Research map / Reinforcement learning / RL from feedback
TROLL: Trust Regions Improve Reinforcement Learning for Large Language Models
Philipp Becker, Niklas Freymuth, Serge Thilges, Fabian Otto, Gerhard Neumann
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 PPO-like clipping with TROLL trust-region projection (GRPO) | GRPO, Qwen3-8B · DAPO-Eval | success rate: 0.640 -> 0.691 | Improved |
| replace clipping with TROLL (GRPO) | GRPO, Qwen2.5-7B-Instruct · DAPO-Eval | success rate: 0.323 -> 0.389 | Improved |
| replace clipping with TROLL (GSPO) | GSPO, Qwen3-8B · DAPO-Eval | success rate: 0.000 -> 0.706 | Improved |
| replace clipping with TROLL (PPO) | PPO, Qwen3-8B · MATH-Eval | success rate: 0.508 -> 0.591 | Improved |
| replace clipping with TROLL (Dr.GRPO) | Dr.GRPO, Qwen3-8B · MATH-Eval | success rate: 0.549 -> 0.546 | No clear effect |
| replace clipping with TROLL (PPO) | PPO, Qwen2.5-7B-Instruct · DAPO-Train | success rate: 0.444 -> 0.431 | Hurt |
| replace clipping with TROLL on code generation | GRPO, Qwen3 sizes · Eurus-Code | success rate: +7 to +18 percentage points | Improved |
| raise the KL bound epsilon (ablation) | TROLL, Qwen3-8B GRPO · DAPO | success rate | Hurt |
| lower the KL bound epsilon (ablation) | TROLL, Qwen3-8B GRPO · DAPO | training speed | Hurt |
| use K=16 kept tokens instead of 64 (ablation) | TROLL, Qwen3-8B GRPO · DAPO | success rate | Hurt |
| use K=256 kept tokens instead of 64 (ablation) | TROLL, Qwen3-8B GRPO · DAPO | success rate | No clear effect |
| use TROLL instead of Clip at larger batch sizes (ablation) | Clip vs TROLL · DAPO | success rate | Improved |
Closest papers by meaning
- Multi-Step Likelihood-Ratio Correction for Reinforcement Learning with Verifiable Rewards (NeurIPS 2026 · similarity 0.88)
- Trustworthy Alignment of Retrieval-Augmented Large Language Models via Reinforcement Learning (ICML 2024 · similarity 0.88)
- On Predictability of Reinforcement Learning Dynamics for Large Language Models (ICLR 2026 · similarity 0.87)
- ReMax: A Simple, Effective, and Efficient Reinforcement Learning Method for Aligning Large Language Models (ICML 2024 · similarity 0.87)
- Reinforcement Learning Finetunes Small Subnetworks in Large Language Models (NeurIPS 2025 · similarity 0.87)
- Segment Policy Optimization: Effective Segment-Level Credit Assignment in RL for Large Language Models (NeurIPS 2025 · 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.
