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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| rescale token loss by token probability (DFT) instead of plain SFT | SFT · Math benchmarks avg | Average@16 accuracy: 18.01 -> 31.58 | Improved |
| rescale token loss by token probability (DFT) instead of plain SFT | Qwen2.5-Math-7B · AIME24 | Average@16 accuracy: 2.48 -> 8.56 | Improved |
| apply DFT instead of SFT for code generation | Qwen2.5-Coder-7B · HumanEval | pass@1: 54.9 -> 67.7 | Improved |
| apply DFT instead of SFT for multimodal math | Qwen2.5-VL-3B · MathVerse overall | accuracy: 35.66 -> 37.54 | Improved |
| use DFT on self-generated correct samples instead of DPO/RFT/GRPO | Qwen2.5-Math-1.5B · Math benchmarks avg | Average@16 accuracy: 32.00 (GRPO) -> 35.43 | Improved |
| apply DFT instead of SFT on factual QA | Qwen2.5 base · Natural Questions | accuracy: 36.62 -> 30.14 | Hurt |
| use token-level weighting instead of sentence-level weighting (ablation) | DFT · Math benchmarks avg | Average@16 accuracy: 15.75 -> 31.58 | Improved |
| use token-level weighting instead of geometric-mean weighting (ablation) | DFT · Math benchmarks avg | Average@16 accuracy: 17.21 -> 31.58 | Improved |
| apply DFT instead of SFT on MathVision | Qwen2.5-VL-3B · MathVision | accuracy: 21.02 -> 22.30 | Improved |
| apply DFT instead of SFT on code, 3B base | Qwen2.5-3B · MultiPL-E avg | accuracy: 39.10 -> 41.84 | Improved |
Closest papers by meaning
- Vanishing Gradients in Reinforcement Finetuning of Language Models (ICLR 2024 · similarity 0.92)
- Implicit Reward as the Bridge: A Unified View of SFT and DPO Connections (NeurIPS 2025 · similarity 0.92)
- SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training (ICML 2025 · similarity 0.91)
- AceReason-Nemotron 1.1: Advancing Math and Code Reasoning through SFT and RL Synergy (ICLR 2026 · similarity 0.90)
- Q-SFT: Q-Learning for Language Models via Supervised Fine-Tuning (ICLR 2025 · similarity 0.90)
- UFT: Unifying Supervised and Reinforcement Fine-Tuning (NeurIPS 2025 · similarity 0.90)
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.
