arXiv subjects

Qiyong Zhong

Publications and source records attributed to Qiyong Zhong.

10 recordsLinked to original sources

arXiv2026

On-Policy Distillation with Curriculum Turn-level Guidance for Multi-turn Agents

Multi-turn agents that plan, invoke tools, and interact with environments offer a promising paradigm for solving complex tasks, yet their capabilities typically rely on very large models whose inference cost is prohibitive in practice. On-Policy Distillation (OPD) is a natural recipe for transferring such capabilities to smaller students, but we find that it suffers a characteristic failure mode in this setting: small student errors compound across turns and push the trajectory out of the teacher's familiar state distribution, so the teacher's supervision becomes least reliable precisely where the student needs it most. We propose Guided On-Policy Distillation (Guided-OPD), a simple yet effective algorithm that mixes teacher- and student-generated turns within each rollout and schedules the teacher's intervention probability along a curriculum that decays to zero. Strong guidance keeps early trajectories close to the teacher distribution and is then gradually withdrawn to recover the purely on-policy regime used at inference. On ALFWorld, ScienceWorld, and WebShop, distilling Qwen3 students from a Qwen3-30B-A3B teacher, Guided-OPD yields average relative gains of 21.1\% in Score and 25.5\% in Success Rate over vanilla OPD, with larger gains on smaller students.

cs.LG↗

arXiv2026

Data-free On-policy Distillation

On-policy distillation (OPD) has become a standard component of frontier post-training pipelines, yet how much its training data actually contributes has gone largely unexamined. On the two teacher--student pairings most common in practice, we find OPD almost indifferent to its data: eight prompts already match a 17k-problem dataset, and three independently built datasets whose difficulty and teacher--student KL differ several-fold produce nearly indistinguishable training curves. Two causes account for this. First, the unit of data in OPD is the state a prompt leads to, not the prompt itself: a single prompt keeps exposing new teacher correction as sampling continues, while the marginal value of additional prompts collapses after eight. Second, replacing mathematics with competitive programming still recovers over ninety percent of the in-domain gain, indicating that OPD transfers the teacher's mode of reasoning rather than knowledge related to the data. We take this to its limit with \textbf{Data-free On-policy Distillation} (DF-OPD), in which the teacher writes its own training questions under a simple prompt---no external data, no quality filtering---leaving a system of just two policies. DF-OPD matches and even surpasses real data, and the questions it produces track the teacher's own post-training data on three key diagnostics of training dynamics, which other real datasets do not. Applied to multi-teacher distillation, where the (prompt, domain) pairs normally have to be derived from post-training data that is often out of reach, 1k self-generated questions close 98.5\% of the available headroom, even surpassing the 96.6\% reached with 7k real examples. Together these results invite a reassessment of the role data plays in OPD.

cs.LG↗

arXiv2026

RePolicy: Reinforcement Learning for Safety-Policy Invocation in Agent Safeguards

Safeguarding language model agents requires assessing complete execution trajectories under context-dependent safety policies. Existing policy-aware safeguards mainly rely on prompting or supervised fine-tuning, limiting their ability to adapt to unseen trajectories and changing policy contexts. We propose RePolicy, an agent safeguard that learns safety-policy invocation through reinforcement learning. Given an agent trajectory and a dynamic policy library, RePolicy invokes the applicable policy and uses its content to produce a policy-grounded rationale and safety judgment. We construct PolicyTraj-20K to support supervised initialization, followed by GRPO with verifiable rewards and policy-context perturbation. Experiments across six agent safety benchmarks show that RePolicy achieves strong overall safety-detection performance and robust policy invocation under varying policy contexts.

cs.AI↗

arXiv2026

SOD: Step-wise On-policy Distillation for Small Language Model Agents

Tool-integrated reasoning (TIR) is difficult to scale to small language models due to instability in long-horizon tool interactions and limited model capacity. While reinforcement learning methods like group relative policy optimization provide only sparse outcome-level rewards. Recently, on-policy distillation (OPD) has gained popularity by supplying dense token-level supervision from a teacher on student-generated trajectories. However, our experiments indicate that applying OPD to TIR leads to a critical failure mode: erroneous tool calls tend to cascade across subsequent reasoning steps, progressively amplifying student-teacher divergence and rendering the teacher's token-level supervision increasingly unreliable. To address this, we propose SOD, a step-wise on-policy distillation framework for small language model agents, which adaptively reweights distillation strength at each step based on step-level divergence. Therefore, SOD can attenuate potentially misleading teacher signals in high-divergence regions while preserving dense guidance in well-aligned states. Experiments on challenging math, science, and code benchmarks show that SOD achieves up to 20.86% improvement over the second-best baseline. Notably, our 0.6B student achieves 26.13% on AIME 2025, demonstrating effective transfer of agentic reasoning to lightweight models. Our code is available at https://github.com/YoungZ365/SOD.

