AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Temperature as a Meta-Policy: Adaptive Temperature in LLM Reinforcement Learning

Haoran Dang, Cuiling Lan, Hai Wan, Xibin Zhao, Yan Lu

ICLR 2026 · Reinforcement learning › RL from feedback · approach: RL for LLM reasoning · 4 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
learn a temperature meta-policy (TAMPO) instead of fixed temperature in GRPOGRPO on DS-Qwen-1.5B · 5 math benchmarks avgPass@1: 42.6 (best fixed T=1.5) -> 44.5Improved
learn a temperature meta-policy instead of fixed temperature in GRPOGRPO on DS-Qwen-1.5B · 5 math benchmarks avgPass@8: 62.1 -> 63.8Improved
learn a temperature meta-policy instead of linearly increasing temperatureGRPO · OlympiadBenchPass@1: 41.0 -> 39.6Hurt
raise EMA coefficient alpha from 0.01 to 0.05 (ablation)TAMPO · 5 math benchmarks avgPass@1: 41.6 -> 44.5Improved
raise EMA coefficient alpha from 0.05 to 0.10 (ablation)TAMPO · 5 math benchmarks avgPass@1: 44.5 -> 43.6Mixed
use greedy selection of temperature from meta-policy instead of top-p 0.7 (ablation)TAMPO · 5 math benchmarks avgPass@1: 44.5 -> 40.9Hurt
use top-p 0.9 instead of 0.7 for meta-policy sampling (ablation)TAMPO · 5 math benchmarks avgPass@1: 44.5 -> 43.0Hurt
apply TAMPO instead of GRPO with a different base model and taskQwen2.5-3B-Instruct · ECQAPass@1: 75.07% -> 76.12%Improved

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