Loading
Loading articles
By IBM Research Team

AI summary
ALTK-Evolve improves LLM-based customer service agents by distilling task attempt history into behavioral guidelines appended to the system prompt. This method enables agents to optimize performance on the τ²-bench retail and airline domains without requiring model weight updates or manual annotation. By using an iterative feedback loop to refine instructions, the approach allows older models to achieve competitive results on complex, policy-bound tasks.
Key points