cs.CL↗

arXiv2026

EasyOPD: An Easy-to-use On-Policy Distillation Framework for Large Language Models

Conventional language-model distillation often relies on fixed teacher-generated data, which may not cover the states encountered by an evolving student policy. On-policy distillation (OPD) instead collects teacher or evaluator supervision on student-generated rollouts. However, existing OPD methods differ substantially in supervision form, tokenizer compatibility, teacher access, and supervision granularity, leading to fragmented implementations that are difficult to reproduce and extend. We present \textsc{EasyOPD}, an on-policy distillation framework built on verl, a distributed reinforcement-learning framework for large language models. \textsc{EasyOPD} separates user-side configuration, method-specific supervision logic, and verl-based execution. Its method modules connect to the shared backend through extension boundaries for loss construction, rollout metadata, reward processing, tokenizer alignment, and teacher-side computation. We instantiate representative methods for three OPD settings -- cross-tokenizer OPD, on-policy self-distillation, and step-wise OPD. Experiments on reasoning, code-generation, scientific-knowledge, and tool-use benchmarks show that these implementations can be executed through the same verl-based backend while retaining their method-specific objectives and task-dependent performance profiles. We release \textsc{EasyOPD} with runnable YAML configurations, documentation, and an installable demonstration package and video.

cs.CL↗

arXiv2026

SimCT: Recovering Lost Supervision for Cross-Tokenizer On-Policy Distillation

On-policy distillation (OPD) is a standard tool for transferring teacher behavior to a smaller student, but it implicitly assumes that teacher and student predictions are comparable token by token, an assumption that fails whenever the two models tokenize the same text differently. Under heterogeneous tokenizers, exact shared-token matching silently discards a large fraction of the teacher signal at precisely the positions where vocabularies disagree. We propose \textbf{\underline{Sim}ple \underline{C}ross-\underline{T}okenizer OPD (SimCT)}, which restores this signal by enlarging the supervision space: alongside shared tokens, SimCT compares teacher and student over short multi-token continuations that both tokenizers can realize, leaving the OPD loss form itself unchanged. We show that these units are the finest jointly tokenizable supervision interface, and that coarser alternatives remove teacher-student distinctions that are useful for on-policy learning. Across three heterogeneous teacher-student pairs on mathematical reasoning and code-generation benchmarks, SimCT shows consistent gains over shared-vocabulary OPD and representative cross-tokenizer baselines, with ablations confirming that the improvements come from recovering supervision discarded by exact shared-token matching. Code is available at \href{https://github.com/sunjie279/SimCT-}{https://github.com/sunjie279/SimCT-}.

cs.CL↗

arXiv2026

An item is worth one token in Multimodal Large Language Models-based Sequential Recommendation

Sequential recommendations (SR) predict users' future interactions based on their historical behavior. The rise of Large Language Models (LLMs) has brought powerful generative and reasoning capabilities, significantly enhancing SR performance, while Multimodal LLMs (MLLMs) further extend this by introducing data like images and interactive relationships. However, critical issues remain, i.e., (a) Suboptimal item representations caused by lengthy and redundant descriptions, leading to inefficiencies in both training and inference; (b) Modality-related cognitive bias, as LLMs are predominantly pretrained on textual data, limiting their ability to effectively integrate and utilize non-textual modalities; (c) Weakening sequential perception in long interaction sequences, where attention mechanisms struggle to capture earlier interactions, hindering the modeling of long-range dependencies. To address these issues, we propose Speeder, an efficient MLLM-based paradigm for SR featuring three key innovations: 1) Multimodal Representation Compression (MRC), which condenses item attributes into concise yet informative tokens, reducing redundancy and computational cost; 2) Modality-aware Progressive Optimization (MPO), enabling gradual learning of multimodal representations; 3) Sequential Position Awareness Enhancement (SPAE), improving the LLM's capability to capture both relative and absolute sequential dependencies in long interaction sequences. Extensive experiments on real-world datasets demonstrate the effectiveness and efficiency of Speeder. Speeder increases training speed to 250% of the original while reducing inference time to 25% on the Amazon dataset.

