AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

The Optimal Token Baseline: Variance Reduction for Long-Horizon LLM-RL

Yingru Li, Jiawei Xu, Ziniu Li, Jiacai Liu, Wei Liu, Yuxuan Tong, Longtao Zheng, Zhenghai Xue, Yaxiang Zhang, Tianle Cai, Ge Zhang, Qian Liu and 1 more

ICML 2026 · Reinforcement learning › RL from feedback · approach: RL for LLM reasoning · 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
use OTB instead of GRPO baselineGRPO (Qwen3-8B-Base) · AIME25 single-turnavg@32 peak score: 25.31 -> 30.31Improved
use OTB instead of GRPO baselineGRPO (Qwen3-8B-Base) · AIME24 single-turnavg@32 peak score: 31.35 -> 37.29Improved
use OTB instead of GRPO baselineGRPO (Qwen2.5-7B TIR) · AMC23 TIRavg@32 peak score: 59.45 -> 79.45Improved
use OTB instead of SimpleTIR trajectory filteringSimpleTIR (Qwen2.5-7B) · AIME24 TIRavg@32 peak score: 37.91 -> 41.46Improved
use OTB token-level baseline instead of optimal global baselineOGB (Qwen3-8B-Base) · AIME24 single-turnavg@32 peak score: 33.13 -> 37.29Improved
use logit-gradient proxy instead of length proxy in OTB (ablation)OTB · DAPO-MATH-17k training stabilitytraining collapseImproved
add logit-gradient proxy to OGB (ablation)OGB · DAPO-MATH-17k training stabilitytraining collapseMixed
reduce group size from N=32 to N=4 (ablation)OTB · single-turn and TIR trainingperformance / token budget: 66.03% and 68.32% fewer tokensNo clear effect
add TIS or MIS importance sampling to GRPOGRPO · single-turn reasoningtraining stabilityImproved
add TIS or MIS importance sampling to GRPOGRPO · TIRtraining stabilityHurt
increase response length from 8k to 16k and turns from 5 to 10 (ablation)OTB · TIRtraining stabilityNo clear effect
use cumulative energy (OTB) instead of isolated energy weighting (ablation)OTB · DAPO-MATH-17kperformanceImproved

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