ALTK-Evolve adds learned guidelines to a GPT-4.1 agent and tops τ²-bench retail
AIIBM Research says ALTK-Evolve extracts behavioral guidelines from a GPT-4.1 agent's training attempts and appends them to the system prompt, with no weight updates. On held-out τ²-bench retail tasks, Pass^4 rises from 57.5% to 67.5% with the full guideline set, and on airline from 35.0% to 45.0% with the compact core.
AIWhy it matters.
The post shows how prompt-appended guidelines mined from agent runs change accuracy and Pass^4 on τ²-bench, with per-domain tradeoffs and cost figures.
Source details
Hugging Face Blog · huggingface.coPublished · added here
Original title: ALTK-Evolve: How learned memory tops the τ²-bench leaderboard on customer service tasks without a frontier model