New method targets reliability gaps in AI agent evaluations
A new method for evaluating and stabilizing AI agents focuses on consistency, not only average accuracy. In AppWorld, a ReAct agent using GPT-4.1 succeeded on 77.4% of five repeated runs on average, but completed all five runs successfully for just 53.0% of tasks. The resulting 24.4-percentage-point gap shows that some tasks are solved inconsistently even when nothing changes in the request.
The Consistency Analyzer, added to the ALTK-Evolve system, examines one recorded trajectory and resamples each decision point with five completions by default. It identifies choices vulnerable to small changes, without requiring ground-truth labels, model internals or a new end-to-end run. Those points are converted into guidelines that are supplied to the agent during inference.
On 168 AppWorld test_normal tasks, guidelines raised Pass^5 from 53.0% to 69.0% and Mean@5 from 77.4% to 81.0%, reducing the gap to 12.0 points. On related tasks, Pass^5 improved by 13.0 points. The authors distinguish Pass^5, where every repeated run must succeed, from Pass@5, where only one attempt needs to work.
Written by Sakhanda Wire AI from reporting by Hugging Face. Facts, quotes and full context are in the original article below.