AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

On the Generalization of SFT: A Reinforcement Learning Perspective with Reward Rectification

Yongliang Wu, Yizhou Zhou, Zhou Ziheng, Yingzhe Peng, Xinyu Ye, Xinting Hu, Wenbo Zhu, Lu Qi, Ming-Hsuan Yang, Xu Yang

ICLR 2026 · Reinforcement learning › RL from feedback · approach: RL for LLM reasoning

Read the paperCheck your idea against 137,217 papers

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
rescale token loss by token probability (DFT) instead of plain SFTSFT · Math benchmarks avgAverage@16 accuracy: 18.01 -> 31.58Improved
rescale token loss by token probability (DFT) instead of plain SFTQwen2.5-Math-7B · AIME24Average@16 accuracy: 2.48 -> 8.56Improved
apply DFT instead of SFT for code generationQwen2.5-Coder-7B · HumanEvalpass@1: 54.9 -> 67.7Improved
apply DFT instead of SFT for multimodal mathQwen2.5-VL-3B · MathVerse overallaccuracy: 35.66 -> 37.54Improved
use DFT on self-generated correct samples instead of DPO/RFT/GRPOQwen2.5-Math-1.5B · Math benchmarks avgAverage@16 accuracy: 32.00 (GRPO) -> 35.43Improved
apply DFT instead of SFT on factual QAQwen2.5 base · Natural Questionsaccuracy: 36.62 -> 30.14Hurt
use token-level weighting instead of sentence-level weighting (ablation)DFT · Math benchmarks avgAverage@16 accuracy: 15.75 -> 31.58Improved
use token-level weighting instead of geometric-mean weighting (ablation)DFT · Math benchmarks avgAverage@16 accuracy: 17.21 -> 31.58Improved
apply DFT instead of SFT on MathVisionQwen2.5-VL-3B · MathVisionaccuracy: 21.02 -> 22.30Improved
apply DFT instead of SFT on code, 3B baseQwen2.5-3B · MultiPL-E avgaccuracy: 39.10 -> 41.84Improved

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