bias. To address these issues, we propose a self-reference-based AI feedback framework that enables a 13B Llama2-Chat to provide high-quality feedback under simple and general principles such as ``best for humanity``. Specifically, we allow the AI to first respond to the user's instructions, then generate criticism of other answers based on its own response as a reference, and finally determine which answer better fits human preferences according to the criticism. Additionally, we use a self-consistency method to further reduce the impact of position bias, and employ semantic perplexity to calculate the preference strength differences between different answers. Experimental results show that our method enables 13B and 70B Llama2-Chat annotators to provide high-quality preference feedback, and the policy models trained based on these preference data achieve significant advantages in benchmark datasets through reinforcement learning.

cs.CL↗

arXiv2024

MetaRM: Shifted Distributions Alignment via Meta-Learning

The success of Reinforcement Learning from Human Feedback (RLHF) in language model alignment is critically dependent on the capability of the reward model (RM). However, as the training process progresses, the output distribution of the policy model shifts, leading to the RM's reduced ability to distinguish between responses. This issue is further compounded when the RM, trained on a specific data distribution, struggles to generalize to examples outside of that distribution. These two issues can be united as a challenge posed by the shifted distribution of the environment. To surmount this challenge, we introduce MetaRM, a method leveraging meta-learning to align the RM with the shifted environment distribution. MetaRM is designed to train the RM by minimizing data loss, particularly for data that can improve the differentiation ability to examples of the shifted target distribution. Extensive experiments demonstrate that MetaRM significantly improves the RM's distinguishing ability in iterative RLHF optimization, and also provides the capacity to identify subtle differences in out-of-distribution samples.

cs.LG↗

arXiv2024

LoRAMoE: Alleviate World Knowledge Forgetting in Large Language Models via MoE-Style Plugin

Supervised fine-tuning (SFT) is a crucial step for large language models (LLMs), enabling them to align with human instructions and enhance their capabilities in downstream tasks. Increasing instruction data substantially is a direct solution to align the model with a broader range of downstream tasks or notably improve its performance on a specific task. However, we find that large-scale increases in instruction data can damage the world knowledge previously stored in LLMs. To address this challenge, we propose LoRAMoE, a novelty framework that introduces several low-rank adapters (LoRA) and integrates them by using a router network, like a plugin version of Mixture of Experts (MoE). It freezes the backbone model and forces a portion of LoRAs to focus on leveraging world knowledge to solve downstream tasks, to alleviate world knowledge-edge forgetting. Experimental results show that, as the instruction data increases, LoRAMoE can significantly improve the ability to process downstream tasks, while maintaining the world knowledge stored in the LLM.

cs.CL↗