arXiv subjects

Zhifang Sui

Publications and source records attributed to Zhifang Sui.

At least 19 recordsLinked to original sources

arXiv2026

MOPD: Multi-Teacher On-Policy Distillation for Capability Integration in LLM Post-Training

Modern large language models (LLMs) rely on reinforcement learning during post-training to push specific capabilities, yet integrating multiple capabilities into one model remains hard. Existing methods, such as Off-Policy Finetune and Mix-RL, are either inefficient or lose performance. In this work, we propose Multi-teacher On-Policy Distillation (MOPD), a post-training paradigm for combining the capabilities of multiple domain RL teachers: we first run per-domain specialised RL to obtain a set of domain teachers, then distill these teachers into the student on its own rollouts. This eliminates exposure bias and provides a dense optimization signal. On Qwen3-30B-A3B, MOPD outperforms Mix-RL, Cascade RL, Off-Policy Finetune, and Param-Merge baselines, inheriting nearly all of each teacher's capability. MOPD also enables parallel, independent development of domain teachers, removing the cross-domain coupling typical of multi-domain post-training. MOPD has been deployed in the post-training of MiMo-V2-Flash, an industrial-scale frontier model, demonstrating its practical value for capability integration in frontier-scale LLMs.

cs.CL↗

arXiv2026

HauntAttack: When Attack Follows Reasoning as a Shadow

Emerging Large Reasoning Models (LRMs) consistently excel in mathematical and reasoning tasks, showcasing remarkable capabilities. However, the enhancement of reasoning abilities and the exposure of internal reasoning processes introduce new safety vulnerabilities. A critical question arises: when reasoning becomes intertwined with harmfulness, will LRMs become more vulnerable to jailbreaks in reasoning mode? To investigate this, we introduce HauntAttack, a novel and general-purpose black-box adversarial attack framework that systematically embeds harmful instructions into reasoning questions. Specifically, we modify key reasoning conditions in existing questions with harmful instructions, thereby constructing a reasoning pathway that guides the model step by step toward unsafe outputs. We evaluate HauntAttack on 11 LRMs and observe an average attack success rate of over 70\%, achieving up to 13 percentage points of absolute improvement over the strongest prior baseline. Our further analysis reveals that even advanced safety-aligned models remain highly susceptible to reasoning-based attacks, offering insights into the urgent challenge of balancing reasoning capability and safety in future model development.

cs.CR↗

arXiv2026

SenseJudge: Human-Centric Preference-Driven Judgment Framework

Large Language Models (LLMs) as judges across various scenarios such as assessing model responses is becoming an increasingly accepted paradigm. However, existing judgment approaches often rely on trained judgers using fixed preference data, which tend to overlook diverse user preferences and struggle to adapt to real-world human-AI dialogue scenarios. To address these limitations, we propose SenseJudge, a customizable judgment framework driven by human preferences and SenseBench, a diverse and challenging instruction-following benchmark derived from real-world multi-turn interactions. We applied the automatic judgment framework and benchmark to two tasks: (1) LLMs as personalized judges, and (2) model ranking. We conducted extensive experiments, and the results demonstrate that the SenseJudge framework surpasses other judgment methods and models in the LLMs-as-personalized-judges task and achieves model ranking that aligns with real human sense. Additionally, we conducted analyses on position bias and consistency, alongside ablation studies, which affirmed the robustness of SenseJudge.

cs.CL↗

arXiv2026

From Refuse to Richness: Rubric Rewards for Long-Form Hallucination Reinforcement Learning

Rewards that penalize unsupported claims can improve grounding in long-form generation, but they can also teach models to answer less. We study this refusal-to-richness trade-off in long-form hallucination RL. Instead of using global richness proxies such as length, claim count, detail, or pairwise relevance, we represent each question with a key-point rubric that specifies the required and optional information a useful answer should cover. These rubrics define coverage directly and are used both for evaluation and as reward signals. Across grounding-only, proxy-based, rubric-only, and combined rewards, we find a stable trade-off: strict grounding rewards improve support but suppress coverage, while unconstrained rubric rewards improve coverage but weaken grounding. A soft combination of grounding, rubric coverage, and relevance gives the best balance in our experiments, improving in-distribution support while transferring better to out-of-distribution checklist tasks than either grounding-only or rubric-only rewards.

