arXiv subjects

Yirong Zeng

Publications and source records attributed to Yirong Zeng.

11 recordsLinked to original sources

arXiv2026

ATI-OPD: Adaptive Teacher Intervention for Stable Long-Sequence On-Policy Distillation

On-policy distillation (OPD) trains a student on its own responses using token-level feedback from a teacher. As responses grow longer, the student's prefix can depart from the teacher's own reasoning, weakening its feedback on later tokens. Updates based on this feedback can reinforce the same deviations in later responses and destabilize training. Prior approaches such as length control and loss reweighting limit learning from problematic responses but leave their prefixes unchanged. Fixed cutoffs can also discard later reasoning. We propose \textbf{Adaptive Teacher Intervention for On-Policy Distillation (ATI-OPD)} to improve subsequent teacher guidance by revising the student's reasoning prefix. ATI-OPD replaces the earliest high-disagreement segment with a short teacher segment, then lets the student resume its own reasoning from the revised text. To handle renewed disagreement, it excludes persistently divergent suffixes from the distillation loss while retaining useful student continuation. Training diagnostics show less teacher-student disagreement, fewer gradient spikes, and fewer extremely long responses. Across five mathematical reasoning benchmarks and two teacher-student pairs, ATI-OPD improves average student-only accuracy over standard OPD.

cs.CL↗

arXiv2026

Towards Scalable RLVR: Multimodal Instruction Following Data Synthesis and Distillation

Multimodal instruction following (MMIF) is crucial for building generalist agents. However, current training paradigms rely heavily on Supervised Fine-Tuning (SFT), which often leads to surface-level pattern matching and degrades general capabilities. While Reinforcement Learning with Verifiable Rewards (RLVR) offers a promising alternative, its scalability in MMIF is severely bottlenecked by the scarcity of high-quality, RL-ready multimodal data. To bridge this gap, we present MIFS (\textbf{M}ultimodal \textbf{I}nstruction \textbf{F}ollowing \textbf{S}ynthesis), a systematic pipeline designed to generate RL-ready multimodal data. Specifically, MIFS introduces a generative constraint protocol to synthesize diverse raw samples, followed by a learnability-aware distillation mechanism that filters data based on RL training dynamics to ensure stable policy optimization. Furthermore, a code-based verifier provides high-precision reward signals for policy learning. The resulting dataset comprises 90k samples across 8 constraint categories and 14 task domains. Empirical evaluations demonstrate that MIFS-trained MLLMs achieve an average improvement of 8.13\% on four MMIF benchmarks and a 3$\times$ faster training convergence compared to using raw data. Crucially, our approach mitigates the generalization trade-offs typical of SFT, preserving core visual capabilities while significantly boosting instruction-following precision.

cs.CL↗

arXiv2026

EnvCraft: Synthesizing Executable Environments in Agentic RL for Claw-like Agent

The paradigm of LLMs has rapidly shifted from passive language interfaces to autonomous Claw-like agents that execute long-horizon tasks across stateful workspaces. While Agentic Reinforcement Learning (Agentic RL) provides a promising path to optimize these agents, its scaling is heavily bottlenecked by the severe scarcity of interactive training environments. Existing synthetic environments are strictly limited to tool-calling endpoints, rendering them insufficient for accommodating the end-to-end real-world demands of claw-like agents. To bridge this gap, we introduce EnvCraft, an automated framework for synthesizing executable environments and scalable training data. Specifically, EnvCraft employs an environment synthesis engine to build sandbox-isolated workspaces, alongside a topology-aware data generation engine to produce coherent task trajectories. Overall, we synthesize 139 interactive environments comprising approximately 20K complex tasks for Agentic RL training. Experiments on Qwen3/3.5 models (8B-32B) show that our method yields gains of up to +11.9% on Claw-style benchmarks and +8.0% on general tool-use benchmarks, with concurrent reductions in inference token cost. The results confirm that synthesized executable environments provide robust and generalizable learning signals for training.

cs.AI↗

arXiv2026

TinyJudge: Unverifiable Constraint Alignment via Lightweight Specialist Ensembles

