TauBench V2 Retail

τ-bench Retail simulates a customer-service agent in a retail business. The model talks to a simulated user and must use the company's tools (orders, refunds, policy rules) to actually solve the customer's problem. Only when the database state ends up exactly right does the task count as solved.

TauBench V2 Retail measures how well a model acts as a general-purpose agent. In detail, it checks:

• calling tools with correct arguments
• following written policies while serving the simulated customer
• multi-turn conversation behaviour (asking, confirming, acting)
• correct final database results, not just polite replies

The score is accuracy: the percentage of tasks answered or solved correctly. 100% means every task was solved perfectly; higher is better.

What is the TauBench V2 Retail benchmark?

τ-bench Retail simulates a customer-service agent in a retail business. The model talks to a simulated user and must use the company's tools (orders, refunds, policy rules) to actually solve the customer's problem. Only when the database state ends up exactly right does the task count as solved.

What does TauBench V2 Retail test?

TauBench V2 Retail mainly measures how well a model acts as a general-purpose agent. Concretely, it checks: calling tools with correct arguments; following written policies while serving the simulated customer; multi-turn conversation behaviour (asking, confirming, acting).

How is the TauBench V2 Retail score calculated?

The score is accuracy: the percentage of tasks answered or solved correctly. 100% means every task was solved perfectly; higher is better. The declared maximum score of this benchmark is 100.0.

Where do I get TauBench V2 Retail data and scores?

The official source is https://github.com/sierra-research/tau2-bench. The official benchmark page/reference is linked on this page. Every score on Open LLM Radar is also available as open data: download the scores JSON (http://open-llm-radar-api:8000/api/v1/benchmarks/taubench-v2-retail/scores) or pull it directly from the Open LLM Radar API with any HTTP client.