AI Social Psychology

Does conditioning LLMs on personal profiles improve prediction?

Persona induction—feeding LLMs participant-specific information—is widely used to make models simulate individuals more accurately. But does it actually work at the individual level where it matters most?

Note · 2026-05-28 · sourced from Cognitive Models Latent
How accurately can language models simulate human personalities?

A common recipe for making LLMs behave like specific people is persona induction: condition the model on participant-specific information — demographics, prior responses, a profile — and expect it to predict that individual more accurately. The Psych-201 study tests this at unusual scale (208,021 participants, ~26 million behavioral responses, hundreds of experiments) and finds it does not work at the level that matters. Persona-induction does not improve predictions for individuals. Conditioning on who the person is fails to sharpen the model's account of what that person will actually do.

The result is damaging precisely because the technique is so widely used and intuitively reasonable. If LLMs are to serve as surrogates — simulating patient responses for clinician training, anticipating population reactions to policy, modeling student learning trajectories — they need to capture individual variation, not just population averages. Persona induction is the standard lever for individuation, and it comes up empty here. The model conditioned on a participant's profile is not meaningfully better at predicting that participant than the model without it.

Why it matters: it converges with a body of vault evidence that LLM persona simulation captures aggregate or modal behavior far better than individual-level behavior. Where prior work showed persona simulations replicate published main effects but falter on marginal ones, and that persona-conditioned annotations are dominated by model uncertainty rather than persona knowledge, this adds a large-scale behavioral confirmation: the individuation lever itself is weak. The counterpoint is scope — persona induction may still shift population-level distributions usefully even when it fails per-individual, so the finding indicts individual prediction specifically, not all uses of conditioning.


— "Post-training makes large language models less human-like", https://arxiv.org/abs/2605.07632

Related concepts in this collection

Concept map
14 direct connections · 88 in 2-hop network ·medium cluster Open in graph ↗

Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph

your link semantically near linked from elsewhere
Original note title

persona induction fails to improve individual-level prediction undercutting a popular human-simulation technique