Emotional Cues Enhance LLM Performance

Adding emotional language or stakes to prompts can improve LLM task performance. Phrases like “This is very important to my career” or “Take a deep breath and think carefully” have been shown to increase accuracy on reasoning tasks.

The Phenomenon

Prompts with emotional framing outperform neutral prompts on various benchmarks. This is counterintuitive, why would a statistical model respond to emotional appeals?

Proposed Explanations

1. Training Data Correlation: High-stakes human text may correlate with more careful, accurate content
2. Attention Allocation: Emotional cues may shift attention weights toward task-relevant tokens
3. Style Matching: Models generate text matching the “register” of the prompt
4. Anchoring: Emotional framing provides implicit quality standards

Practical Implications

• Include importance/stakes language in production prompts
• “Please” and “thank you” may actually help (marginally)
• Match prompt tone to desired output quality level

Caveats

• Effects are modest (few percentage points)
• May not generalize across all tasks/models
• Mechanism not well understood
• Could be prompt engineering artifact

Related: 00-source—schulhoff-2024-prompt-report, 05-molecule—chain-of-thought-prompting