Leaderboard

Tool-calling (TC) agents on τ-Emotion-Bench: executable tasks across ten domains, where an emotion-conditioned user agent can end the conversation if the agent handles it badly. Each agent is evaluated with two user agents.

• Pass^1 is the average success rate over three trials. Pass^3 is the share of tasks solved in all three trials, averaged over the ten domains. Both are in %. The bar shows both on a shared 0 to 100 scale: the solid part is Pass^3 and the light part extends to Pass^1.
• An episode fails if the user agent ends it with ###UNSATISFIED### before the task is complete.
• Each user agent has its own ranking: scores are not comparable across user agents. Click a metric to sort by it.
• custom marks a submission that changed the agent side, such as its own agent loop, tools or prompts, or a model trained on τ-Emotion-Bench tasks.
• To add your model, follow the submission guide.