ALTK-Evolve: How learned memory tops the τ²-bench leaderboard on customer service tasks without a frontier model
ALTK-Evolve, a customer service agent, has recently outperformed retail competitors on the τ²-bench benchmark by utilizing GPT-4.1 without updating model weights. This non-reasoning model achieved a Pass^4 of 67.5% across four independent attempts on tasks within the retail benchmark. The agent operates using GPT-4.1's stock ReAct scaffold with learned guidelines appended to its system prompt.
ALTK-Evolve extracts these guidelines from the agent's training-task attempts and incorporates them into the agent's prompt for subsequent conversations, thereby improving its performance without the need for additional weight updates or human annotation.
Written by urgent.news from Hugging Face's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
