Research map / Reinforcement learning / RL from feedback
Beyond Markovian: Reflective Exploration via Bayes-Adaptive RL for LLM Reasoning
Shenao Zhang, Yaqing Wang, Yinxiao Liu, Tianqi Liu, Peter Grabowski, Eugene Ie, Zhaoran Wang, Yunxuan Li
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 BARL instead of GRPO | Qwen2.5-Math-1.5B · six math benchmarks (average) | accuracy (%): 51.0 -> 53.3 | Improved |
| use BARL instead of GRPO | Qwen2.5-Math-7B · six math benchmarks (average) | accuracy (%): 57.1 -> 59.4 | Improved |
| use BARL instead of GRPO | DeepSeek-R1-Distill-Llama-8B · six math benchmarks (average) | accuracy (%): 52.1 -> 52.7 | Improved |
| use BARL instead of progress-reward baseline | Qwen2.5-Math-1.5B · AIME | accuracy (%): 14.4 -> 17.8 | Improved |
| use BARL instead of progress-reward baseline | DeepSeek-R1-Distill-Llama-8B · MATH / GSM8K | accuracy (%): GSM8K 85.9 -> 85.4 | No clear effect |
| use BARL over GRPO and progress for token efficiency (ablation) | Qwen2.5-Math-1.5B · GSM8K and MATH pass@k | average tokens: up to 39% fewer than progress, 50% fewer than GRPO | Improved |
| add Bayes-adaptive hypothesis weighting (BARL) vs Markovian RL in didactic task | small transformer · synthetic repeat-token task | evaluation accuracy | Improved |
| use more informative candidate set (|M|=3 repeated patterns) (ablation) | BARL · synthetic repeat-token task | evaluation accuracy | Improved |
| increase frequency of reflection behaviors | LLMs · math benchmarks | accuracy | No clear effect |
| restrict GRPO response length to 32 tokens (ablation) | GRPO · training accuracy | training accuracy | No clear effect |
Closest papers by meaning
- Reshaping Reasoning in LLMs: A Theoretical Analysis of RL Training Dynamics through Pattern Selection (ICLR 2026 · similarity 0.91)
- Efficient Reinforcement Learning with Large Language Model Priors (ICLR 2025 · similarity 0.90)
- REA-RL: Reflection-Aware Online Reinforcement Learning for Efficient Reasoning (ICLR 2026 · similarity 0.90)
- RL of Thoughts: Navigating LLM Reasoning with Inference-time Reinforcement Learning (ICLR 2026 · similarity 0.90)
- Beyond Accuracy: Dissecting Mathematical Reasoning for LLMs Under Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
- Resource-Efficient Reinforcement for Reasoning Large Language Models via Dynamic One-Shot Policy Refinement (ICML 2026 · similarity 0.90)
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.
