Research map / Learning theory and algorithms / Prediction and decision losses
Why Tree-Style Branching Matters for Thought Advantage Estimation in GRPO
Hongcheng Wang, Yinuo Huang, Sukai Wang, Guanghui Ren, Hao Dong
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 |
|---|---|---|---|
| sample 4 answers per thought (T4A4) instead of 1 (ablation) | GRPO T4A1 · GSM8K | Pass@1: 62.54 -> 70.58 | Improved |
| sample 4 answers per thought instead of 1 (ablation) | GRPO T4A1 · HumanEval | Pass@1: 56.10 -> 58.66 | Improved |
| sample 4 answers per thought instead of 1 (ablation) | GRPO T4A1 · Trajectory Prediction | DFD: 187.99 -> 151.10 | Improved |
| answer-level branching (T4A4) instead of scaling thoughts to 16 (T16A1) (ablation) | GRPO · Math, Code, vision tasks | performance and cost: GSS@10 15 -> 5 on Math | Improved |
| increase answers per thought from 1 to 4 and 8 (ablation) | GRPO-MA · Trajectory Prediction | DFD: 187.99 -> 151.10 -> 146.37 | Improved |
| remove chain-of-thought (GRPO-NoThink with 16 answers) (ablation) | GRPO-MA · Simulated manipulation (seen) | success rate: 31.40% -> 10.60% | Hurt |
| use token-length branching (TreeRPO-64) instead of XML-tag branching (ablation) | GRPO-MA · Trajectory Prediction | DFD: 151.10 -> 157.03 | Hurt |
| scale model from 3B to 7B with T8A4 vs T8A1 (ablation) | GRPO-MA · Trajectory Prediction | DFD: 167.32 -> 134.67 | Improved |
| add loss or normalization refinements (DAPO, Dr.GRPO, GRPO-CARE) without branching (ablation) | GRPO · Trajectory Prediction | DFD: 187.99 -> 180.56 (Dr.GRPO), 188.23 (GRPO-CARE) | Mixed |
Closest papers by meaning
- S-GRPO: Early Exit via Reinforcement Learning in Reasoning Models (NeurIPS 2025 · similarity 0.88)
- Unveiling Implicit Advantage Symmetry: Why GRPO Struggles with Exploration and Difficulty Adaptation (NeurIPS 2026 · similarity 0.88)
- All Roads Lead to Rome: Incentivizing Divergent Thinking in Vision-Language Models (CVPR 2026 · similarity 0.87)
- Learn from the Gap: Differential-Aware Advantage Pruning with Adaptive Rollout Sampling for GRPO (NeurIPS 2026 · similarity 0.87)
- Think or Not? Selective Reasoning via Reinforcement Learning for Vision-Language Models (NeurIPS 2025 · similarity 0.86)
- WS-GRPO: Weakly-Supervised Group-Relative Policy Optimization for Rollout-Efficient Reasoning (ICML 2026 · similarity 0.86)
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.
