Agents & Tool Use
tau2-bench
tau2-bench puts a model in the role of a customer support agent for an airline, retail or telecom account, using the company tools and database while talking to a simulated customer it must also guide through steps only they can perform.
423items
15subjects
MITlicense
agents_and_tool_usedomain
textmodality
item-level responses released
Saturation status: No
Response matrix
Rasch analysis p = σ(θ − z + c)
23,210 responses, 80/20 split over cells · 15 subjects · 423 items · 32 conditions
AUC train
0.882
AUC test
0.857
Loading session strips…
Correct (1)Incorrect (0)Unobserved
Scale: 1 = correct · 0 = incorrect