Study Finds LLMs Underestimate Missing Information; Question Order Affects Diagnostic Accuracy

A joint study by Harvard Medical School, Broad Institute, Kempner Institute and Google DeepMind formalizes multi-turn information seeking as k-underdetermined constraint satisfaction and introduces the MT-INFOSEEK benchmark. It finds that LLMs can detect but severely underestimate missing information, and that question order affects diagnostic accuracy.

2026-08-19 ~ 2026-08-19 · 3 related posts