Research map / Reinforcement learning / RL from feedback
Lookahead Tree-Based Rollouts for Enhanced Trajectory-Level Exploration in Reinforcement Learning with Verifiable Rewards
Shangyu Xing, Siyuan Wang, Chenyuan Yang, Xinyu Dai, Xiang Ren
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 |
|---|---|---|---|
| replace stochastic sampling rollouts with LATR | DAPO (Qwen2.5-3B) · Countdown | Pass@1: 70.7 -> 74.7 | Improved |
| replace stochastic sampling rollouts with LATR | GRPO (Qwen2.5-3B) · Countdown | Pass@1: 65.9 -> 70.9 | Improved |
| replace stochastic sampling rollouts with LATR | DAPO (Qwen2.5-3B) · AMC-2023 | Pass@1: 37.8 -> 45.3 | Improved |
| replace stochastic sampling rollouts with LATR | DAPO (Qwen2.5-3B) · MATH-500 | Pass@8: 79.2 -> 79.0 | No clear effect |
| replace stochastic sampling rollouts with LATR | DAPO (Qwen2.5-3B) · Countdown | average completion length: 483 -> 367 | Improved |
| use LATR on larger/other base models | Qwen2.5-7B · Countdown | Pass@1: 70.9 -> 76.0 | Improved |
| use random branching instead of probability-threshold branching (ablation) | LATR · Countdown (DAPO) | Pass@1: 69.6 -> 74.7 | Improved |
| add similarity-based pruning (ablation) | LATR · Countdown (DAPO) | Pass@1: 71.0 -> 74.7 | Improved |
| use trajectory-level lookahead instead of token-level lookahead (ablation) | LATR · Countdown (DAPO) | Pass@1: 72.1 -> 74.7 | Improved |
| use embedding-based similarity for pruning instead of edit distance (ablation) | LATR · Countdown (DAPO) | Pass@1: 74.7 -> 72.9 | Hurt |
| use ROUGE-L or suffix matching instead of edit distance (ablation) | LATR · Countdown (DAPO) | Pass@1: 74.7 vs 73.9 (ROUGE-L), 74.9 (suffix) | No clear effect |
| use rollout down-sampling (RDS) with DAPO instead of stochastic (ablation) | DAPO · Countdown | Pass@1: 70.7 -> 68.5 | Hurt |
Closest papers by meaning
- ROLLVERIFY: BRIDGING EFFICIENCY AND ACCURACY IN LONG-TAIL ROLLOUT REINFORCEMENT LEARNING (NeurIPS 2026 · similarity 0.90)
- Learn from the Gap: Differential-Aware Advantage Pruning with Adaptive Rollout Sampling for GRPO (NeurIPS 2026 · similarity 0.89)
- SetPO: Set-Level Policy Optimization for Diversity-Preserving LLM Reasoning (ICML 2026 · similarity 0.88)
- Learning to Explore with Parameter-Space Noise: A Deep Dive into Parameter-Space Noise for Reinforcement Learning with Verifiable Rewards (NeurIPS 2026 · similarity 0.87)
- SALT: When More Rollouts Don’t Help in Group-Based Policy Optimization and How to Make Them Matter (NeurIPS 2026 · similarity 0.87)
- Depth-Breadth Synergy in RLVR: Unlocking LLM Reasoning Gains with Adaptive Exploration (ICML 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.
