AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Proximal Supervised Fine-Tuning

Wenhong Zhu, Ruobing Xie, Rui Wang, Xingwu Sun, Di Wang, Pengfei Liu

ICLR 2026 · Reinforcement learning › RL from feedback · approach: RL fine-tuning of generative models

Read the paper

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
clip the policy ratio (PSFT) instead of plain SFTQwen2.5-7B-Instruct · Out-of-domain average (7 benchmarks)accuracy: 57.90 -> 61.26Improved
clip the policy ratio (PSFT) instead of plain SFTLlama3.1-8B-Instruct · Out-of-domain average (7 benchmarks)accuracy: 50.49 -> 59.25Improved
clip the policy ratio (PSFT) instead of plain SFTQwen2.5-7B-Instruct · IFEval looseaccuracy: 54.42 -> 73.03Improved
use PSFT instead of SFT, in-domainQwen2.5-7B-Instruct · In-domain math averageaccuracy: 47.99 -> 46.98No clear effect
add a warm-up SFT phase before PSFT (ablation)PSFT · In-domain math averageaccuracy: 46.98 -> 48.17 (Qwen); 35.93 -> 38.36 (Llama)Improved
add warm-up to PSFT (ablation)PSFT · IFEval looseaccuracy: 73.03 -> 55.07Hurt
add KL regularization (coefficient 0.5) to SFTSFT · In-domain and out-of-domain averagesaccuracy: 57.90 -> 57.38 (OOD, Qwen)No clear effect
use PSFT as cold start before GRPO instead of SFTQwen2.5-7B-Instruct · Out-of-domain average after RLaccuracy: 59.90 -> 64.06Improved
use PSFT as cold start before GRPO instead of SFTQwen2.5-7B-Instruct · In-domain math average after RLaccuracy: 52.40 -> 53.31Improved
use PSFT instead of SFT before DPOQwen3-4B-Base · AlpacaEval2 LCLC win rate (%): 16.96 -> 23.29Improved
use PSFT instead of SFT on a vision-language modelQwen2.5-7B-VL-Instruct · MMLU (text-only)accuracy: 45.39 -> 62.44Improved
remove ratio clipping from PSFT (ablation)PSFT · Qwen2.5-7B-Instruct SFT traininggradient stability and downstream resultsHurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy