Can Emotion Break AI Agents? τ-Emotion-Bench Tests Tool-Use Reliability

Introduction

When an LLM agent performs a long sequence of tool calls, evaluation usually focuses on whether it selects the right tool, supplies valid arguments, and reaches the final objective. Real users add another source of uncertainty. They may become impatient with repeated questions, react negatively to slow or inefficient progress, or abandon an interaction altogether. τ-Emotion-Bench studies this missing dimension: how does user emotion affect the reliability of tool-calling agents?

What the benchmark changes

The benchmark covers ten domains and keeps the underlying task objective fixed while varying the emotional conditions of the user interaction. This makes it possible to separate task difficulty from interaction difficulty. Rather than assuming a cooperative and behaviorally stable user, the evaluation models a setting in which the user’s response can change as the interaction unfolds.

The authors also introduce Tracer, a scalable pipeline for creating training data. Tracer constructs grounded, policy-valid tasks and collects trajectories through an interactive agent–user–environment harness. Its most important feature is the ability of a dissatisfied user to terminate an inefficient interaction. That signal helps filter trajectories in which the agent asks for too much help, delays tool execution, or reaches the answer only after excessive interaction. The resulting supervision rewards effective collaboration rather than success measured in isolation.

Key findings

• Tests with two user simulators indicate that current frontier agents remain unreliable under varied user behavior.
• Average task success does not fully reflect consistent completion; an agent may succeed in aggregate while failing repeatedly under certain interaction conditions.
• Stronger performance is associated with a better balance between tool execution and the user follow-up that is genuinely necessary.
• Tracer-generated supervision improves Qwen3 models across scales and generalizes to domains not seen during training.
• The reported gains transfer to τ-Bench and τ-Trait, suggesting that interaction-aware supervision has value beyond one benchmark.

Why it matters

The work broadens the definition of agent reliability. A system that completes a task with a cooperative user may still be unsuitable for deployment if it consumes too many turns, depends on the user to resolve avoidable uncertainty, or fails when frustration appears. Completion rate therefore needs to be considered alongside interaction efficiency and the likelihood that the user will continue participating.

The benchmark also points to a practical direction for training. User emotion can serve as an interaction-level signal for selecting trajectories and shaping agent behavior, complementing conventional tool-call correctness. The approach does not eliminate the need for evaluations with real users, and simulated emotions or termination rules may not capture every cultural or contextual difference. Still, τ-Emotion-Bench makes a useful distinction explicit: a capable agent must not only execute tools correctly, but also manage the collaboration well enough to finish before the user gives up.

Source: Hugging Face Daily Papers