arXiv subjects

Zezhong Tan

Publications and source records attributed to Zezhong Tan.

3 recordsLinked to original sources

arXiv2026

I-SDPO: Instance-Level Adaptive Self-Distillation Policy Optimization

Group Relative Policy Optimization (GRPO) learns from reward differences within a rollout group, but receives no useful relative signal when every sampled response is incorrect. Privileged self-distillation can fill this gap with dense token supervision, yet applying it throughout training creates a different failure mode: the teacher is a biased, low-variance surrogate for the reward objective, so persistent imitation can oppose reward-improving updates after the policy becomes capable of producing successful trajectories. We introduce I-SDPO (Instance-Level Adaptive Self-Distillation Policy Optimization), which treats teacher reliance as capability-dependent. I-SDPO makes one routing decision per input instance and shares it across that instance's rollout group: all-incorrect groups use a privileged self-distillation objective, whereas any-success groups remain intact for GRPO. This design uses imitation only where group-relative rewards are uninformative. A local analysis characterizes when teacher and reward directions align and shows that a non-vanishing biased distillation weight induces an optimization bias floor. The routing rule automatically reduces the expected distillation rate as success probability rises, withdrawing teacher influence without a hand-designed schedule. On SciKnowEval, I-SDPO obtains the best result in all four scientific domains and improves average mean@16 accuracy from 56.67% with GRPO to 70.31%, with a maximum domain gain of 18.24 points.

cs.LG↗

arXiv2025

ADHint: Adaptive Hints with Difficulty Priors for Reinforcement Learning

To address the limited capability expansion and low sample efficiency of Reinforcement Learning (RL), recent methods have integrated ''hints'' into post-training, which are prefix segments of complete reasoning trajectories, aiming for powerful knowledge expansion and reasoning generalization. However, existing hint-based RL methods often neglect the role of difficulty in the hint-ratio schedule and relative-advantage estimation, resulting in unstable learning and excessive imitation of off-policy hints. To address this, we propose ADHint, which explicitly integrates difficulty into both processes to achieve a better trade-off between exploration and imitation. Specifically, we propose Adaptive Hint with Sample Difficulty Prior, which evaluates the difficulty of each sample under the current policy to schedule an appropriate hint ratio for rollout generation. Furthermore, we introduce Consistency-based Gradient Modulation alongside Selective Masking for Hint Preservation, which jointly modulate token-level gradients within hints to prevent biased and destructive updates. Additionally, we propose Advantage Estimation with Rollout Difficulty Posterior, which leverages the relative difficulty of rollouts with and without hints to compute their respective advantages, yielding more balanced updates. Extensive experiments across diverse modalities, scales, model families, and domains show that ADHint achieves superior reasoning capabilities and out-of-distribution generalization. Code will be released upon paper acceptance.

cs.CV↗

arXiv2025

Towards Flash Thinking via Decoupled Advantage Policy Optimization

Recent Large Reasoning Models (LRMs) have achieved remarkable performance in solving complex problems via supervised fine-tuning (SFT) and reinforcement learning (RL). Although existing RL algorithms significantly enhance model accuracy, they still suffer from excessively lengthy responses and overthinking issues, resulting in increased inference latency and computational consumption, especially for simple tasks that require minimal reasoning. To address this, we propose a novel RL framework, DEPO, to reduce inefficient reasoning for models. Our method mainly consists of three core components: (1) an innovative advantage decoupled algorithm to guide model reduction of inefficient tokens; (2) a difficulty-aware length penalty to lower the overall length of model responses; (3) an advantage clipping method to prevent bias in policy optimization. In our experiments, applied to DeepSeek-Distill-Qwen-7B and DeepSeek-Distill-Qwen-1.5B as base models, DEPO achieves a significant reduction in sequence length by 39% and reduces excessive reasoning paths in inefficient tokens, while outperforming the base model in overall accuracy.

cs.AI↗