cs.CL↗

arXiv2026

Claw-Eval: Towards Trustworthy Evaluation of Autonomous Agents

Large language models are increasingly deployed as autonomous agents for multi-step workflows in real-world software environments. However, existing agent benchmarks are limited by trajectory-opaque grading, underspecified safety and robustness evaluation, and narrow coverage of modalities and interaction paradigms. We introduce Claw-Eval, an end-to-end evaluation suite addressing these gaps with 300 human-verified tasks spanning 9 categories across three groups: general service orchestration, multimodal perception and interaction, and multi-turn professional dialogue. To enable trajectory-aware grading, each run is recorded through three independent evidence channels: execution traces, audit logs, and environment snapshots, yielding 2,159 fine-grained rubric items. The scoring protocol evaluates Completion, Safety, and Robustness, with Average Score, Pass@k, and Pass^k across three trials to distinguish genuine capability from lucky outcomes. Experiments on 14 frontier models show that: (1) Trajectory-opaque evaluation is systematically unreliable, missing 44% of safety violations and 13% of robustness failures detected by our framework. (2) Capability does not imply consistency, with Pass@3 remaining stable under error injection while Pass^3 dropping by up to 24 percentage points. (3) Agent capability is strongly multi-dimensional, with model rankings varying across task groups and metrics, indicating that our heterogeneous evaluation coverage is essential. Claw-Eval highlights directions for developing agents that are not only capable but reliably deployable.

cs.AI↗

arXiv2026

SelfBudgeter: Adaptive Token Allocation for Efficient LLM Reasoning

Recently, large reasoning models demonstrate exceptional performance on various tasks. However, reasoning models always consume excessive tokens even for simple queries, leading to resource waste and prolonged user latency. To address this challenge, we propose SelfBudgeter - a self-adaptive reasoning strategy for efficient and controllable reasoning. Specifically, we first train the model to self-estimate the required reasoning budget based on the query. We then introduce budget-guided GRPO for reinforcement learning, which effectively maintains accuracy while reducing output length. Experimental results demonstrate that SelfBudgeter dynamically allocates budgets according to problem complexity, achieving an average response length compression of 61% on math reasoning tasks while maintaining accuracy. Furthermore, SelfBudgeter allows users to see how long generation will take and decide whether to continue or stop. Additionally, users can directly control the reasoning length by setting token budgets upfront.

cs.AI↗

arXiv2026

HistLens: Mapping Idea Change across Concepts and Corpora

Language change both reflects and shapes social processes, and the semantic evolution of foundational concepts provides a measurable trace of historical and social transformation. Despite recent advances in diachronic semantics and discourse analysis, existing computational approaches often (i) concentrate on a single concept or a single corpus, making findings difficult to compare across heterogeneous sources, and (ii) remain confined to surface lexical evidence, offering insufficient computational and interpretive granularity when concepts are expressed implicitly. We propose HistLens, a unified, SAE-based framework for multi-concept, multi-corpus conceptual-history analysis. The framework decomposes concept representations into interpretable features and tracks their activation dynamics over time and across sources, yielding comparable conceptual trajectories within a shared coordinate system. Experiments on long-span press corpora show that HistLens supports cross-concept, cross-corpus computation of patterns of idea evolution and enables implicit concept computation. By bridging conceptual modeling with interpretive needs, HistLens broadens the analytical perspectives and methodological repertoire available to social science and the humanities for diachronic text analysis.

cs.CL↗

arXiv2026

GroundingME: Exposing the Visual Grounding Gap in MLLMs through Multi-Dimensional Evaluation

