Emotional Cues Enhance LLM Performance

EmotionPrompt appends emotional stimulus phrases to prompts, and LLM performance improves significantly.

The approach draws from psychological research showing language affects human performance. Phrases like “This is important for my career” or “I really need your help” apparently activate something in models too.

Results (GPT-4):

• Instruction induction: +8.00% relative improvement
• BIG-Bench tasks: +115% improvement
• Human evaluation (n=106): +10.9% in performance, truthfulness, and responsibility

This is strange but consistent: models trained on human text have absorbed human motivational patterns. Whether this reflects genuine understanding or pattern matching on “serious request → serious response” is unclear.

Either way, the practical implication is clear: tone matters, even for machines.

Related: [None yet]