ALTK-Evolve lifts a GPT-4.1 agent to the top retail score on τ²-bench
IBM Research says its ALTK-Evolve method has a GPT-4.1 customer-service agent reach 67.5% Pass^4 on the τ²-bench retail tasks, which is higher than every retail entry on the benchmark's October 12, 2026 leaderboard snapshot. The method has the model write short behavioral rules from the agent's own training attempts and add them to the prompt, with no weight updates or human annotation. On airline tasks, IBM reports the same approach with gpt-5.5 raises accuracy by 10 points on held-out test tasks, from 72.5% to 82.5%.