Research map / Reinforcement learning / RL from feedback
Squeeze the Soaked Sponge: Efficient Off-policy RFT for Large Language Model
Jing Liang, Jinyi Liu, Yi Ma, Hongyao Tang, YAN ZHENG, Shuyue Hu, LEI BAI, Jianye HAO
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 increased UTD ratio (ablation) | ReMix-PPO · Math avg Pass@1 (5 benchmarks, 1.5B) | avg Pass@1: 50.63 -> 52.10 | Improved |
| add KL-Convex policy constraint (ablation) | ReMix-PPO · Math avg Pass@1 (5 benchmarks, 1.5B) | avg Pass@1: 49.36 -> 52.10 | Improved |
| add policy reincarnation (ablation) | ReMix-PPO · Math avg Pass@1 (5 benchmarks, 1.5B) | avg Pass@1: 47.26 -> 52.10 | Improved |
| add UTD, KL-Convex and reincarnation on top of Mix-PPG (ablation) | Mix-PPG · Math avg Pass@1 (5 benchmarks, 1.5B) | avg Pass@1: 48.70 -> 52.10 | Improved |
| replace on-policy PPO with ReMix-PPO (1.5B, rollout efficiency) | PPO · Math avg Pass@1 (5 benchmarks, 1.5B) | avg Pass@1 at matched score: 50.61 at 0.230M -> 50.61 at 0.020M | Improved |
| replace PPO (200 steps) with ReMix-PPO (75 steps) | PPO · Math avg Pass@1 (5 benchmarks, 7B) | avg Pass@1: 60.34 -> 64.39 | Improved |
| apply ReMix to GRPO | GRPO · Math avg Pass@1 (5 benchmarks, 1.5B) | avg Pass@1 (200 steps): 48.64 -> 50.53 | Improved |
| apply ReMix-PPO to code generation (1.5B) | PPO · LiveCodeBench | Pass@1: 19.35 -> 24.01 | Improved |
| apply ReMix-PPO to code generation (7B) | PPO · LiveCodeBench | Pass@1: 33.33 -> 34.41 | Improved |
| ReMix-PPO training on 1.5B | R1-Distill-Qwen-1.5B · Math avg Pass@8 | Pass@8: 37.68 -> 51.72 | Improved |
| ReMix-PPO training on 7B | R1-Distill-Qwen-7B · Math avg Pass@8 | Pass@8: 68.63 -> 75.94 | Improved |
| vary reincarnation step T from 50 to 25 (ablation) | ReMix-PPO · Math avg Pass@1 (1.5B) | avg Pass@1: 51.52 -> 49.29 | Hurt |
Closest papers by meaning
- VESPO: Variational Sequence-level Soft Policy Optimization for Off-Policy LLM Training (NeurIPS 2026 · similarity 0.91)
- Asymmetric REINFORCE for off-Policy Reinforcement Learning: Balancing positive and negative rewards (NeurIPS 2025 · similarity 0.91)
- Reinforcement Mid-Training (ICLR 2026 · similarity 0.91)
- ROLLVERIFY: BRIDGING EFFICIENCY AND ACCURACY IN LONG-TAIL ROLLOUT REINFORCEMENT LEARNING (NeurIPS 2026 · similarity 0.90)
- On Predictability of Reinforcement Learning Dynamics for Large Language Models (ICLR 2026 · similarity 0.90)
- d1: Scaling Reasoning in Diffusion Large Language Models via Reinforcement Learning (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.
