Stanford's Zitnik: Biomedical Discovery Lacks Hard Verifiers, Blocking AI Scientist Loops

marinkazitnik · x · 2026-09-01

Marinka Zitnik (Stanford) discusses why verification is the key bottleneck for closing the loop in AI-driven biomedical discovery: unlike math and program search, where proof checkers or test suites can evaluate candidates exactly and cheaply in real time, biomedicine has no such verifiers.

Her three points:

Training must account for sparse, delayed, noisy feedback, learning across loops that run weeks or months, with credit assignment both within a single reasoning trajectory and across repeated cycles.
Verification needs both soft verifiers (simulations, surrogate/world models, retrieval, learned judges) to guide at scale, and hard verifiers that interrogate the biological system directly — stronger but graded, noisy, context-dependent, costly and delayed.
Evaluation must test whether the AI scientist closes the loop: not just generating good hypotheses, but selecting informative experiments, revising hypotheses, and changing course when results contradict its explanation.

Original post →

Stanford's Zitnik: Biomedical Discovery Lacks Hard Verifiers, Blocking AI Scientist Loops

More from AGI Musings