Research map / Reinforcement learning / RL from feedback
Learning Structured Reasoning via Tractable Trajectory Control
Po-Nien Kung, Zhen Yang, Jeffrey Luo, Cheng-Fu Yang, Haikang Deng, Zi-Yi Dou, Yinfei Yang, Nanyun Peng, Zhe Gan, Kai-Wei Chang
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 intermediate power scaling beta=0.2 instead of no importance sampling (beta=0) (ablation) | Ctrl-R Qwen3-1.7B-Base · Math reasoning average (6 benchmarks) | average score: 31.79 -> 33.85 | Improved |
| use intermediate power scaling beta=0.2 instead of full IS (beta=1) (ablation) | Ctrl-R Qwen3-1.7B-Base · Math reasoning average (6 benchmarks) | average score: 32.46 -> 33.85 | Improved |
| add Ctrl-R guidance to DAPO | Qwen3-8B-Base · Math reasoning average (6 benchmarks) | average score: 54.89 -> 56.27 | Improved |
| add Ctrl-R guidance to DAPO | Qwen3-1.7B-Base · Math reasoning average (6 benchmarks) | average score: 32.58 -> 33.85 | Improved |
| add NL guidance hints to DAPO | Qwen3-8B-Base · Math reasoning average (6 benchmarks) | average score: 54.89 -> 55.02 | No clear effect |
| add reward shaping to DAPO | Qwen3-8B-Base · Math reasoning average (6 benchmarks) | average score: 54.89 -> 55.28 | No clear effect |
| add NL guidance hints to DAPO | Qwen3-1.7B-Base · Math reasoning average (6 benchmarks) | average score: 32.58 -> 33.02 | No clear effect |
| add reward shaping to DAPO | Qwen3-1.7B-Base · Math reasoning average (6 benchmarks) | average score: 32.58 -> 32.88 | No clear effect |
| add Ctrl-R guidance to GRPO | Qwen2.5-VL-7B-Instruct · VLM math average (6 benchmarks) | average score: 44.57 -> 47.17 | Improved |
| add Ctrl-R guidance instead of Trajectory-Hint | Qwen2.5-VL-7B-Instruct · VLM math average (6 benchmarks) | average score: 46.49 -> 47.17 | Improved |
| add Ctrl-R guidance instead of OpenVLThinker SFT pre-alignment | Qwen2.5-VL-7B-Instruct · VLM math average (6 benchmarks) | average score: 46.60 -> 47.17 | Improved |
| add Ctrl-R guidance to DAPO | Qwen3-1.7B-Base · Minerva | average score: 23.16 -> 23.16 | No clear effect |
Closest papers by meaning
- RL Squeezes, SFT Expands: A Comparative Study of Reasoning LLMs (ICLR 2026 · similarity 0.90)
- Modeling Hierarchical Thinking in Large Reasoning Models (ICML 2026 · similarity 0.89)
- Attention as a Compass: Efficient Exploration for Process-Supervised RL in Reasoning Models (ICLR 2026 · similarity 0.89)
- ProRL: Prolonged Reinforcement Learning Expands Reasoning Boundaries in Large Language Models (NeurIPS 2025 · similarity 0.89)
- Reshaping Reasoning in LLMs: A Theoretical Analysis of RL Training Dynamics through Pattern Selection (ICLR 2026 · similarity 0.89)
- Prioritize the Process, Not Just the Outcome: Rewarding Latent Thought Trajectories Improves Reasoning in Looped Language Models (ICML 2026 · 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.
