CasePlay: Self-Play Reinforcement Learning from Case Reports for Medical Reasoning

Key points

• We study whether document-grounded self-play can convert these reports into effective RL training signals.1
• We identify two challenges that limit existing self-play methods in this setting: knowledge-point fixation}, where the proposer repeatedly asks about a narrow set of salient clues in a source report, and difficulty-quality mismatch, where appropriately difficult questions may still be poorly formed or shallow, limiting their value for reasoner training.1
• Extensive experiments across seven medically related benchmarks demonstrate that CasePlay outperforms existing self-play baselines and improves the SFT warm-start model by 3.1 points on average, highlighted by a 6.3-point gain on LiveClin.1
• This approach offers an effective path toward continually improving medical reasoning ability through self-play RL.1

Sources (1)

• [1]CasePlay: Self-Play Reinforcement Learning from Case Reports for Medical Reasoning
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 18, 02:06 AM

  Reinforcement learning (RL) is a promising way to improve the clinical reasoning abilities of large language models (LLMs), but current pipelines often rely on expert-labeled questions or costly synthetic-data generation.

  We study whether document-grounded self-play can convert these reports into effective RL training signals.

Extractive summary: sentences quoted from the sources.