R0023/2026-03-25/Q001/SRC03/E02¶
Low-knowledge personas actively reduce accuracy in specific models.
URL: https://gail.wharton.upenn.edu/research-and-insights/playing-pretend-expert-personas/
Extract¶
On GPQA Diamond: - o4-mini: All three low-knowledge personas (Layperson, Young Child, Toddler) underperformed baseline - GPT-4o: "Toddler" persona produced statistically significant accuracy reduction
Gemini 2.5 Flash frequently declined answering when given out-of-domain expert personas, refusing an average of 10.56 out of 25 trials per question, typically citing lack of relevant expertise.
Relevance to Hypotheses¶
| Hypothesis | Relationship | Strength |
|---|---|---|
| H1 | Supports | Low-knowledge personas actively degrade performance — personas can harm in both directions |
| H2 | Contradicts | Consistent degradation pattern across multiple conditions |
| H3 | Supports | Effect varies by model, demonstrating context-dependence |
Context¶
The refusal behavior of Gemini 2.5 Flash is a particularly noteworthy failure mode — the model takes the persona assignment so literally that it refuses to answer questions outside the persona's domain, making it less useful than the baseline with no persona.