Searcharxiv⌕ Search

arXiv subjects

Pengyu Cheng

Publications and source records attributed to Pengyu Cheng.

At least 19 recordsLinked to original sources
arXiv2026

Fewer Tokens, Smaller Cache: Reward-Coordinated Efficient Reasoning

Large Reasoning Models (LRMs) excel on complex tasks through long chain-of-thought (CoT) reasoning, but their lengthy intermediate steps cause severe overthinking that inflates inference cost. KV-cache compression is a common solution, yet existing reasoning-oriented methods apply a uniform policy across the trajectory and judge compression only by what it removes from the cache. Two observations point the other way. First, a reasoning state's tolerance to context loss varies along the trajectory, and process reward tracks it: deleting tokens at high-reward steps preserves accuracy far better than deleting the same budget at random. Second, compression is not free on the generation side, since a smaller cache leads the model to generate more tokens, partly canceling the saving. Together these motivate coordinating both sides under a single process reward. We propose ReCo (Reward-Coordinated Compression), a step-wise framework in which a lightweight process-reward estimator scores each completed step and drives three components: (1) reward-adaptive KV-cache compression that shrinks the retained cache harder at high-reward steps and less at low-reward ones, (2) a reward-banded penalty on reflection tokens that curbs redundant generation, and (3) confidence-based early stopping that triggers when the reasoning is reliable. Across three reasoning models and six benchmarks, ReCo reduces generated tokens by 37%-65% and end-to-end latency by 2.08x-2.35x over Full CoT, all while largely preserving accuracy.

cs.AI↗
arXiv2026

Skill Self-Play: Pushing the Frontier of LLM Capability with Co-Evolving Skills

LLM training is shifting from manual design and annotation to interaction-driven self-evolution. However, existing self-evolutionary methods face a fundamental dilemma between task diversity and verification reliability: environment-bound methods obtain precise feedback but confine learning to narrow domains, while open-ended self-generation broadens the task space but lacks reliable verification, allowing misleading rewards to pollute the training loop. We identify agent skills as a powerful middle ground to reconcile this tension: each skill ensures deep, verifiable execution in a specific scenario, while dynamic routing across skills maintains open-ended task variety. Leveraging this insight, we introduce Skill Self-Play (Skill-SP), a co-evolutionary framework comprising a proposer, a solver, and a dynamic skill controller. Orchestrated via a reinforcement learning loop, these components co-evolve in a continuous self-play loop: the proposer generates challenging tasks conditioned on dynamically sampled skills; the solver explores candidate solutions to push its capability boundaries; and the skill controller collects execution feedback to update and expand the skill library. This interactive co-evolution effectively bridges the gap between structured verification and open-ended exploration. Empirical evaluations on tool-use and reasoning benchmarks demonstrate that Skill-SP, serving as a robust evolution engine, consistently pushes the performance ceiling of competent backbones while catalyzing striking turnarounds for initially misaligned models. Our code is available at https://github.com/Qwen-Applications/skill-self-play.

cs.CL↗
arXiv2026

Harness TTS: Towards Context-Aware Expressive Speech Synthesis with Harness Layer

Expressive speech synthesis for voice assistants requires flexible style control that adapts to explicit requests and broader interaction context. We propose Harness TTS, a lightweight control layer that wraps around a TTS engine to externalize and govern its expressive behavior. It reformulates style control as closed-set prompt-tool routing: offline, a compact registry of stylistic prompt tools is constructed with structured metadata; online, an LLM planner selects the appropriate tool based on a priority-aware observation schema, and the TTS executor synthesizes speech using the corresponding prompt audio. We evaluate Harness TTS on both routing and synthesis tasks. In routing, Qwen3-4B achieves Top-1 accuracies of 74.3%, 43.0%, and 64.6% on explicit, implicit, and conflict subsets. For synthesis, experiments on CosyVoice3 and VoxCPM2 show that Harness TTS outperforms instruction-only control, achieving higher instruction-following win rates (margins of 23.1-35.6 points on CosyVoice3 and 13.8-20.0 points on VoxCPM2) and improving UTMOSv2 scores by 0.11-0.38. Moreover, the 4B planner delivers its first tool recommendation in under 50 ms in standard mode, introducing negligible latency for real-time interaction. These results demonstrate that equipping TTS engines with a dedicated Harness layer offers a practical, auditable, and context-aware solution for voice assistant expression control.

