τ²-bench
AssessTools
A benchmark for evaluating policy-constrained customer-service agents on realistic tasks.
Why it's here
Placed in Assess: 1 article(s) of evidence from 1 source(s), led by research-stage coverage, with 1 in the last 30 days. Confidence 24%. Low accumulated evidence, so it defaults conservatively pending more signal.
Evidence (1)
- 7Hugging Face Blog·10/20/2026researchALTK-Evolve tops τ²-bench retail without weight updates
IBM researchers report that ALTK-Evolve uses learned memory to improve a GPT-4.1 customer-service agent and reach the top of the τ²-bench retail leaderboard. The method distills successful and failed attempts into short guidelines added to the system prompt, with no model weight updates or human annotation.