AI and the World's Hard Problems · Theoretical computer science · arXiv · 2026-10-20
CADS: Measuring the Incremental Contribution of LLM Proposers in Sequential Biological Screens
Raw hit counts in sequential biological screens mix an LLM proposer's value with the discovery capacity of the underlying algorithm, so the authors present CADS, an error-controlled framework that measures the proposer's incremental contribution against an identical LLM-free loop. In DepMap-based simulations, the LLM-free loop alone clears the feedback gate, and raw counts overstate LLM contributions. Qwen2.5-7B showed a positive net contribution across 20 seeds, but the effect did not persist in a 50-seed replication.