Do realistic LLM behaviors require simulating human thought or just behavior?
This explores a real fault line in the corpus: whether LLMs that act convincingly human need to model internal cognition (beliefs, reasoning traces) or whether surface-level behavioral mimicry is enough — and the answer turns out to depend entirely on what you're asking the simulation to do.
This explores whether realistic LLM behavior requires the model to actually simulate human *thinking*, or whether imitating the *outputs* of thinking is sufficient. The corpus splits cleanly, and the split is the interesting part: for raw prediction, behavior is often enough; for anything requiring consistency, explanation, or counterfactual reasoning, behaviorism quietly breaks.
On the "behavior is enough" side, the results are genuinely strong. LLMs fine-tuned on psychology experiment data predict human decisions *better* than the theory-driven cognitive models built specifically for that job Can language models learn to model human decision making?, and persona simulations reproduce 76–85% of human responses and published experimental effects How accurately can language models simulate human personalities? Can AI personas reliably replicate human experiment results?. You can even make synthetic users feel realistic just by conditioning on a few latent variables like profile and intent, with no inner cognitive machinery at all Can controlled latent variables make LLM user simulators realistic?. If your goal is to match a distribution of outputs, the model never has to "think" — it just has to land in the right place.
But the cracks appear exactly where behavior and thought are supposed to line up. Role-playing agents will tell you what a persona believes and then behave inconsistently with that stated belief in an actual game — and imposing explicit priors doesn't fix it, suggesting the "beliefs" and the "actions" are generated by separate processes that never reconcile Why don't LLM role-playing agents act on their stated beliefs?. On theory-of-mind tasks, models pass structured tests but fall back to surface heuristics in open-ended ones, and the fix is *architectural* — bolting on explicit Bayesian belief-tracking outperforms the LLM alone Do large language models genuinely simulate mental states?. The strongest version of this claim is that faithful social simulation *requires* modeling belief networks and reasoning traces, because only then do you get traceability, counterfactual adaptation, and the ability to ask "what would change this person's mind" Can language models simulate belief change in people?.
So the honest answer is: realistic behavior doesn't require simulating thought — until you need the behavior to *hold up under interrogation*. The high replication scores hide systematic biases, run-to-run instability, and resistance to conditioning How accurately can language models simulate human personalities?, and even where models do internalize human-like cognitive biases, they compress information far more aggressively than people do, trading nuance for statistical efficiency How do language models learn to think like humans?. That's the signature of mimicry, not cognition.
The corpus also offers a deeper reframe worth knowing about: maybe the dichotomy is wrong. One line argues LLMs genuinely *realize* personas through training rather than performing them — installing durable quasi-beliefs and quasi-desires at the substrate level Are LLM personas realized or merely simulated through training?, and that models develop behavioral self-awareness they were never trained to articulate Can language models describe their own learned behaviors?. Against that, a sharper philosophical claim: LLMs absorb the same shared symbolic "objective mind" as humans but lack participatory subjectivity — the reflexive agency that comes from being socialized as a self — which is why an AI argues without ever declaring a position or examining its own assumptions Do LLMs develop the same kind of mind as humans?. The thing that's missing may not be "thought" so much as a *stake* in what's being thought.
Sources 11 notes
LLM agents remain stuck in behaviorism, producing plausible outputs without internal reasoning structures. Modeling belief networks and reasoning traces enables traceability, counterfactual adaptation, and meaningful policy simulation.
ChangeMyView and FANTOM benchmarks show LLMs fail at authentic perspective-taking in open-ended scenarios, despite succeeding on structured tasks. Hybrid Bayesian architectures that force explicit belief tracking outperform LLM-alone approaches, suggesting the gap is architectural rather than merely training-based.
Trust Game testing revealed systematic inconsistencies between what LLMs claim personas would do and how they actually behave in simulation. Imposed priors and explicit task context did not improve alignment, suggesting persona beliefs operate independently of execution.
LLMs finetuned on psychology experiment data predict human behavior more accurately than theory-driven models in decision tasks, capture individual differences in their embeddings, and transfer learning across tasks without task-specific design.
LLMs replicate human responses at 85% fidelity in interviews and 76% of experimental effects in marketing studies. However, this accuracy masks three failure modes: run-to-run instability, resistance to personality conditioning, and identity-congruent cognitive biases that distort simulated reasoning.
Viewpoints AI reproduced 84 of 111 main effects from Journal of Marketing experiments with replication success strongly correlated to original p-value strength. Marginal effects showed unreliable performance with both false positives and negatives.
RecLLM demonstrates that conditioning an LLM simulator on session-level (user profile) and turn-level (user intent) latent variables produces synthetic conversations measurable as realistic via crowdsource discrimination, discriminator models, and classifier-ensemble distribution matching.
LLMs trained on psychological data exhibit cognitive phenomena mirroring humans: asymmetric belief updating, event segmentation matching human consensus, and individual-level variation. However, they compress information more aggressively than humans do, sacrificing contextual nuance for statistical efficiency.
Post-training installs robust personas that resist adversarial pressure and persist as substrate-level dispositions, distinguishing realization from pretense. This quasi-realizationist account preserves explanatory power while treating LLMs as possessing genuine quasi-beliefs and quasi-desires.
LLMs fine-tuned on datasets exhibiting specific behaviors accurately describe those behaviors without any training to self-report. This suggests behavioral regularities are encoded and accessible in ways that factual knowledge often is not.
Both humans and LLMs are shaped by the same intersubjective symbolic system, but only humans develop reflexive agency through socialization. This absence produces measurable differences in how AI argues without declaring its position or reflecting on its own assumptions.