Visual grounding, localizing objects from natural language descriptions, represents a critical bridge between language and vision understanding. While multimodal large language models (MLLMs) achieve impressive scores on existing benchmarks, a fundamental question remains: can MLLMs truly visually ground with human-like sophistication, or are they merely pattern-matching on simplified datasets? Current benchmarks fail to capture real-world complexity where humans effortlessly navigate intricate references and recognize when grounding is impossible. To rigorously assess MLLMs' true capabilities, we introduce GroundingME, a benchmark that systematically challenges models across four critical dimensions: (1) Discriminative: distinguishing highly similar objects, (2) Spatial: understanding complex relational descriptions, (3) Limited: handling occlusions or tiny objects, and (4) Rejection: recognizing ungroundable queries. Through careful curation combining automated generation with human verification, we create 1,005 challenging examples mirroring real-world complexity. Evaluating 25 state-of-the-art MLLMs reveals a profound capability gap: the best model achieves only 45.1% accuracy, while most score 0% on rejection tasks. We explore two strategies for improvements: (1) test-time scaling selects optimal response by thinking trajectory to improve overall performance by up to 4.5%, and (2) data-mixture training boosts rejection accuracy from 0% to 27.9%. GroundingME thus serves as both a diagnostic tool revealing current limitations in MLLMs and a roadmap toward human-level visual grounding. Project page: https://groundingme.github.io

cs.CV↗

arXiv2026

Taking a Deep Breath: Enhancing Language Modeling of Large Language Models with Sentinel Tokens

Large language models (LLMs) have shown promising efficacy across various tasks, becoming powerful tools in numerous aspects of human life. However, Transformer-based LLMs suffer a performance degradation when modeling long-term contexts due to they discard some information to reduce computational overhead. In this work, we propose a simple yet effective method to enable LLMs to take a deep breath, encouraging them to summarize information contained within discrete text chunks. Specifically, we segment the text into multiple chunks and insert special token at the end of each chunk. We then modify the attention mask to integrate the chunk's information into the corresponding token. This facilitates LLMs to interpret information not only from historical individual tokens but also from the token, aggregating the chunk's semantic information. Experiments on language modeling and out-of-domain downstream tasks validate the superiority of our approach.

cs.CL↗

arXiv2026

Sparse-BitNet: 1.58-bit LLMs are Naturally Friendly to Semi-Structured Sparsity

Semi-structured N:M sparsity and low-bit quantization (e.g., 1.58-bit BitNet) are two promising approaches for improving the efficiency of large language models (LLMs), yet they have largely been studied in isolation. In this work, we investigate their interaction and show that 1.58-bit BitNet is naturally more compatible with N:M sparsity than full-precision models. To study this effect, we propose Sparse-BitNet, a unified framework that jointly applies 1.58-bit quantization and dynamic N:M sparsification while ensuring stable training for the first time. Across multiple model scales and training regimes (sparse pretraining and dense-to-sparse schedules), 1.58-bit BitNet consistently exhibits smaller performance degradation than full-precision baselines at the same sparsity levels and can tolerate higher structured sparsity before accuracy collapse. Moreover, using our custom sparse tensor core, Sparse-BitNet achieves substantial speedups in both training and inference, reaching up to 1.30X. These results highlight that combining extremely low-bit quantization with semi-structured N:M sparsity is a promising direction for efficient LLMs. Code available at https://github.com/AAzdi/Sparse-BitNet

cs.CL↗

arXiv2026

Towards Better RL Training Data Utilization via Second-Order Rollout

Reinforcement Learning (RL) has empowered Large Language Models (LLMs) with strong reasoning capabilities, but vanilla RL mainly focuses on generation capability improvement by training with only first-order rollout (generating multiple responses for a question), and we argue that this approach fails to fully exploit the potential of training data because of the neglect of critique capability training. To tackle this problem, we further introduce the concept of second-order rollout (generating multiple critiques for a response) and propose a unified framework for jointly training generation and critique capabilities. Extensive experiments across various models and datasets demonstrate that our approach can utilize training data more effectively than vanilla RL and achieve better performance under the same training data. Additionally, we uncover several insightful findings regarding second-order rollout and critique training, such as the importance of label balance in critique training and the noise problem of outcome-based rewards, which can be mitigated through sampling techniques. Our work offers a preliminary exploration of dynamic data augmentation and joint generation-critique training in RL, providing meaningful inspiration for the further advancement of RL training

