ASU τ-Emotion-Bench Shows Top Agent DeepSeek V4.1-Flash Hits Only 32.2% Pass³ Under Emotional Users

ASU's τ-Emotion-Bench evaluates tool-calling agents against 594 emotion conditions across 580 executable tasks in 10 domains, with users empowered to terminate interactions via an UNSATISFIED signal. DeepSeek V4.1-Flash tops non-thinking models at 55.9% Avg@3 but only 32.2% Pass³ for consistent completion; the authors' Tracer pipeline lifts Qwen3-8B by 19.7 points to 38.1% using 1,498 trajectories across 4 domains.