Instruction Following (IF) is a core capability of LLMs, requiring strict adherence to diverse constraints, ranging from verifiable ones (e.g., output length) to unverifiable ones (e.g., tone). Reinforcement learning with verifiable rewards has emerged as a paradigm for IF tasks, leveraging LLM-as-a-judge to assess unverifiable constraints. However, we empirically find that this approach remains a significant bottleneck, suffering from severe reward hacking and higher computational overhead. In this work, we first analyze the generalization capabilities of unverifiable constraints and discover that specific constraints exhibit distinct, high-generalization patterns. Motivated by this, we propose TinyJudge, a framework that employs an ensemble of specialized tiny language models ($\sim0.6B$) to provide rewards for soft constraints. By distilling expertise from frontier models into these tiny models, it achieves high-precision, lightweight evaluation. Extensive evaluations across five benchmarks demonstrate that TinyJudge outperforms the baselines by $\sim10\%$ in average performance and $12\%$ in reward precision. Crucially, it also achieves a $3\times$ speedup in total training time. Our work provides a scalable and robust path for aligning LLMs with unverifiable human instructions.

cs.CL↗

arXiv2026

AutoTool: Automatic Scaling of Tool-Use Capabilities in RL via Decoupled Entropy Constraints

Tool use represents a critical capability for AI agents, with recent advances focusing on leveraging reinforcement learning (RL) to scale up the explicit reasoning process to achieve better performance. However, there are some key challenges for tool use in current RL-based scaling approaches: (a) direct RL training often struggles to scale up thinking length sufficiently to solve complex problems, and (b) scaled-up models tend to overthink simpler problems, resulting in substantial token inefficiency. To address these challenges, we propose a novel training paradigm that first employs warm-up supervised fine-tuning to help models distinguish between simple and complex problems, followed by RL that enable models to automatically determine appropriate reasoning trajectories. Furthermore, to tackle the issue of automatic thinking-length scaling, we discover that entropy-based optimization objectives effectively maintain model diversity while successfully unlocking the model's scaling capabilities. Based on this insight, we introduce an entropy-based long-short reasoning fusion RL strategy. Our experiments on three benchmarks demonstrate that model successfully achieves auto-scaling for efficient tool use, achieving significant 9.8\% accuracy improvements while reducing computational overhead by \textasciitilde81\%.

cs.AI↗

arXiv2026

The Tool-Overuse Illusion: Why Does LLM Prefer External Tools over Internal Knowledge?

Equipping LLMs with external tools effectively addresses internal reasoning limitations. However, it introduces a critical yet under-explored phenomenon: tool overuse, the unnecessary tool-use during reasoning. In this paper, we first reveal this phenomenon is pervasive across diverse LLMs. We then experimentally elucidate its underlying mechanisms through two key lenses: (1) First, by analyzing tool-use behavior across different internal knowledge availability regions, we identify a \textit{knowledge epistemic illusion}: models misjudge internal knowledge boundaries and fail to accurately perceive their actual knowledge availability. To mitigate this, we propose a knowledge-aware epistemic boundary alignment strategy based on direct preference optimization, which reduces tool usage in by 82.8\% while yielding an accuracy improvement. (2) Second, we establish a causal link between reward structures and tool-use behavior by visualizing the tool-augmented training process. It reveals that \textit{outcome-only rewards} inadvertently encourage tool overuse by rewarding only final correctness, regardless of tool efficiency. To verify this, we balance reward signals during training rather than relying on outcome-only rewards, cutting unnecessary tool calls by 66.7\% (7B) and 60.7\% (32B) without sacrificing accuracy. Finally, we provide theoretical justification in this two lenses to understand tool overuse.

cs.AI↗

arXiv2026

Precision over Diversity: High-Precision Reward Generalizes to Robust Instruction Following

A central belief in scaling reinforcement learning with verifiable rewards for instruction following (IF) tasks is that, a diverse mixture of verifiable hard and unverifiable soft constraints is essential for generalizing to unseen instructions. In this work, we challenge this prevailing consensus through a systematic empirical investigation. Counter-intuitively, we find that models trained on hard-only constraints consistently outperform those trained on mixed datasets. Extensive experiments reveal that reward precision, rather than constraint diversity, is the primary driver of effective alignment. The LLM judge suffers from a low recall rate in detecting false response, which leads to severe reward hacking, thereby undermining the benefits of diversity. Furthermore, analysis of the attention mechanism reveals that high-precision rewards develop a transferable meta-skill for IF. Motivated by these insights, we propose a simple yet effective data-centric refinement strategy that prioritizes reward precision. Evaluated on five benchmarks, our approach outperforms competitive baselines by 13.4\% in performance while achieving a 58\% reduction in training time, maintaining strong generalization beyond instruction following. Our findings advocate for a paradigm shift: moving away from the indiscriminate pursuit of data diversity toward high-precision rewards.

