AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

TGPO: Efficient Policy Optimization through Sequence Anchor and Information Gating

Hang Ding, Dongqi Liu, Qiming Feng, Jian Li, Tong Lei, Jiafu Wu, Shuo Wang, Jiangning Zhang, Chengjie Wang, Yabiao Wang

ICML 2026 · Reinforcement learning › RL from feedback · approach: Scaling RL with verifiable rewards

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
add sequence anchor + information gate (TGPO) vs best baseline DAPOTGPO · Avg of 7 math benchmarksgreedy@1: 46.7 -> 50.5Improved
add TGPO to Llama-3.2-3B-Instruct vs DAPOTGPO · Avg of 7 math benchmarksgreedy@1: 26.2 -> 27.0Improved
add TGPO vs DAPO under samplingTGPO · Avg@32 / Pass@32Avg@32, Pass@32: 37.9 -> 38.8; 58.0 -> 62.2Improved
add sequence anchor (ablation)TGPO · Qwen2.5-Math-7B ablation avggreedy@1: 37.9 -> 42.8Improved
add information gate (ablation)TGPO · Qwen2.5-Math-7B ablation avggreedy@1: 36.9 -> 42.8Improved
add consistency trust term (ablation)TGPO · Qwen2.5-Math-7B ablation avggreedy@1: 37.2 -> 42.8Improved
add drift trust term (ablation)TGPO · Qwen2.5-Math-7B ablation avggreedy@1: 39.7 -> 42.8Improved
use tau=2 rather than 1 or 3 (ablation)TGPO · Qwen2.5-Math-7B ablation avggreedy@1: 38.9 (tau=1), 39.1 (tau=3) vs 42.8Improved
replace GSPO sequence-level weighting with TGPOGSPO · Qwen2.5-Math-7B ablation avggreedy@1: 33.8 -> 42.8Improved
retain imperfect rollouts via down-weighting vs sequence-level discardingTGPO vs GSPO · trainingnon-zero gradient fractionImproved
use token-level IS (GRPO) instead of TGPOGRPO · training dynamicsIS variance, entropyHurt

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