cs.SD↗
arXiv2026

PolicyAlign: Direct Policy-Based Safety Alignment for Large Language Models

Safety alignment of large language models (LLMs) typically depends on high-quality supervision data, such as safe demonstrations or preference pairs. However, in real-world deployment, emerging safety requirements are often specified as natural-language policies, while corresponding supervision data may be costly, delayed, or unavailable. This creates a mismatch between rapidly evolving safety policies and conventional data-driven alignment methods. To address this, we propose PolicyAlign, a simple yet effective framework for directly aligning LLMs with safety policies. Given a safety policy, PolicyAlign first synthesizes policy-violating instructions and then performs on-policy self-distillation to internalize policy-guided behavior. To improve training stability and data efficiency, we further introduce Policy-Sensitive Filtering, which selects instructions where the policy induces the largest behavioral shift. Experiments across multiple models show that PolicyAlign consistently improves safety while maintaining low over-refusal and preserving general capabilities. PolicyAlign also generalizes to medical, legal, and financial safety scenarios, highlighting its potential as a scalable and maintainable approach to policy-based LLM safety alignment. The code is released at https://github.com/Qwen-Applications/PolicyAlign.

cs.CL↗
arXiv2026

GD$^2$PO: Mitigating Multi-Reward Conflicts via Group-Dynamic reward-Decoupled Policy Optimization

As LLMs advance, post-training reinforcement learning (RL) increasingly relies on multi-dimensional rewards to cultivate comprehensive capabilities. This shift demands new algorithms capable of optimizing diverse and potentially competing objectives simultaneously. To address this, existing methods such as Group reward-Decoupled Policy Optimization (GDPO) decompose the overall score into independent reward groups, then compute the RL loss separately within each group. However, this strategy still encounters multi-reward conflicts: a single rollout can yield positive advantages on certain reward dimensions but negative ones on others, causing opposing signals to cancel each other out during aggregation, further hindering RL training efficiency. Inspired by Dynamic sAmpling Policy Optimization (DAPO), which improves RL training efficiency by filtering out ineffective rollouts with near-zero advantages, we propose Group-Dynamic reward-Decoupled Policy Optimization (GD$^2$PO). Specifically, GD$^2$PO employs a conflict-aware filtering mechanism to mask out rollouts suffering from severe reward-wise disagreement. By preventing conflicting signals from canceling each other out, this masking strategy preserves and enhances the magnitude of effective RL advantages, thereby significantly accelerating learning efficiency. Furthermore, we introduce query-level reweighting to dynamically adjust the update intensity of each query based on its overall reward consensus. Experiments on various multi-reward scenarios, including tool calling and human preference alignment, demonstrate that GD$^2$PO consistently and significantly outperforms existing baselines. The code is available at https://github.com/Qwen-Applications/GD2PO.

cs.LG↗
arXiv2026

F3-Tokenizer: Taming Audio Autoencoder Latents for Understanding and Generation

Continuous audio autoencoders reconstruct waveforms well but often produce latents with weak structure for understanding, while self-supervised audio encoders capture semantics but are not directly decodable. This mismatch complicates a single audio tokenizer that must support both understanding and generation. We adapt continuous autoencoder latents to this setting with two components: a noise-regularized autoencoder bottleneck and a latent-side representation encoder. The bottleneck uses channel normalization and stochastic perturbation instead of KL-based variational training, yielding scale-controlled continuous latents for reconstruction and autoregressive generation. The representation encoder is trained on frozen autoencoder latents with RQ-MTP and frozen-LLM supervision. The resulting tokenizer provides high-dimensional representations for understanding while preserving normalized continuous latents as generation targets

