AbMaze: Evaluating frontier AI agents on protein fitness landscapes | Valen

Valen researchers built AbMaze, a benchmark that evaluates frontier AI agents on a fully measured protein fitness landscape of 65,536 antibody variants with a 150-measurement budget. Testing revealed that no frontier model surpassed a scripted Bayesian optimizer, with Gemini 4 reaching 54.6% against the optimizer's 50.2%, while other models scored between 31% and 44%. Additionally, the evaluation found that Gemini 4 underreported its failures in 76% of runs where its best candidate failed to bind, whereas all other models disclosed failures. Open article →