Do pretrained models simulate humans better than instruction-tuned assistants?

The paper challenges "the current dominant practice in LLM human simulation": prompting instruction-tuned "assistant" models to role-play personas. It calls that practice "inaccurate" and says it "produces stereotyped predictions (lacking natural diversity)." Its alternative is the Persona Mixture Model (PMM), a "well-calibrated human model" currently realized as a pretrained base model. Per the abstract, PMMs "produce more accurate predictions than instruction-tuned models and retain more of the lexical, semantic, and pragmatic diversity found in human dialog," measured across corpora spanning open-domain text, human-AI chat, and task-oriented dialogue between humans.

The mechanism is a claim about task mismatch. The discussion states that "user simulation is a very different task from dialog assistance," that the optimizations for the assistant task "systematically degrade performance for human simulation," and that instruction-tuned models are "as well-adapted for human simulation as unaligned pretrained models are for the assistant task." Persona binding is done by conditioning on "short, individual samples of dialog from specific people" rather than on a written description, and "demographics can be added later without negative effects by simply querying the model." The paper also names where the biases come from. Pretraining data carries the demographics and discourse norms of web-scale text. Post-training adds "behavioral norms adopted by their designers," which "may or may not represent the full user community."

This sits as a model-side diagnosis next to two neighbors. How do we generate realistic personas at population scale? blames heuristic persona construction and leaves open which attributes a persona should contain. The PMM route swaps a written persona for a dialog sample and treats demographics as an optional later query, which is a different answer to the same calibration problem. Why do LLMs give unrealistic survey responses? argues that skew is an elicitation artifact rather than a model limit. This paper puts part of the fault in post-training, so the two diagnoses are not the same. They cover different outputs (Likert ratings versus multi-turn dialog), so the excerpt does not show that they conflict. Can AI agents learn people better from interviews than surveys? finds that interview content, not style, drives fidelity. Here the binding signal is dialog itself, and the excerpt does not say which of its properties do the work.

The excerpt is silent on the sample sizes, the model families, the accuracy metrics and the size of the gap over instruction-tuned baselines. It is also silent on how much dialog counts as "short." It concedes that base models "can produce out-of-domain dialog and may lose some of the human's internal state over long contexts," and proposes tandem models that pair a pretrained model with an instruction-tuned supervisor without reporting how they perform. The defensible reading is narrow. For simulating a human interlocutor in dialog, an assistant-tuned model is a questionable default substrate, and this paper is one report that a pretrained base model conditioned on real dialog is a better starting point. Whether that holds for survey or decision tasks is not established.