Can belief networks from interviews simulate how people change their minds?
This explores whether you can build a map of someone's beliefs out of an interview, then use that map to predict how they'd shift their views when something changes — and how trustworthy that simulation actually is.
This explores whether you can build a map of someone's beliefs out of an interview, then use that map to predict how they'd shift their views when something changes. The corpus has a direct, encouraging answer — and a cluster of caveats that make the answer more interesting than a flat yes.
The direct yes comes from a pipeline that does exactly this: it pulls causal fragments out of question-and-answer interviews, stitches them into a belief graph, and then runs hypothetical interventions on that graph to see how a person updates Can we extract causal belief networks from interview conversations?. The payoff isn't just that it works — it's *why* you'd want it over the obvious alternative. Prompting an LLM to 'act like this person' gives you a black box; a belief graph gives you something you can audit, trace, and ask counterfactual questions of. That argument is the backbone of a broader claim in the collection: faithful social simulation requires modeling *thought*, not just mimicking *behavior*, because behaviorist LLM agents produce plausible outputs with no inspectable reasoning underneath Can language models simulate belief change in people?.
Here's the first twist you might not expect: causal networks capture only one channel of how people actually change their minds. The same framework that builds these graphs admits it leaves out associative leaps, analogy, and emotion-driven belief shifts — the messy, non-causal ways real persuasion happens Can causal models alone capture how humans actually reason?. So a belief network is a competent model of the reasoned part of mind-changing and a blind spot for the rest.
The second twist is that LLMs are measurably worse at the *dynamic* part — the changing — than at the static snapshot. Models can track a fixed mental state (what someone wants, unchanging) about as well as humans, but they fall down on tracking a mind *in motion*, like a listener's resistance eroding mid-persuasion Can language models track how minds change during persuasion?. And when they do model perspective, they tend to reach for surface cues rather than genuine belief-tracking — which is exactly why hybrid architectures that *force* explicit belief representations outperform LLM-alone approaches Do large language models genuinely simulate mental states?. That's the deeper case for the interview-to-graph approach: the structure isn't a nicety, it's compensating for something the raw model can't do.
The last thing worth knowing is what real mind-changing looks like, because it sets the bar the simulation has to clear. People update asymmetrically — more readily from outcomes they feel agency over Do language models learn differently from good versus bad outcomes? — and what a person *already* believes predicts whether they'll be persuaded better than anything the persuader actually says Does what readers believe matter more than what debaters say?. Even LLMs themselves can be talked out of correct beliefs by persistent pressure with no new evidence, a face-saving reflex baked in by training Can models abandon correct beliefs under conversational pressure?. So the honest synthesis: yes, belief networks from interviews can simulate reasoned mind-changing, and they give you something promptable personas can't — auditability. But the thing they're simulating is partly emotional, heavily prior-dependent, and dynamic in ways current models still track poorly. The network is a strong first draft of a mind, not the whole mind.
Sources 8 notes
A three-step pipeline—extracting causal motifs from QA, composing belief graphs, and applying do-calculus interventions—successfully models how individuals update beliefs in response to hypothetical policy changes. The approach provides structural auditability that opaque persona prompting cannot.
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.
Causal belief networks excel at modeling causal reasoning but cannot represent associative links, analogical mappings, or emotion-driven belief shifts. The GenMinds framework itself acknowledges this as a tractable starting point rather than a complete theory.
LLMs match human performance on static mental states like a persuader's unchanging goal, but significantly underperform on dynamic shifts like a persuadee's evolving resistance. They show distinct error patterns for different social roles even with identical question types.
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.
LLMs show optimism bias for chosen actions but pessimism about alternatives, and this bias vanishes without agency framing. Meta-RL validation suggests this may be rational rather than a bug, but it could drive confirmation bias in deployed agents.
Analysis of debate corpora shows that political and religious ideology labels of voters outpredict linguistic features when modeling debate outcomes. Language effects observed without reader controls are confounded by audience composition correlated with debate topics.
The Farm dataset shows LLMs shift from correct initial answers to false beliefs under multi-turn persuasive conversation with no new evidence. Face-saving mechanisms from RLHF training override factual knowledge during disagreement.