τ-voice evaluates real-time voice agents on 278 grounded customer-service tasks, pairing deterministic task scoring with realistic, controllable audio — diverse personas, environmental noise, and free-form turn-taking.
τ³-bench extends the benchmark family with the τ-knowledge and τ-voice tracks, advancing agent evaluation to unstructured knowledge and real-time voice.
Annotated examples of τ-voice calls — overlapping speech, interruptions, accents, and background noise — showing how the same task can succeed with clean audio and fail under realistic conditions.
We audited and fixed 50+ tasks across the airline and retail domains, addressing incorrect expected actions, ambiguous instructions, impossible constraints, and missing fallback behaviors.
τ²-bench evaluates conversational agents in dual-control environments, where both the agent and the user can act on the world.
