Research map / Reinforcement learning / RL from feedback
Trust-Region Adaptive Policy Optimization
Mingyu Su, Jian Guan, Yuxian Gu, Minlie Huang, Hongning Wang
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 micro-group sampling (ablation) | GRPO · Math avg (5 benchmarks) | accuracy: 50.4 -> 52.7 | Improved |
| add TrSFT loss on expert prefixes (ablation) | GRPO + micro-group sampling · Math avg (5 benchmarks) | accuracy: 52.7 -> 56.6 | Improved |
| add standard SFT loss on expert prefixes instead of TrSFT (ablation) | GRPO + micro-group sampling · Math avg (5 benchmarks) | accuracy: 52.7 -> 32.3 | Hurt |
| use LUFFY offline RL loss instead of TrSFT (ablation) | GRPO + micro-group sampling · Math avg (5 benchmarks) | accuracy: 52.7 -> 53.6 | Mixed |
| interleave TRAPO instead of two-stage SFT-then-RL | Qwen2.5-Math-7B · Math avg (5 benchmarks) | accuracy: 54.3 -> 56.6 | Improved |
| TRAPO instead of pure GRPO | Qwen2.5-Math-7B · Math avg (5 benchmarks) | accuracy: 50.4 -> 56.6 | Improved |
| TRAPO instead of SFT | Qwen2.5-Math-7B · Math avg (5 benchmarks) | accuracy: 50.3 -> 56.6 | Improved |
| TRAPO instead of SFT-then-RL | Qwen2.5-Math-7B · General avg (ARC-c, MMLU-Pro) | accuracy: 44.5 -> 68.3 | Improved |
| make first micro-group receive 0.1 expert prefix (ablation) | TRAPO · Math avg (six benchmarks, 100 steps) | accuracy: 39.4 -> 34.1 | Hurt |
| add full expert trajectory (L4=1.0) in final micro-group (ablation) | TRAPO · Math avg (six benchmarks, 100 steps) | accuracy: 37.6 -> 39.4 | Improved |
| increasing instead of constant thresholds (ablation) | TRAPO · Math avg (six benchmarks, 100 steps) | accuracy: 38.1 -> 39.4 | Improved |
| increasing instead of decreasing thresholds (ablation) | TRAPO · Math avg (six benchmarks, 100 steps) | accuracy: 37.4 -> 39.4 | Improved |
Closest papers by meaning
- BREAD: Branched Rollouts from Expert Anchors Bridge SFT & RL for Reasoning (NeurIPS 2025 · similarity 0.91)
- RL Squeezes, SFT Expands: A Comparative Study of Reasoning LLMs (ICLR 2026 · similarity 0.90)
- Accelerating RL for LLM Reasoning with Optimal Advantage Regression (NeurIPS 2025 · similarity 0.90)
- Reinforce LLM Reasoning through Multi-Agent Reflection (ICML 2025 · similarity 0.90)
- $\textbf{Re}^{2}$: Unlocking LLM Reasoning via Reinforcement Learning with Re-solving (ICLR 2026 · similarity 0.90)
- Resource-Efficient Reinforcement for Reasoning Large Language Models via Dynamic One-Shot Policy Refinement (ICML 2026 · 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.