cs.SD↗
arXiv2026

Skill-RM: Unifying Heterogeneous Evaluation Criteria via Agent Skill

Reward models (RMs) provide critical feedback signals for LLM post-training, notably in reinforced fine-tuning (RFT) and reinforcement learning (RL) pipelines. However, current reward evaluation relies on heterogeneous criteria such as rule-based verifiers, ground-truth references, procedural checklists, and complex rubrics, where a unified mechanism to integrate all types of evidence remains unexplored. To this end, we propose Skill Reward Model (Skill-RM), a unified framework that reformulates reward modeling as the execution of a reusable Reward-Evaluation Skill. By treating reward computation as a structured agentic task, Skill-RM provides a consistent interface to orchestrate heterogeneous resources, dynamically selecting and aggregating evidence tailored to the specific requirements of each input. This approach enables the reward model to move beyond static evaluation, ensuring consistency and transparency across diverse tasks. Extensive experiments on reward benchmarks and downstream applications, including best-of-N selection and reinforcement learning, demonstrate that Skill-RM consistently outperforms traditional judge baselines. Our findings suggest that Skill-RM not only provides a unified solution for reward modeling but also achieves superior performance through the strategic and dynamic orchestration of evidence. The code is at https://github.com/Qwen-Applications/Skill-RM.

cs.LG↗
arXiv2026

Search Self-play: Pushing the Frontier of Agent Capability without Supervision

Reinforcement learning with verifiable rewards (RLVR) has become the mainstream technique for training LLM agents. However, RLVR highly depends on well-crafted task queries and corresponding ground-truth answers to provide accurate rewards, which requires significant human effort and hinders the scaling of RL processes, especially in agentic scenarios. Although a few recent works explore task synthesis methods, the difficulty of generated agentic tasks can hardly be controlled to provide effective RL training advantages. To achieve agentic RLVR with higher scalability, we explore self-play training for deep search agents, in which the learning LLM utilizes multi-turn search engine calling and acts simultaneously as both a task proposer and a problem solver. The task proposer aims to generate deep search queries with well-defined ground-truth answers and increasing task difficulty. The problem solver tries to handle the generated search queries and output the correct answer predictions. To ensure that each generated search query has accurate ground truth, we collect all the searching results from the proposer's trajectory as external knowledge, then conduct retrieval-augmentation generation (RAG) to test whether the proposed query can be correctly answered with all necessary search documents provided. In this search self-play (SSP) game, the proposer and the solver co-evolve their agent capabilities through both competition and cooperation. With substantial experimental results, we find that SSP can significantly improve search agents' performance uniformly on various benchmarks without any supervision under both from-scratch and continuous RL training setups. The code is at https://github.com/Qwen-Applications/SSP.

cs.LG↗
arXiv2026

Eliminating Inductive Bias in Reward Models with Information-Theoretic Guidance

Reward models (RMs) are essential in reinforcement learning from human feedback (RLHF) to align large language models (LLMs) with human values. However, RM training data is commonly recognized as low-quality, containing inductive biases that can easily lead to overfitting and reward hacking. For example, more detailed and comprehensive responses are usually human-preferred but with more words, leading response length to become one of the inevitable inductive biases. A limited number of prior RM debiasing approaches either target a single specific type of bias or model the problem with only simple linear correlations, \textit{e.g.}, Pearson coefficients. To mitigate more complex and diverse inductive biases in reward modeling, we introduce a novel information-theoretic debiasing method called \textbf{D}ebiasing via \textbf{I}nformation optimization for \textbf{R}M (DIR). Inspired by the information bottleneck (IB), we maximize the mutual information (MI) between RM scores and human preference pairs, while minimizing the MI between RM outputs and biased attributes of preference inputs. With theoretical justification from information theory, DIR can handle more sophisticated types of biases with non-linear correlations, broadly extending the real-world application scenarios for RM debiasing methods. In experiments, we verify the effectiveness of DIR with three types of inductive biases: \textit{response length}, \textit{sycophancy}, and \textit{format}. We discover that DIR not only effectively mitigates target inductive biases but also enhances RLHF performance across diverse benchmarks, yielding better generalization abilities. The code and training recipes are available at https://github.com/Qwen-Applications/DIR.

