Tell a language model to respond as an expert. It becomes more helpful, more aligned with human preferences, safer in its outputs. It also becomes less accurate.
Note (2026-11-13; read against the paper's full text): the paper's damage finding is not limited to discriminative tasks. It reports that expert personas damage pretraining-dependent knowledge retrieval, including within the generative MT-Bench (for example Humanities down by 0.20 and Math by 0.10), and attributes this to the persona engaging an instruction-following mode that would otherwise serve factual recall, with shorter personas reducing but not removing the effect; the essay's explanation about confident, authoritative distributions is the essay's own. PRISM, as the paper describes it, self-distills an intent-conditioned expert persona into a single gated LoRA adapter through a bootstrapping process that needs no external data or models, rather than routing between two separate responses. The experiments use six models and the limitations section says they are limited to 7 to 8 billion parameters.
Hu, Rostami, and Thomason measure the tradeoff across models, task types, and prompt configurations. Expert personas consistently improve alignment metrics — the responses read better, match human preferences more closely, avoid harmful content more reliably. The same personas consistently damage accuracy on discriminative tasks — factual questions, classification, anything with a verifiable right answer.
The mechanism: the persona shifts the model's distribution toward confident, authoritative, well-structured responses. An expert explains clearly. An expert avoids hedging. An expert sounds right. But sounding right and being right are different distributions, and the persona optimizes for the former at the expense of the latter.
PRISM routes between persona-conditioned and baseline responses through a gated adapter. Generative tasks — where alignment matters — get the persona. Discriminative tasks — where accuracy matters — get the unmodified model. The routing preserves the alignment benefit while avoiding the accuracy cost.
The structural insight: alignment and accuracy are not the same axis. Improving how the model communicates (clarity, safety, helpfulness) can degrade what it communicates (correctness, precision, faithfulness). The persona is not adding capability — it is redistributing probability mass from accurate-but-flat responses toward fluent-but-overconfident ones. The expert sounds like an expert precisely because it commits more strongly to its answers, including the wrong ones.