Research map / Large language models / Reasoning and chain of thought
Learning to Reason in Structured In-context Environments with Reinforcement Learning
Peng Yu, Zeyuan Zhao, Shao Zhang, Luoyi Fu, Xinbing Wang, Ying Wen
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes learning to reason in structured in-context environments with RL, where the abstract argues an ideal LLM reasoning environment should be scalable and general (full text unavailable, so details are limited).
Problem
RL finetuning of LLMs for reasoning depends on the environment, and an ideal reasoning environment should be scalable and general; the abstract is truncated so the specific gap is not fully stated.
Approach
Uses structured in-context environments as the RL training environment for LLM reasoning; the specific design is not stated in the available abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add structured in-context environment during RL (ablation) | GRPO on Qwen2.5-7B-Instruct · WebQSP | accuracy: 59.7 -> 93.4 | Improved |
| add structured in-context environment during RL (ablation) | GRPO on Qwen2.5-7B-Instruct · GrailQA | accuracy: 20.8 -> 85.8 | Improved |
| use RL in SIE instead of SFT on distilled structured reasoning data | Qwen2.5-7B-Instruct · WebQSP | accuracy: 40.5 -> 93.4 | Improved |
| RL fine-tuning in SIE vs CoT prompting for out-of-domain math | Qwen2.5-7B-Instruct · GSM8K | accuracy: 29.2 -> 87.4 | Improved |
| RL fine-tuning in SIE vs CoT on out-of-domain logic | Qwen2.5-7B-Instruct · KK-hard | accuracy: 19.5 -> 29.0 | Improved |
| remove all supporting triples (SIE-0% vs SIE-100%) (ablation) | RL w/ SIE Qwen2.5-7B-Instruct · WebQSP | accuracy: 93.4 -> 72.8 | Hurt |
| remove supporting triples (SIE-0% vs SIE-100%) (ablation) | RL w/ SIE Qwen2.5-7B-Instruct · GSM8K (OOD) | accuracy: 87.4 -> 87.1 | No clear effect |
| use PPO instead of GRPO (ablation) | RL w/ SIE Qwen2.5-7B-Instruct · WebQSP | accuracy: 93.4 -> 85.4 | Hurt |
| use REINFORCE++ instead of GRPO (ablation) | RL w/ SIE Qwen2.5-7B-Instruct · WebQSP | accuracy: 93.4 -> 93.1 | No clear effect |
| start RL from SFT checkpoint (ablation) | RL w/ SIE Qwen2.5-7B-Instruct · WebQSP | accuracy: 93.4 -> 88.5 (WebQSP); 29.0 -> 33.5 (KK-hard) | Mixed |
| add correctness reward instead of format-only reward (ablation) | RL w/ SIE Qwen2.5-7B-Instruct · WebQSP | accuracy: 31.6 -> 93.4 | Improved |
| replace correctness reward with random noise (ablation) | RL w/ SIE Qwen2.5-7B-Instruct · WebQSP | accuracy: 26.3 -> 6.3 | Hurt |
Closest papers by meaning
- General-Reasoner: Advancing LLM Reasoning Across All Domains (NeurIPS 2025 · similarity 0.90)
- Compositional Reasoning in Language Models under Reinforcement Learning Post-Training (NeurIPS 2026 · similarity 0.90)
- Native Reasoning Models: Training Language Models to Reason on Unverifiable Data (ICLR 2026 · similarity 0.89)
- Revisiting Reinforcement Learning for LLM Reasoning from A Cross-Domain Perspective (NeurIPS 2025 · similarity 0.89)
- RL Squeezes, SFT Expands: A Comparative Study of Reasoning LLMs (ICLR 2026 · similarity 0.89)
- Resource-Efficient Reinforcement for Reasoning Large Language Models via Dynamic One-Shot Policy Refinement (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.
