Research: AI emotions can boost Agent performance significantly

量子位 · wechat · 2026-08-16

Research from USTC and Oxford shows that allowing AI to identify and act based on emotions significantly improves performance. Emotions like curiosity, confusion, and frustration correlate with specific skill choices. In error-prone tasks, using "negative emotions" as metacognitive signals boosted success rates from 4.4% to 31.3%. Additionally, Tianjin University's LEWM integrates emotions into world models, using them to predict future states and improving accuracy by up to 45.72%.

Original post →

More from coding & agent

coding & agent channel →