IBM ALTK-Evolve gives customer-service agents learned memory without changing model weights

What is changing?

IBM Research said on October 20 that it had adapted ALTK-Evolve to customer-service tasks. The system reviews successful and failed training-task attempts by a GPT-4.1 agent, then extracts a compact playbook for steps such as authentication, transaction confirmation, and returns. The rules are appended to the system prompt in later conversations; model weights are not updated and no human annotation is required.

In IBM’s τ²-bench retail test, the method raised Pass⁴—the share of tasks solved in all four independent trials—to 67.5%. IBM says this exceeded every entry in its October 12 snapshot of the retail leaderboard. In the airline domain, the same GPT-4.1 agent reached 44% Pass⁴, while a stronger GPT-5.5 base using the same learned guidelines reached 82%.

The approach could improve agent behavior through inspectable text rules rather than retraining. The results, however, come from a specific benchmark setup; teams still need to test policy compliance, cost, latency, and incorrect tool use on their own production data.