Large language models are increasingly used to guide sequential biological screens, but standard evaluations based on raw hit counts may be misleading. A new arXiv paper argues that such counts conflate the proposer's value with the discovery capacity of the underlying algorithmic harness.
The authors introduce CADS, a metric designed to measure the incremental contribution of LLM proposers. By isolating the proposer's effect, CADS offers a way to evaluate whether an LLM adds value beyond what the screening harness would achieve on its own.
The paper highlights the need for evaluation methods that distinguish between the tool and the system around it. This distinction matters as LLMs become more common in scientific discovery pipelines.