τ²-bench

ALTK-Evolve tops customer-service benchmark via learned memory

Hugging Face introduced ALTK-Evolve, which uses GPT-4.1 to distill short behavioral rules from an agent's training-task successes and failures and append them to the system prompt, without updating model weights. On Sierra's τ²-bench retail customer-service tasks, a basic ReAct agent reached 67.5% Pass^4, beating every retail entry on the leaderboard snapshot.