cs.LG↗
arXiv2026

Borderless Long Speech Synthesis

Most existing text-to-speech (TTS) systems either synthesize speech sentence by sentence and stitch the results together, or drive synthesis from plain-text dialogues alone. Both approaches leave models with little understanding of global context or paralinguistic cues, making it hard to capture real-world phenomena such as multi-speaker interactions (interruptions, overlapping speech), evolving emotional arcs, and varied acoustic environments. We introduce the Borderless Long Speech Synthesis framework for agent-centric, borderless long audio synthesis. Rather than targeting a single narrow task, the system is designed as a unified capability set spanning VoiceDesigner, multi-speaker synthesis, Instruct TTS, and long-form text synthesis. On the data side, we propose a "Labeling over filtering/cleaning" strategy and design a top-down, multi-level annotation schema we call Global-Sentence-Token. On the model side, we adopt a backbone with a continuous tokenizer and add Chain-of-Thought (CoT) reasoning together with Dimension Dropout, both of which markedly improve instruction following under complex conditions. We further show that the system is Native Agentic by design: the hierarchical annotation doubles as a Structured Semantic Interface between the LLM Agent and the synthesis engine, creating a layered control protocol stack that spans from scene semantics down to phonetic detail. Text thereby becomes an information-complete, wide-band control channel, enabling a front-end LLM to convert inputs of any modality into structured generation commands, extending the paradigm from Text2Speech to borderless long speech synthesis.

cs.SD↗
arXiv2026

Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills

Large Language Model (LLM) agents increasingly rely on domain-specific skills, yet manually authoring such skills does not scale, and skills generated purely from parametric knowledge often miss critical operational pitfalls. We introduce Trace2Skill, a framework that consolidates broad execution trajectories in parallel into a unified skill directory through inductive reasoning over agent experience. Trace2Skill supports both deepening existing human-written skills and creating useful skills from weak LLM-generated drafts. Experiments demonstrate the effectiveness of Trace2Skill across diverse domains, including office workflows, math reasoning, and vision QA. Importantly, the evolved skills are not merely memorized artifacts of the trajectories used to create them: they often transfer across model scales, across model families, and to out-of-distribution settings. For example, skills evolved from Qwen3.5-35B trajectories improve a Qwen3.5-122B agent by up to $57.65$ percentage points on WikiTableQuestions. Further analyses show that Trace2Skill outperforms sequential skill editing and ReasoningBank-style retrieval memories, compresses recurring failures and workarounds into standard operating procedures (SoPs), and yields portable skills that can be reused without parameter updates or test-time retrieval.

cs.AI↗
arXiv2026

MARCH: Multi-Agent Reinforced Self-Check for LLM Hallucination

Hallucination remains a critical bottleneck for large language models (LLMs), undermining their reliability in real-world applications, especially in Retrieval-Augmented Generation (RAG) systems. While existing hallucination detection methods employ LLM-as-a-judge to verify LLM outputs against retrieved evidence, they suffer from inherent confirmation bias, where the verifier inadvertently reproduces the errors of the original generation. To address this, we introduce Multi-Agent Reinforced Self-Check for Hallucination (MARCH), a framework that enforces rigorous factual alignment by leveraging deliberate information asymmetry. MARCH orchestrates a collaborative pipeline of three specialized agents: a Solver, a Proposer, and a Checker. The Solver generates an initial RAG response, which the Proposer decomposes into claim-level verifiable atomic propositions. Crucially, the Checker validates these propositions against retrieved evidence in isolation, deprived of the Solver's original output. This well-crafted information asymmetry scheme breaks the cycle of self-confirmation bias. By training this pipeline with multi-agent reinforcement learning (MARL), we enable the agents to co-evolve and optimize factual adherence. Extensive experiments across hallucination benchmarks demonstrate that MARCH substantially reduces hallucination rates. Notably, an 8B-parameter LLM equipped with MARCH achieves performance competitive with powerful closed-source models. MARCH paves a scalable path for factual self-improvement of LLMs through co-evolution. The code is at https://github.com/Qwen-Applications/MARCH.

