AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Approximation of Log-Partition Function in Policy Mirror Descent Induces Implicit Regularization for LLM Post-Training

Zhenghao Xu, Qin Lu, Changlong Yu, Tuo Zhao

ICML 2026 · Reinforcement learning › RL from feedback · approach: RL fine-tuning of generative models · 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 GRPO with PMD-MEAN under staleness 16Qwen2.5-7B · AIME 2025Avg@32: 10.52 -> 19.48Improved
replace GRPO with PMD-MEAN under staleness 16Qwen2.5-7B · AIME 2024Avg@32: 17.08 -> 19.69Improved
replace GRPO with PMD-MEAN (tau 0.1) under staleness 16Qwen3-30B-A3B-Base · AIME 2024Avg@32: 36.56 -> 50.83Improved
use larger global rollout batch (stale rollouts) instead of on-policy gradientPMD-MEAN · Qwen2.5-7Boverall time (ms/token): 0.0569 -> 0.0126 ms/tokenImproved
compare PMD-MEAN to on-policy gradient at matched qualityPMD-MEAN · Qwen2.5-7BAvg@32: 18.49 -> 19.58 averageNo clear effect
compare PMD-MEAN to GSPOPMD-MEAN · Qwen3-30B-A3B-BaseAvg@32: 43.96 -> 44.01 averageNo clear effect
compare PMD-MEAN to GSPOPMD-MEAN · Qwen2.5-7BAvg@32: 13.75 -> 19.58 averageImproved
fit partition-normalized target (PMD-PART) instead of mean-reward approximation (ablation)PMD-MEAN · Qwen2.5-7B training stabilitytraining stability: unstable, may collapseHurt
vary tau of PMD-MEAN (ablation)PMD-MEAN · Qwen2.5-7BAvg@32: average 19.58, 17.55, 19.58 for tau 0.005, 0.01, 0.02Mixed

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