MedPAIR: Physician and AI Relevance Alignment | Snorkel AI Reading Group

Yuexing Hao presents MedPAIR, a dataset that compares which sentences physician trainees and LLMs treat as relevant when answering medical exam questions. Physician trainees and board-certified physicians labeled every sentence in 933 clinical QA questions (drawn from MMLU, JAMA Clinical Challenge, MedBullets, and MedXpertQA) as highly relevant, low relevance, or irrelevant. Six LLMs were then asked the same thing, once by prompting and once through ContextCite attribution scores (for the open models). The question: when a model gets a clinical question right, is it reading the case the way a physician does? The headline result: humans and LLMs agree on only 50 to 60% of relevance labels, and keeping only physician-labeled relevant sentences improves accuracy for most, but not all, models. The talk covers the annotation design, the spurious rate metric, human vs. LLM agreement, and a Q&A on human-in-the-loop AI, simulated users, and evaluation methodology. 🔗 Links 📄 Paper: https://arxiv.org/abs/2505.24040 🌐 Dataset: http://medpair.csail.mit.edu/ 🔗 Yuexing Hao: https://yuexinghao.github.io/Yuexing-Hao/ 🐦 Yuexing Hao on X: https://x.com/YuexingHao 🔗 Yuexing Hao on LinkedIn: https://www.linkedin.com/in/yuexing-hao/ 📖 Blog post with full transcript: https://snorkel.ai/blog/medpair-measuring-whether-physicians-and-ai-agree-on-what-matters-in-medical-qa/ 🎙️ Reading Group info: https://snorkel.ai/reading-group/ Previous Reading Group sessions: ▶️ OSWorld 2.0: Why Long-Horizon Computer-Use Agents Still Fail Four Out of Five Tasks: https://snorkel.ai/blog/osworld-2-0-why-computer-use-agents-fail-most-tasks/ ▶️ Train-to-Test (T²) Scaling Laws: https://snorkel.ai/blog/train-to-test-scaling-laws-reading-group/ ▶️ Senior SWE-Bench: Evaluating Coding Agents: https://snorkel.ai/blog/senior-swe-bench-evaluating-coding-agents-like-senior-engineers/ ▶️ Agents' Last Exam: https://snorkel.ai/blog/agents-last-exam-reading-group/ ▶️ JudgmentBench: Rubric vs. Preference Evaluation: https://snorkel.ai/blog/judgment-bench-comparing-rubric-and-preference-evaluation-for-quality-assessment-legal-ai/ ▶️ Collaborative Gym: Human-Agent Collaboration: https://snorkel.ai/blog/collaborative-gym-a-framework-for-enabling-and-evaluating-human-agent-collaboration/ ▶️ Code World Models & AutoHarness: https://snorkel.ai/blog/code-world-models-and-autoharness-for-llm-agents/ ▶️ Olmix: Data Mixing for LM Development: https://snorkel.ai/blog/olmix-data-mixing-reading-group/