Reproduce “$τ$-Emotion-Bench: Exposing the Interaction Fragility of LLM Agents through Real-World User Emotions” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Rishitosh Kumar Singh, Amir Saeidi, Shamanthak Hegde, Shrinidhi Kumbhar, Chitta Baral

The authors introduce Tracer and τ-Emotion-Bench for tool-calling agents that must complete tasks while keeping simulated users engaged. Tracer retains trajectories that complete tasks without user dissatisfaction. On τ-Emotion-Bench, Qwen3-8B fine-tuned on Tracer data improves from 37.6% to 54.7% Avg@3 with GLM-5.2 as user simulator; Avg@3 averages success over three trials, not all-trial success.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.