cs.CL↗
arXiv2026

RLHF in an SFT Way: From Optimal Solution to Reward-Weighted Alignment

Reinforcement Learning from Human Feedback (RLHF) is crucial for aligning Large Language Models (LLMs) with human values. However, RLHF has been continuously challenged by its high complexity in implementation and computation consumption, specifically for online sampling-based methods like Proximal Policy Optimization (PPO) and Group Relative Policy Optimization (GRPO). Even with recent simplifications, such as Direct Preference Optimization (DPO) that designs an offline implicit reward learning objective relying on pre-collected preference datasets, the problems of over-fitting and training instability remain hindering the alignment process from the expected optimal performance. To address the existing challenges, we propose a novel simplification of RLHF from the perspective of variational inference, called Variational Alignment with Re-weighting (VAR). Specifically, by directly minimizing the distribution gap between the learning LLM policy and the optimal solution of RLHF, we transform the alignment objective into an offline reward-driven re-weighted supervised fine-tuning (SFT) form, which only requires minor adjustment on the SFT loss to obtain noticeable improvement on training stability and effectiveness. In comprehensive evaluation benchmarks, our objective empowers LLMs to outperform offline alignments, demonstrating superior performance in both helpfulness and harmlessness metrics (avg. $\uparrow7.16\%$ than DPO). Meanwhile, when compared to online sampling methods, our method is also comparable even better while significantly reducing computational overhead and accelerating convergence speed (over $5\times$ faster than GRPO), suggesting our approach as an efficient and effective solution in bridging the gap between efficiency and performance in LLM alignment.

cs.LG↗
arXiv2026

CLIPO: Contrastive Learning in Policy Optimization Generalizes RLVR

Reinforcement Learning with Verifiable Rewards (RLVR) has significantly advanced the reasoning capacity of Large Language Models (LLMs). However, RLVR solely relies on final answers as outcome rewards, neglecting the correctness of intermediate reasoning steps. Training on these process-wrong but outcome-correct rollouts can lead to hallucination and answer-copying, severely undermining the model's generalization and robustness. To address this, we incorporate a Contrastive Learning mechanism into the Policy Optimization (CLIPO) to generalize the RLVR process. By optimizing a contrastive loss over successful rollouts, CLIPO steers the LLM to capture the invariant structure shared across correct reasoning paths. This provides a more robust cross-trajectory regularization than the original single-path supervision in RLVR, effectively mitigating step-level reasoning inconsistencies and suppressing hallucinatory artifacts. In experiments, CLIPO consistently improves multiple RLVR baselines across diverse reasoning benchmarks, demonstrating uniform improvements in generalization and robustness for policy optimization of LLMs. Our code and training recipes are available at https://github.com/Qwen-Applications/CLIPO.

cs.LG↗
arXiv2025

Kimi-VL Technical Report

We present Kimi-VL, an efficient open-source Mixture-of-Experts (MoE) vision-language model (VLM) that offers advanced multimodal reasoning, long-context understanding, and strong agent capabilities - all while activating only 2.8B parameters in its language decoder (Kimi-VL-A3B). Kimi-VL demonstrates strong performance across challenging domains: as a general-purpose VLM, Kimi-VL excels in multi-turn agent tasks (e.g., OSWorld), matching flagship models. Furthermore, it exhibits remarkable capabilities across diverse challenging vision language tasks, including college-level image and video comprehension, OCR, mathematical reasoning, and multi-image understanding. In comparative evaluations, it effectively competes with cutting-edge efficient VLMs such as GPT-4o-mini, Qwen2.5-VL-7B, and Gemma-3-12B-IT, while surpassing GPT-4o in several key domains. Kimi-VL also advances in processing long contexts and perceiving clearly. With a 128K extended context window, Kimi-VL can process diverse long inputs, achieving impressive scores of 64.5 on LongVideoBench and 35.1 on MMLongBench-Doc. Its native-resolution vision encoder, MoonViT, further allows it to see and understand ultra-high-resolution visual inputs, achieving 83.2 on InfoVQA and 34.5 on ScreenSpot-Pro, while maintaining lower computational cost for common tasks. Building upon Kimi-VL, we introduce an advanced long-thinking variant: Kimi-VL-Thinking-2506. Developed through long chain-of-thought (CoT) supervised fine-tuning (SFT) and reinforcement learning (RL), the latest model exhibits strong long-horizon reasoning capabilities (64.0 on MMMU, 46.3 on MMMU-Pro, 56.9 on MathVision, 80.1 on MathVista, 65.2 on VideoMMMU) while obtaining robust general abilities. Code and models are publicly accessible at https://github.com/MoonshotAI/Kimi-VL.