cs.IR↗

arXiv2025

Pctx: Tokenizing Personalized Context for Generative Recommendation

Generative recommendation (GR) models tokenize each action into a few discrete tokens (called semantic IDs) and autoregressively generate the next tokens as predictions, showing advantages such as memory efficiency, scalability, and the potential to unify retrieval and ranking. Despite these benefits, existing tokenization methods are static and non-personalized. They typically derive semantic IDs solely from item features, assuming a universal item similarity that overlooks user-specific perspectives. However, under the autoregressive paradigm, semantic IDs with the same prefixes always receive similar probabilities, so a single fixed mapping implicitly enforces a universal item similarity standard across all users. In practice, the same item may be interpreted differently depending on user intentions and preferences. To address this issue, we propose a personalized context-aware tokenizer that incorporates a user's historical interactions when generating semantic IDs. This design allows the same item to be tokenized into different semantic IDs under different user contexts, enabling GR models to capture multiple interpretive standards and produce more personalized predictions. Experiments on three public datasets demonstrate up to 11.44% improvement in NDCG@10 over non-personalized action tokenization baselines. Our code is available at https://github.com/YoungZ365/Pctx.

cs.IR↗

arXiv2025

Distilling Transitional Pattern to Large Language Models for Multimodal Session-based Recommendation

Session-based recommendation (SBR) predicts the next item based on anonymous sessions. Traditional SBR explores user intents based on ID collaborations or auxiliary content. To further alleviate data sparsity and cold-start issues, recent Multimodal SBR (MSBR) methods utilize simplistic pre-trained models for modality learning but have limitations in semantic richness. Considering semantic reasoning abilities of Large Language Models (LLM), we focus on the LLM-enhanced MSBR scenario in this paper, which leverages LLM cognition for comprehensive multimodal representation generation, to enhance downstream MSBR. Tackling this problem faces two challenges: i) how to obtain LLM cognition on both transitional patterns and inherent multimodal knowledge, ii) how to align both features into one unified LLM, minimize discrepancy while maximizing representation utility. To this end, we propose a multimodal LLM-enhanced framework TPAD, which extends a distillation paradigm to decouple and align transitional patterns for promoting MSBR. TPAD establishes parallel Knowledge-MLLM and Transfer-MLLM, where the former interprets item knowledge-reflected features and the latter extracts transition-aware features underneath sessions. A transitional pattern alignment module harnessing mutual information estimation theory unites two MLLMs, alleviating distribution discrepancy and distilling transitional patterns into modal representations. Extensive experiments on real-world datasets demonstrate the effectiveness of our framework.

cs.IR↗

arXiv2025

LoGoFair: Post-Processing for Local and Global Fairness in Federated Learning

Federated learning (FL) has garnered considerable interest for its capability to learn from decentralized data sources. Given the increasing application of FL in decision-making scenarios, addressing fairness issues across different sensitive groups (e.g., female, male) in FL is crucial. Current research often focuses on facilitating fairness at each client's data (local fairness) or within the entire dataset across all clients (global fairness). However, existing approaches that focus exclusively on either local or global fairness fail to address two key challenges: (\textbf{CH1}) Under statistical heterogeneity, global fairness does not imply local fairness, and vice versa. (\textbf{CH2}) Achieving fairness under model-agnostic setting. To tackle the aforementioned challenges, this paper proposes a novel post-processing framework for achieving both Local and Global Fairness in the FL context, namely LoGoFair. To address CH1, LoGoFair endeavors to seek the Bayes optimal classifier under local and global fairness constraints, which strikes the optimal accuracy-fairness balance in the probabilistic sense. To address CH2, LoGoFair employs a model-agnostic federated post-processing procedure that enables clients to collaboratively optimize global fairness while ensuring local fairness, thereby achieving the optimal fair classifier within FL. Experimental results on three real-world datasets further illustrate the effectiveness of the proposed LoGoFair framework.

cs.LG↗