cs.CL↗

arXiv2026

CoLT: Reasoning with Chain of Latent Tool Calls

Chain-of-Thought (CoT) is a critical technique in enhancing the reasoning ability of Large Language Models (LLMs), and latent reasoning methods have been proposed to accelerate the inefficient token-level reasoning chain. We notice that existing latent reasoning methods generally require model structure augmentation and exhaustive training, limiting their broader applicability. In this paper, we propose CoLT, a novel framework that implements latent reasoning as ``tool calls''. Instead of reasoning entirely in the latent space, CoLT generates seed tokens that contain information of a reasoning step. When a latent tool call is triggered, a smaller external model will take the hidden states of seed tokens as its input, and unpack the seed tokens back to a full reasoning step. In this way, we can ensure that the main model reasons in the explicit token space, preserving its ability while improving efficiency. Experimental results on four mathematical datasets demonstrate that CoLT achieves higher accuracy and shorter reasoning length than baseline latent models, and is compatible with reinforcement learning algorithms and different decoder structures.

cs.CL↗

arXiv2026

Decoding in Geometry: Alleviating Embedding-Space Crowding for Complex Reasoning

Sampling-based decoding underlies complex reasoning in large language models (LLMs), where decoding strategies critically shape model behavior. Temperature- and truncation-based methods reshape the next-token distribution through global probability reweighting or thresholding to balance the quality-diversity tradeoff. However, they operate solely on token probabilities, ignoring fine-grained relationships among tokens in the embedding space. We uncover a novel phenomenon, embedding-space crowding, where the next-token distribution concentrates its probability mass on geometrically close tokens in the embedding space. We quantify crowding at multiple granularities and find a statistical association with reasoning success in mathematical problem solving. Motivated by this finding, we propose CraEG, a plug-and-play sampling method that mitigates crowding through geometry-guided reweighting. CraEG is training-free, single-pass, and compatible with standard sampling strategies. Experiments on multiple models and benchmarks demonstrate improved generation performance, with gains in robustness and diversity metrics.

cs.AI↗

arXiv2026

TeachBench: A Syllabus-Grounded Framework for Evaluating Teaching Ability in Large Language Models

Large language models (LLMs) show promise as teaching assistants, yet their teaching capability remains insufficiently evaluated. Existing benchmarks mainly focus on problem-solving or problem-level guidance, leaving knowledge-centered teaching underexplored. We propose a syllabus-grounded evaluation framework that measures LLM teaching capability via student performance improvement after multi-turn instruction. By restricting teacher agents to structured knowledge points and example problems, the framework avoids information leakage and enables reuse of existing benchmarks. We instantiate the framework on Gaokao data across multiple subjects. Experiments reveal substantial variation in teaching effectiveness across models and domains: some models perform well in mathematics, while teaching remains challenging in physics and chemistry. We also find that incorporating example problems does not necessarily improve teaching, as models often shift toward example-specific error correction. Overall, our results highlight teaching ability as a distinct and measurable dimension of LLM behavior.

cs.AI↗

arXiv2026

Towards Stable and Effective Reinforcement Learning for Mixture-of-Experts

Recent advances in reinforcement learning (RL) have substantially improved the training of large-scale language models, leading to significant gains in generation quality and reasoning ability. However, most existing research focuses on dense models, while RL training for Mixture-of-Experts (MoE) architectures remains underexplored. To address the instability commonly observed in MoE training, we propose a novel router-aware approach to optimize importance sampling (IS) weights in off-policy RL. Specifically, we design a rescaling strategy guided by router logits, which effectively reduces gradient variance and mitigates training divergence. Experimental results demonstrate that our method significantly improves both the convergence stability and the final performance of MoE models, highlighting the potential of RL algorithmic innovations tailored to MoE architectures and providing a promising direction for efficient training of large-scale expert models.

