AI × 生物医药· arXiv · cs.AI·· 4 hours agoPickScore60
Explore or Confirm? LLM Agents Misallocate Measurements Under Noise
In brief
The authors introduce the Wellocate benchmark, testing how five language-model agents allocate a fixed measurement budget between exploration and confirmation on a generated landscape and two real protein datasets, GB1 and PhoQ. Three models over-confirmed, and on the generated landscape one trusted single readings too much. Giving the over-confirmers a list of untried candidates markedly improved the final choices of two models on real protein data, while deferring re-testing to a final round recovered only about a quarter of the loss from wrong picks.
Why it matters
Wellocate turns the measurement-allocation biases of five language-model agents on a generated landscape and GB1 and PhoQ into reproducible readings.
Source: AI × 生物医药 · arxiv.org