Filter, Then Reweight: Rethinking Optimization Granularity in On-Policy Distillation

2026-06-01

AI Summary

Existing online policy distillation methods often suffer from information loss and training instability due to insufficiently granular supervision signals at both the trajectory and token levels. To address this, this work proposes FiRe-OPD, a novel framework that uniquely integrates trajectory-level filtering with token-level soft reweighting. This approach preserves high-quality trajectories while enabling fine-grained token reweighting, thereby avoiding the information loss associated with hard selection strategies. By incorporating KL divergence-based supervision and multi-teacher distillation, FiRe-OPD consistently outperforms current methods across strong-to-weak, single-teacher, and multi-teacher settings, achieving significant performance gains of 6.25 and 18.81 points on the AIME 2024 and Miner benchmarks, respectively.

Abstract

On-Policy distillation (OPD) in large language models is shifting from full-trace KL supervision toward more selective training paradigms. Recent OPD methods increasingly focus on selecting which trajectories to learn from, which tokens are most informative, and which supervision signals are most reliable. Motivated by this trend, we rethink optimization granularity of OPD and propose \fireicon\ FiRe-OPD (Filter, then Reweight), which jointly adjusts supervision signals at both trajectory and token levels. In details, FiRe-OPD first filters trajectories to remove low-quality rollout samples, and then applies soft reweighting within the retained trajectories to emphasize informative tokens. Compared with hard token selection, FiRe-OPD leverages a soft-weighting mechanism to effectively mitigate information loss and enhance optimization stability, thereby achieving finer-grained OPD optimization. We validate the effectiveness of FiRe-OPD across strong-to-weak, single-teacher, and multi-teacher settings, and demonstrate its superiority over recent token-level OPD methods ( (e.g., +6.25 on AIME 2024 in strong-to-weak, +18.81 on Miner in multi-teacher). Our code is available at https://github.com/YuYingLi0/FiRe-OPD.

Problem

• On-Policy Distillation
• Optimization Granularity
• Trajectory Filtering
• Token Reweighting
• Supervision Signal Selection

Innovation

• On-Policy Distillation
• Optimization Granularity
• Trajectory Filtering
• Soft Reweighting
• Token-level Supervision