cs.CV↗
arXiv2025

Atoxia: Red-teaming Large Language Models with Target Toxic Answers

Despite the substantial advancements in artificial intelligence, large language models (LLMs) remain being challenged by generation safety. With adversarial jailbreaking prompts, one can effortlessly induce LLMs to output harmful content, causing unexpected negative social impacts. This vulnerability highlights the necessity for robust LLM red-teaming strategies to identify and mitigate such risks before large-scale application. To detect specific types of risks, we propose a novel red-teaming method that $\textbf{A}$ttacks LLMs with $\textbf{T}$arget $\textbf{Toxi}$c $\textbf{A}$nswers ($\textbf{Atoxia}$). Given a particular harmful answer, Atoxia generates a corresponding user query and a misleading answer opening to examine the internal defects of a given LLM. The proposed attacker is trained within a reinforcement learning scheme with the LLM outputting probability of the target answer as the reward. We verify the effectiveness of our method on various red-teaming benchmarks, such as AdvBench and HH-Harmless. The empirical results demonstrate that Atoxia can successfully detect safety risks in not only open-source models but also state-of-the-art black-box models such as GPT-4o.

cs.CL↗
arXiv2025

Self-playing Adversarial Language Game Enhances LLM Reasoning

We explore the potential of self-play training for large language models (LLMs) in a two-player adversarial language game called Adversarial Taboo. In this game, an attacker and a defender communicate around a target word only visible to the attacker. The attacker aims to induce the defender to speak the target word unconsciously, while the defender tries to infer the target word from the attacker's utterances. To win the game, both players must have sufficient knowledge about the target word and high-level reasoning ability to infer and express in this information-reserved conversation. Hence, we are curious about whether LLMs' reasoning ability can be further enhanced by Self-Playing this Adversarial language Game (SPAG). With this goal, we select several open-source LLMs and let each act as the attacker and play with a copy of itself as the defender on an extensive range of target words. Through reinforcement learning on the game outcomes, we observe that the LLMs' performances uniformly improve on a broad range of reasoning benchmarks. Furthermore, iteratively adopting this self-play process can continuously promote LLMs' reasoning abilities. The code is available at https://github.com/Linear95/SPAG.

cs.CL↗
arXiv2024

On Diversified Preferences of Large Language Model Alignment

Aligning large language models (LLMs) with human preferences has been recognized as the key to improving LLMs' interaction quality. However, in this pluralistic world, human preferences can be diversified due to annotators' different tastes, which hinders the effectiveness of LLM alignment methods. This paper presents the first quantitative analysis of the experimental scaling law for reward models with varying sizes, from 1.3 billion to 7 billion parameters, trained with human feedback exhibiting diverse preferences. Our analysis reveals that the impact of diversified human preferences depends on both model size and data size. Larger models with sufficient capacity mitigate the negative effects of diverse preferences, while smaller models struggle to accommodate them. To mitigate the impact of diverse preferences, we introduce a new metric, Expected Calibration Error (ECE), to evaluate RMs and show their obvious positive correlation with the alignment performance of LLMs. Furthermore, we propose a Multi-Objective Reward learning method (MORE) to enhance the calibration performance of RMs on shared preferences. Through experiments on four models and five human preference datasets, we find the calibration error can be adopted as a key metric for evaluating RMs and MORE can obtain superior alignment performance.

cs.AI↗