Research map / Large language models / Instruction and fine-tuning
Beyond Two-Stage Training: Cooperative SFT and RL for LLM Reasoning
Liang Chen, Xueting Han, Li Shen, Jing Bai, Kam-Fai Wong
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 |
|---|---|---|---|
| add the meta-objective J_meta (bilevel upper-level update) (ablation) | BRIDGE on Qwen3-8B-Base · Math average (5 benchmarks) | average accuracy: 40.3 -> 49.9 | Improved |
| learn SFT guidance via bilevel LoRA teacher instead of naive SFT+RL mixing | SFT+RL on Qwen2.5-3B · Math average (5 benchmarks) | average accuracy: 29.4 -> 36.0 | Improved |
| add SFT warm-up before RL (SFT->RL vs RL) | RL on Llama-3.2-3B-Instruct · Math average (5 benchmarks) | average accuracy: 21.8 -> 18.9 | Hurt |
| add SFT warm-up before RL (SFT->RL vs RL) | RL on Qwen3-8B-Base · Math average (5 benchmarks) | average accuracy: 42.9 -> 45.5 | Improved |
| add SFT warm-up before RL (SFT->RL vs RL) | RL on Qwen2.5-3B · Math average (5 benchmarks) | average accuracy: 32.2 -> 32.2 | No clear effect |
| add periodic SFT updates during RL stage (freq=10) (ablation) | SFT->RL on Qwen2.5-3B · Math average (5 benchmarks) | average accuracy: 32.2 -> 31.4 | No clear effect |
| add periodic SFT updates during RL stage (freq=5) (ablation) | SFT->RL on Qwen2.5-3B · Math average (5 benchmarks) | average accuracy: 32.2 -> 31.5 | No clear effect |
| change mixing weight lambda from 0.5 to 0.0 (pure SFT) (ablation) | BRIDGE on Qwen3-8B-Base · Math average (5 benchmarks) | average accuracy: 49.9 -> 37.6 | Hurt |
| change mixing weight lambda from 0.5 to 1.0 (pure RL) (ablation) | BRIDGE on Qwen3-8B-Base · Math average (5 benchmarks) | average accuracy: 49.9 -> 42.9 | Hurt |
| vary lambda within 0.3-0.7 (ablation) | BRIDGE on Qwen3-8B-Base · Math average (5 benchmarks) | average accuracy: 47.2--49.9 | No clear effect |
| change LoRA rank/alpha from 32/16 to 16/32 (ablation) | BRIDGE on Qwen3-8B-Base · Math average (5 benchmarks) | average accuracy: 49.9 -> 49.5 | No clear effect |
| use BRIDGE instead of RL on Knights-and-Knaves | RL on Qwen2.5-3B · Knights-and-Knaves (avg) | accuracy: 37.0 -> 55.3 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SRFT — SRFT: A Single-Stage Method with Supervised and Reinforcement Fine-Tuning for Reasoning (ICLR 2026) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- RL Squeezes, SFT Expands: A Comparative Study of Reasoning LLMs (ICLR 2026 · similarity 0.93)
- Quagmires in SFT-RL Post-Training: When High SFT Scores Mislead and What to Use Instead (ICLR 2026 · similarity 0.92)
- AceReason-Nemotron 1.1: Advancing Math and Code Reasoning through SFT and RL Synergy (ICLR 2026 · similarity 0.92)
- BREAD: Branched Rollouts from Expert Anchors Bridge SFT & RL for Reasoning (NeurIPS 2025 · similarity 0.92)
- Revisiting Reinforcement Learning for LLM Reasoning from A Cross-Domain Perspective (NeurIPS 2025 · similarity 0.91)
- First SFT, Second RL, Third UPT: Continual Improving Multi-Modal LLM Reasoning via Unsupervised Post-Training (NeurIPS 2025 · similarity 0.91)
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.
