ALTK-Evolve: How learned memory tops the τ²-bench leaderboard on customer service tasks without a frontier model

What happened: ALTK-Evolve uses GPT-4.1 to distill an agent's training-task attempts into short behavioral guidelines that are appended to the system prompt, with no weight updates or human annotation. Why it matters: The retail score is above every retail entry in the October 12, 2026 leaderboard snapshot, achieved with an older, non-reasoning model. The content frames Pass^4 as the reliability measure, since an agent that succeeds on only some attempts still fails some customers.

ALTK-Evolve uses GPT-4.1 to distill an agent's training-task attempts into short behavioral guidelines that are appended to the system prompt, with no weight updates or human annotation.

The retail score is above every retail entry in the October 12, 2026 leaderboard snapshot, achieved with an older, non-reasoning model. The content frames Pass^4 as the reliability measure, since an agent that succeeds on only some attempts still fails some customers.