Research map / Reinforcement learning / RL from feedback
DRIFT: Decoupled Rollouts and Importance-Weighted Fine-Tuning for Efficient Multi-Turn Optimization
Jian Mu, Tianyi Lin, Chengwei Qin, Zhongxiang Dai, Yao Shu
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 |
|---|---|---|---|
| replace online RL rollouts (UFO-5turn) with offline weighted SFT (DRIFT-5turn) | Qwen2.5-3B-Instruct · Average over 6 benchmarks | multi@5 accuracy (%): 60.2 -> 60.5 | Improved |
| replace online RL rollouts (UFO-5turn) with offline weighted SFT (DRIFT-5turn) | Llama3.1-8B-Instruct · Average over 6 benchmarks | multi@5 accuracy (%): 56.1 -> 55.6 | Hurt |
| replace UFO with DRIFT | Llama3.1-8B-Instruct · MATH500 | multi@5 accuracy (%): 46.4 -> 48.2 | Improved |
| replace UFO with DRIFT | Llama3.1-8B-Instruct · General average | multi@5 accuracy (%): 72.3 -> 70.1 | Hurt |
| train with 5-turn offline trajectories instead of single-turn SFT | SFT · Average over 6 benchmarks | multi@5 accuracy (%): 51.7 -> 54.9 | Improved |
| train with 5-turn offline trajectories instead of single-turn SFT | SFT · Average over 6 benchmarks | multi@5 accuracy (%): 49.9 -> 50.0 | No clear effect |
| multi-turn RL (UFO-5turn) instead of single-turn PPO | PPO · Average over 6 benchmarks | multi@5 accuracy (%): 51.2 -> 60.2 | Improved |
| use terminal-step-only supervision instead of weighting all turns (ablation) | DRIFT · MATH500 | accuracy | Improved |
| use moderate beta (0.1) rather than larger or smaller (ablation) | DRIFT · cumulative accuracy per turn | accuracy | Improved |
| use smaller discount gamma (ablation) | DRIFT · solved rate per turn | solved rate | Improved |
| vary feedback string (e.g. 'Incorrect' vs 'Incorrect. Please think again.') (ablation) | DRIFT · cumulative accuracy per turn | accuracy | No clear effect |
| DRIFT instead of UFO for training cost as turns grow | DRIFT-5turn · GPU time | GPU time | Improved |
Closest papers by meaning
- ROLLVERIFY: BRIDGING EFFICIENCY AND ACCURACY IN LONG-TAIL ROLLOUT REINFORCEMENT LEARNING (NeurIPS 2026 · similarity 0.88)
- Q-SFT: Q-Learning for Language Models via Supervised Fine-Tuning (ICLR 2025 · similarity 0.88)
- Efficient Online Reinforcement Learning Fine-Tuning Need Not Retain Offline Data (ICLR 2025 · similarity 0.87)
- Regressing the Relative Future: Efficient Policy Optimization for Multi-turn RLHF (ICLR 2025 · similarity 0.87)
- The Importance of Online Data: Understanding Preference Fine-tuning via Coverage (NeurIPS 2024 · similarity 0.87)
- SUF: Stabilized Unconstrained Fine-Tuning for Offline-to-Online Reinforcement Learning (AAAI 2024 · similarity 0.87)
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.
