← 研踪 PaperTrailarXiv:2607.11505 · v3
Proxy OPD: On-Policy Distillation with Transferable Relative Proxy Update for Large Language Models
首次提交 2026-07-13 20:56
版本更新 2026-11-16 20:58
雷达观测 2026-11-17 17:11
摘要
Reward-based reinforcement learning (RL) improves large language models but incurs substantial costs when repeated across targets. On-policy distillation (OPD) reuses experts by matching their distributions on target-generated trajectories, yet differing base policies can confound reward-induced updates with pre-existing policy differences, potentially distorting transferred update directions. We propose Proxy OPD (P-OPD), a post-training method that transfers reward-induced relative policy updates using the target's frozen base policy as a reference. A proxy undergoes reward-guided exploration, yielding token-level log-probability changes between its initial and optimized policies. These signals guide learning on target-generated trajectories, with calibration controlling signal absorption, enabling reuse of reward-induced improvements across targets. Experiments on Qwen3-family targets in mathematical reasoning, code generation, and instruction following show that P-OPD improves targets and outperforms conventional OPD in most settings. For 1.7B-to-8B transfer, average Math Mean@16 and Code Mean@8 scores exceed OPD by 8.3 and 3.4 points, respectively. Mechanism analyses show more faithful reward-induced supervision and controllable adaptation. Cross-scale and cross-generation experiments confirm signal transferability and reusability, while multi-domain distillation reveals synergistic gains and cross-domain interference. In the measured 4B-to-8B setting, target adaptation after proxy exploration takes approximately 15\% of direct target RL time; repeated signal reuse amortizes the one-time exploration cost across compatible models. These findings support relative policy updates as effective, transferable, and composable supervision for an ``explore once, transfer multiple times'' post-training paradigm.
观测到的版本
相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。
v3 · 2026-11-16 20:58 · 最新观测版本
Proxy OPD: On-Policy Distillation with Transferable Relative Proxy Update for Large Language Models
Reward-based reinforcement learning (RL) improves large language models but incurs substantial costs when repeated across targets. On-policy distillation (OPD) reuses experts by matching their distributions on target-generated trajectories, yet differing base policies can confound reward-induced updates with pre-existing policy differences, potentially distorting transferred update directions. We propose Proxy OPD (P-OPD), a post-training method that transfers reward-induced relative policy updates using the target's frozen base policy as a reference. A proxy undergoes reward-guided exploration, yielding token-level log-probability changes between its initial and optimized policies. These signals guide learning on target-generated trajectories, with calibration controlling signal absorption, enabling reuse of reward-induced improvements across targets. Experiments on Qwen3-family targets in mathematical reasoning, code generation, and instruction following show that P-OPD improves targets and outperforms conventional OPD in most settings. For 1.7B-to-8B transfer, average Math Mean@16 and Code Mean@8 scores exceed OPD by 8.3 and 3.4 points, respectively. Mechanism analyses show more faithful reward-induced supervision and controllable adaptation. Cross-scale and cross-generation experiments confirm signal transferability and reusability, while multi-domain distillation reveals synergistic gains and cross-domain interference. In the measured 4B-to-8B setting, target adaptation after proxy exploration takes approximately 15\% of direct target RL time; repeated signal reuse amortizes the one-time exploration cost across compatible models. These findings support relative policy updates as effective, transferable, and composable supervision for an ``explore once, transfer multiple times'' post-training paradigm.
查看此版本解读 →查看此版本原文 ↗