$τ$-Emotion-Bench: Exposing the Interaction Fragility of LLM Agents through Real-World User Emotions

Key points

• In long-horizon tool-calling tasks, LLM agents can fail not only because of tool-use errors, but also because of ineffective interaction with users.1
• Yet existing benchmarks largely assume cooperative and behaviorally stable users, leaving the effect of emotional variation on agent reliability underexplored.1
• To scale both benchmark construction and agent training, we introduce Tracer, a scalable agentic data-generation pipeline that constructs grounded, policy-valid tasks and collects trajectories through an interactive agent--user--environment harness.1
• Overall, $τ$-Emotion-Bench and Tracer provide a scalable framework for evaluating agent reliability under diverse user behaviors and improving tool-calling agents through interaction-aware supervision.1

Sources (2)

• [1]$τ$-Emotion-Bench: Exposing the Interaction Fragility of LLM Agents through Real-World User Emotions
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 14, 04:19 AM

  We introduce $τ$-Emotion-Bench, a ten-domain benchmark for evaluating tool-calling agents under controlled emotion-conditioned user interactions while preserving the underlying task objective.

  In long-horizon tool-calling tasks, LLM agents can fail not only because of tool-use errors, but also because of ineffective interaction with users.

• [2]τ-Emotion-Bench: Exposing the Interaction Fragility of LLM Agents through Real-World User Emotions
  Hugging Face Daily Papers · Oct 14, 04:19 AM · same content

Extractive summary: sentences quoted from the sources.