ALTK-Evolve tops customer-service agent benchmark with learned memory
What happened
On 20 October 2026, the Hugging Face blog published a first-hand report introducing ALTK-Evolve, a method that uses GPT-4.1 to distil behavioural guidelines from an agent's own task trajectories and append them to the system prompt. The method reaches 67.5% Pass^4 on the retail domain of the τ²-bench customer-service task, beating every retail entry on the leaderboard and the fine-tuned model RAFT-30B-A3B at 61.4%. The report says it tops the benchmark.
Written by AI from the reports · updated 1 h ago
Timeline
Follow the reports to see the event from each side.
- Hugging Face BlogALTK-Evolve: How learned memory tops the τ²-bench leaderboard on customer service tasks without a frontier model
ALTK-Evolve learns and applies a playbook using GPT-4.1, outperforming the retail entries on τ²-bench, a customer service agent benchmark—without updating model weights.
Heat of this event
Not enough continuous observations yet to draw a trend.