cs.LG↗

arXiv2025

Tool Zero: Training Tool-Augmented LLMs via Pure RL from Scratch

Training tool-augmented LLMs has emerged as a promising approach to enhancing language models' capabilities for complex tasks. The current supervised fine-tuning paradigm relies on constructing extensive domain-specific datasets to train models. However, this approach often struggles to generalize effectively to unfamiliar or intricate tool-use scenarios. Recently, reinforcement learning (RL) paradigm can endow LLMs with superior reasoning and generalization abilities. In this work, we address a key question: Can the pure RL be used to effectively elicit a model's intrinsic reasoning capabilities and enhance the tool-agnostic generalization? We propose a dynamic generalization-guided reward design for rule-based RL, which progressively shifts rewards from exploratory to exploitative tool-use patterns. Based on this design, we introduce the Tool-Zero series models. These models are trained to enable LLMs to autonomously utilize general tools by directly scaling up RL from Zero models (i.e., base models without post-training). Experimental results demonstrate that our models achieve over 7% performance improvement compared to both SFT and RL-with-SFT models under the same experimental settings. These gains are consistently replicated across cross-dataset and intra-dataset evaluations, validating the effectiveness and robustness of our methods.

cs.LG↗

arXiv2025

iTool: Reinforced Fine-Tuning with Dynamic Deficiency Calibration for Advanced Tool Use

Augmenting large language models (LLMs) with external tools is a promising approach to enhance their capabilities, especially for complex tasks. Synthesizing tool-use data through real-world simulations is an effective way to achieve this. However, our investigation reveals that training gains significantly decay as synthetic data increases. The model struggles to benefit from additional synthetic data, which fails to endow it with advanced tool-use capabilities in complex scenarios Moreover, we discovered that the above limitation usually manifests as a fragment deficiency (i.e., parameter errors) in response. To this end, we propose an iterative reinforced fine-tuning strategy designed to alleviate this limitation. This strategy involves: (1) enhancing the diversity of response for synthetic data through path exploration of Monte Carlo Tree Search. (2) iteratively pinpointing the model's deficiency by constructing fine-grained preference pairs, and then improving it by preference optimization algorithms for targeted improvement. The experiments show that our method achieves 13.11% better performance than the same-size base model. It achieves an improvement of 6.5% in complex scenarios compared to the baseline, and it also outperforms larger open-source and closed-source models.

cs.CL↗

arXiv2025

ToolACE-DEV: Self-Improving Tool Learning via Decomposition and EVolution

The tool-using capability of large language models (LLMs) enables them to access up-to-date external information and handle complex tasks. Current approaches to enhancing this capability primarily rely on distilling advanced models by data synthesis. However, this method incurs significant costs associated with advanced model usage and often results in data compatibility issues, led by the high discrepancy in the knowledge scope between the advanced model and the target model. To address these challenges, we propose ToolACE-DEV, a self-improving framework for tool learning. First, we decompose the tool-learning objective into sub-tasks that enhance basic tool-making and tool-using abilities. Then, we introduce a self-evolving paradigm that allows lightweight models to self-improve, reducing reliance on advanced LLMs. Extensive experiments validate the effectiveness of our approach across models of varying scales and architectures.

cs.CL↗

arXiv2024

RU22Fact: Optimizing Evidence for Multilingual Explainable Fact-Checking on Russia-Ukraine Conflict

Fact-checking is the task of verifying the factuality of a given claim by examining the available evidence. High-quality evidence plays a vital role in enhancing fact-checking systems and facilitating the generation of explanations that are understandable to humans. However, the provision of both sufficient and relevant evidence for explainable fact-checking systems poses a challenge. To tackle this challenge, we propose a method based on a Large Language Model to automatically retrieve and summarize evidence from the Web. Furthermore, we construct RU22Fact, a novel multilingual explainable fact-checking dataset on the Russia-Ukraine conflict in 2022 of 16K samples, each containing real-world claims, optimized evidence, and referenced explanation. To establish a baseline for our dataset, we also develop an end-to-end explainable fact-checking system to verify claims and generate explanations. Experimental results demonstrate the prospect of optimized evidence in increasing fact-checking performance and also indicate the possibility of further progress in the end-to-end claim verification and explanation generation tasks.

cs.CL↗