Do audience-framed prompts work better than speaker-framed prompts?

Beyond which persona is named, the paper tests how the persona is framed. It defines two prompt types: a "Speaker-Specific Prompt" that assigns the role to the model itself ("You are a lawyer"), and an "Audience-Specific Prompt" that assigns the role to whoever the model is addressing ("You are talking to a fireman"), alongside a no-persona control and an "Imagine"-prefixed paraphrase of each as a robustness check. Running a mixed-effects model on accuracy by prompt type, with a random effect per model, the authors report that "audience-specific prompts perform better than speaker-specific prompts, and the difference is statistically significant," while cautioning that "the effect size is relatively small, suggesting that different framings of the prompt have limited impacts on model performance."

The paper does not offer a causal account of why addressing the model's imagined audience outperforms assigning the model an identity; it reports the regression result (Figure 3) without further decomposition. The framing is measured inside the same experimental pipeline as the paper's broader persona analysis — 162 roles, 2,410 factual questions, nine models from four families — so the effect holds under the same conditions as the paper's headline null result on persona identity.

The finding is a secondary result inside Do personas in prompts actually improve LLM factual accuracy?, which establishes that no persona identity reliably beats the no-persona control and that persona effects are largely unpredictable; this result instead concerns prompt structure (who is addressed) rather than persona content (who is named), and is one of the few effects in the paper's analysis that reaches significance rather than washing out. It is a narrower, structural counterpart to Do expert personas actually improve LLM factual accuracy?, which varies the named role's domain expertise rather than the grammatical frame of address — together the two notes suggest that how a system prompt addresses its subject may matter slightly more than who it names.

The effect is explicitly described as small, and the paper does not report the raw accuracy difference between the two framings, only that a mixed-effects regression coefficient is statistically significant; nor does it test whether the audience-framing advantage holds for subjective tasks, longer conversations, or models outside the nine tested here. The result should be read as a minor, significant-but-small structural signal inside a study whose larger point is that persona content does not reliably help — not as a standalone prompting technique with a demonstrated practical effect size.