AI Paper Atlas

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

ICLR 2026 · Reinforcement learning › RL from feedback · approach: RL for LLM reasoning

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 changedWhereMeasuredEffect
replace PPO-like clipping with TROLL trust-region projection (GRPO)GRPO, Qwen3-8B · DAPO-Evalsuccess rate: 0.640 -> 0.691Improved
replace clipping with TROLL (GRPO)GRPO, Qwen2.5-7B-Instruct · DAPO-Evalsuccess rate: 0.323 -> 0.389Improved
replace clipping with TROLL (GSPO)GSPO, Qwen3-8B · DAPO-Evalsuccess rate: 0.000 -> 0.706Improved
replace clipping with TROLL (PPO)PPO, Qwen3-8B · MATH-Evalsuccess rate: 0.508 -> 0.591Improved
replace clipping with TROLL (Dr.GRPO)Dr.GRPO, Qwen3-8B · MATH-Evalsuccess rate: 0.549 -> 0.546No clear effect
replace clipping with TROLL (PPO)PPO, Qwen2.5-7B-Instruct · DAPO-Trainsuccess rate: 0.444 -> 0.431Hurt
replace clipping with TROLL on code generationGRPO, Qwen3 sizes · Eurus-Codesuccess rate: +7 to +18 percentage pointsImproved
raise the KL bound epsilon (ablation)TROLL, Qwen3-8B GRPO · DAPOsuccess rateHurt
lower the KL bound epsilon (ablation)TROLL, Qwen3-8B GRPO · DAPOtraining speedHurt
use K=16 kept tokens instead of 64 (ablation)TROLL, Qwen3-8B GRPO · DAPOsuccess rateHurt
use K=256 kept tokens instead of 64 (ablation)TROLL, Qwen3-8B GRPO · DAPOsuccess rateNo clear effect
use TROLL instead of Clip at larger batch sizes (ablation)Clip vs TROLL · DAPOsuccess rateImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy