Research map / Reinforcement learning / RL from feedback
TreePO: Enhancing Policy Efficacy and Inference Efficiency with Tree Modeling
Yizhi LI, Qingshui Gu, Zhoufutu Wen, Ziniu Li, Ruibin Yuan, Tianshun Xing, Shuyue Guo, Tuney Zheng, Xin Zhou, Xingwei Qu, Wangchunshu Zhou, Zheng Zhang and 7 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 |
|---|---|---|---|
| add TreePO tree sampling to GRPO (ablation) | GRPO · Overall (5 math benchmarks) | Maj@16 accuracy: 46.63% -> 54.61% | Improved |
| add TreePO advantage estimator on top of tree sampling (Fixed Init Divergence) (ablation) | TreePO sampling · Overall (5 math benchmarks) | Maj@16 accuracy: 2.27% ↑ | Improved |
| add TreePO advantage estimator on top of tree sampling (More Init Divergence) (ablation) | TreePO sampling · Overall (5 math benchmarks) | Maj@16 accuracy: 3.6% ↑ | Improved |
| use tree sampling b=2 instead of sequential (Fixed Init Divergence) | TreePO Fixed Init Divergence · Overall Maj@16 and GPU hours | GPU hours: 5.78 -> 4.29 (↓26%) | Improved |
| use tree sampling b=2 instead of sequential (More Init Divergence) | TreePO More Init Divergence · Overall Maj@16 and GPU hours | Maj@16 accuracy and GPU hours: 58.21% -> 54.67%; 6.40 -> 3.65 (↓43%) | Mixed |
| use tree sampling b=8 instead of sequential (More Init Divergence) | TreePO More Init Divergence · Overall Maj@16 and GPU hours | Maj@16 accuracy and GPU hours: 58.21% -> 58.06%; 6.40 -> 5.05 (↓22%) | Mixed |
| use tree sampling b=4 instead of sequential (Fixed Init Divergence) | TreePO Fixed Init Divergence · Overall Maj@16 and GPU hours | Maj@16 accuracy and GPU hours: 56.88% -> 57.50%; 5.78 -> 4.82 (↓17%) | Mixed |
| replace size-weighted subgroup aggregation with simple averaging (ablation) | TreePO advantage · MATH and AIME training curves | accuracy, entropy | Improved |
| add subgroup-level dynamic rejection sampling (ablation) | TreePO advantage · MATH and AIME training curves | accuracy | Hurt |
| drop the root-group advantage term (ablation) | TreePO advantage · MATH and AIME training curves | accuracy | No clear effect |
| misaligned 512-token fallback with 7x1024 segments (ablation) | TreePO sampling · AIME training curve | AIME accuracy, response length | Hurt |
| use 14x512 depth-segment instead of 7x1024 (ablation) | TreePO with size-weighted advantage · MATH and AIME training curves | accuracy | Improved |
Closest papers by meaning
- Policy Guided Tree Search for Enhanced LLM Reasoning (ICML 2025 · similarity 0.89)
- SPOT: Scalable Policy Optimization with Trees for Markov Decision Processes (NeurIPS 2025 · similarity 0.88)
- DAPO : Improving Multi-Step Reasoning Abilities of Large Language Models with Direct Advantage-Based Policy Optimization (NeurIPS 2025 · similarity 0.88)
- Segment Policy Optimization: Effective Segment-Level Credit Assignment in RL for Large Language Models (NeurIPS 2025 · similarity 0.88)
- Agentic Reinforced Policy Optimization (ICLR 2026 · similarity 0.88)
- HAPO: Training Language Models to Reason Concisely via History-Aware Policy Optimization (AAAI 2026 · similarity 0.87)
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.
