IBM's ALTK-Evolve Tops Customer Service Benchmark With GPT-4.1

IBM Research's ALTK-Evolve method propelled a GPT-4.1-based customer-service agent to the top of the τ²-bench leaderboard without fine-tuning or human labels. The agent automatically writes its own playbook by reviewing past trajectories and distilling behavioral rules into the system prompt. On retail tasks, Pass^4 reliability jumped from 57.5% to 67.5%; on airline tasks, a lean 14-rule core lifted Pass^4 from 35% to 45%. The approach adds only context tokens, no training overhead.