AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

QUATRO: Query-Adaptive Trust Region Policy Optimization for LLM Fine-tuning

Doyeon Lee, Eunyi Lyou, Hyunsoo Cho, Soo Kyung Kim, Joonseok Lee, Jaemoo Choi

ICML 2026 · Reinforcement learning › RL from feedback · approach: GRPO stability and variants · 0 citations (OpenAlex)

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 GSPO with QUATRO (delta=0.001) at large kQwen2.5-Math-1.5B · Avg of 6 math benchmarksPass@256: 60.85 -> 68.68Improved
replace GSPO with QUATRO (delta=0.001)Qwen2.5-Math-7B · Avg of 6 math benchmarksPass@256: 65.70 -> 73.41Improved
use smaller trust-region radius delta (0.1 -> 0.001) (ablation)QUATRO · Avg of 6 math benchmarks, 1.5BPass@1: 32.63 -> 29.73Hurt
use smaller trust-region radius delta (0.1 -> 0.001) (ablation)QUATRO · Avg of 6 math benchmarks, 1.5BPass@256: 63.65 -> 68.68Improved
use smaller delta (0.1 -> 0.001) (ablation)QUATRO · Final policy entropyentropy: 0.14 -> 0.23Improved
QUATRO vs GSPO entropy controlQUATRO · Final policy entropyentropy: 0.05 -> 0.20Improved
replace GSPO with QUATRO to increase unique correct solutionsQwen2.5-Math-1.5B · MATH500UCC@256: 26.66 -> 97.20Improved
increase inner-loop updates / learning rateGSPO · math reasoning trainingreward stabilityHurt
increase inner-loop updates / learning rateQUATRO · math reasoning trainingreward stabilityNo clear effect
add log-ratio stabilizer term (ablation)QUATRO (delta=0.1, 2 inner-loop updates) · math reasoning trainingreward and KLImproved
replace GSPO with QUATRO on codeQwen2.5-Coder-1.5B · HumanEvalPass@64: 40.85 -> 50.00Improved
replace GSPO with QUATRO on code, k=1Qwen2.5-Coder-1.5B · MBPP zero-shotPass@1: 5.41 -> 6.19Improved

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