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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| use OTB instead of GRPO baseline | GRPO (Qwen3-8B-Base) · AIME25 single-turn | avg@32 peak score: 25.31 -> 30.31 | Improved |
| use OTB instead of GRPO baseline | GRPO (Qwen3-8B-Base) · AIME24 single-turn | avg@32 peak score: 31.35 -> 37.29 | Improved |
| use OTB instead of GRPO baseline | GRPO (Qwen2.5-7B TIR) · AMC23 TIR | avg@32 peak score: 59.45 -> 79.45 | Improved |
| use OTB instead of SimpleTIR trajectory filtering | SimpleTIR (Qwen2.5-7B) · AIME24 TIR | avg@32 peak score: 37.91 -> 41.46 | Improved |
| use OTB token-level baseline instead of optimal global baseline | OGB (Qwen3-8B-Base) · AIME24 single-turn | avg@32 peak score: 33.13 -> 37.29 | Improved |
| use logit-gradient proxy instead of length proxy in OTB (ablation) | OTB · DAPO-MATH-17k training stability | training collapse | Improved |
| add logit-gradient proxy to OGB (ablation) | OGB · DAPO-MATH-17k training stability | training collapse | Mixed |
| reduce group size from N=32 to N=4 (ablation) | OTB · single-turn and TIR training | performance / token budget: 66.03% and 68.32% fewer tokens | No clear effect |
| add TIS or MIS importance sampling to GRPO | GRPO · single-turn reasoning | training stability | Improved |
| add TIS or MIS importance sampling to GRPO | GRPO · TIR | training stability | Hurt |
| increase response length from 8k to 16k and turns from 5 to 10 (ablation) | OTB · TIR | training stability | No clear effect |
| use cumulative energy (OTB) instead of isolated energy weighting (ablation) | OTB · DAPO-MATH-17k | performance | Improved |
Closest papers by meaning
- Do Not Let Low-Probability Tokens Over-Dominate in RL for LLMs (ICLR 2026 · similarity 0.91)
- STAPO: Stabilizing Reinforcement Learning for LLMs by Silencing Rare Spurious Tokens (NeurIPS 2026 · similarity 0.91)
- Accelerating RL for LLM Reasoning with Optimal Advantage Regression (NeurIPS 2025 · similarity 0.90)
- Trajectory Bellman Residual Minimization: A Simple Value-Based Method for LLM Reasoning (NeurIPS 2025 · similarity 0.90)
- Reinforcement Mid-Training (ICLR 2026 · similarity 0.90)
- Diversity-Aware Policy Optimization for Large Language Model Reasoning (NeurIPS 2025 · similarity 0.89)
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.