cs.LG↗

arXiv2025

Enhancing Reliability across Short and Long-Form QA via Reinforcement Learning

While reinforcement learning has unlocked unprecedented complex reasoning in large language models, it has also amplified their propensity for hallucination, creating a critical trade-off between capability and reliability. This work confronts this challenge by introducing a targeted RL framework designed to mitigate both intrinsic and extrinsic hallucinations across short and long-form question answering. We address extrinsic hallucinations (flawed internal knowledge) by creating a novel training set from open-ended conversions of TriviaQA. Concurrently, we tackle intrinsic hallucinations (unfaithfulness to context) by leveraging long-form texts from FineWeb in a fact-grounding reward scheme. To further bolster reliability, our framework explicitly rewards the model for refusing to answer unanswerable questions, thereby cultivating crucial cautiousness. Extensive experiments demonstrate that our methodology yields significant performance gains across a diverse suite of benchmarks, substantially reducing both hallucination types. Ultimately, this research contributes a practical framework for resolving the critical tension between advanced reasoning and factual trustworthiness, paving the way for more capable and reliable large language models.

cs.CL↗

arXiv2025

Stabilizing MoE Reinforcement Learning by Aligning Training and Inference Routers

Reinforcement learning (RL) has emerged as a crucial approach for enhancing the capabilities of large language models. However, in Mixture-of-Experts (MoE) models, the routing mechanism often introduces instability, even leading to catastrophic RL training collapse. We analyze the training-inference consistency of MoE models and identify a notable discrepancy in routing behaviors between the two phases. Moreover, even under identical conditions, the routing framework can yield divergent expert selections across repeated forward passes. To address this foundational inconsistency, we propose Rollout Routing Replay (R3), a method that records routing distributions from the inference engine and replays them during training. R3 significantly reduces training-inference policy KL divergence and mitigates extreme discrepancies without compromising training speed. Extensive experiments on various settings confirm that R3 succeeds in stabilizing RL training, preventing collapse and outperforming methods such as GSPO and TIS. We believe this work can offer a new solution for stabilizing RL in MoE models.

cs.CL↗

arXiv2025

LLM-REVal: Can We Trust LLM Reviewers Yet?

The rapid advancement of large language models (LLMs) has inspired researchers to integrate them extensively into the academic workflow, potentially reshaping how research is practiced and reviewed. While previous studies highlight the potential of LLMs in supporting research and peer review, their dual roles in the academic workflow and the complex interplay between research and review bring new risks that remain largely underexplored. In this study, we focus on how the deep integration of LLMs into both peer-review and research processes may influence scholarly fairness, examining the potential risks of using LLMs as reviewers by simulation. This simulation incorporates a research agent, which generates papers and revises, alongside a review agent, which assesses the submissions. Based on the simulation results, we conduct human annotations and identify pronounced misalignment between LLM-based reviews and human judgments: (1) LLM reviewers systematically inflate scores for LLM-authored papers, assigning them markedly higher scores than human-authored ones; (2) LLM reviewers persistently underrate human-authored papers with critical statements (e.g., risk, fairness), even after multiple revisions. Our analysis reveals that these stem from two primary biases in LLM reviewers: a linguistic feature bias favoring LLM-generated writing styles, and an aversion toward critical statements. These results highlight the risks and equity concerns posed to human authors and academic research if LLMs are deployed in the peer review cycle without adequate caution. On the other hand, revisions guided by LLM reviews yield quality gains in both LLM-based and human evaluations, illustrating the potential of the LLMs-as-reviewers for early-stage researchers and enhancing low-quality papers.

cs.CL↗