SettledAgents
IBM's ALTK-Evolve tops τ²-bench by learning guidelines from past runs
IBM Research released ALTK-Evolve, a library that distills guidelines from an agent's past attempts into its system prompt. IBM reports 67.5% Pass^4 on τ²-bench retail and 82.0% on airline with gpt-5.5, both first on the leaderboard as of Oct. 12. The tests are IBM's own runs, and guidelines raise per-conversation cost. No license is stated.
- HEAT
- OUTLETS
- 0
- FIRST SEEN
- LAST SIGNAL
Primary source
Hugging Facehuggingface.co/blog/ibm-research/altk-evolve-bcCoverage · 1 article, oldest first
Tuesday, October 20, 2026