Can models track dynamic mental state changes better than static beliefs?
This explores whether LLMs are better at tracking how someone's mind shifts in real time (a changing belief, growing resistance) than at holding a fixed mental state steady — and the corpus says it's actually the reverse.
This explores whether LLMs handle moving mental states better than static ones, and the most direct evidence flips the premise: models are *worse* at the dynamic case. In persuasion settings, LLMs match human performance when tracking a fixed mental state — say, a persuader's unchanging goal — but fall off sharply when asked to follow a persuadee's *evolving* resistance as it shifts turn by turn Can language models track how minds change during persuasion?. So the intuition that motion is easier than stasis doesn't hold; the harder problem is keeping up with a mind that won't sit still.
Why the gap? A recurring theme in the corpus is that models reach for surface strategies instead of genuinely simulating a mind. On open-ended benchmarks like ChangeMyView and FANTOM, LLMs fail at real perspective-taking even while acing structured tasks — and the fix that works is *architectural*: bolting on explicit Bayesian belief-tracking outperforms the LLM alone Do large language models genuinely simulate mental states?. That suggests dynamic tracking fails precisely because there's no persistent internal belief state being updated — just pattern completion. The point lands harder when you learn many theory-of-mind benchmarks can be solved by pattern matching alone, with templated artifacts letting models pass without any real reasoning Can language models solve ToM benchmarks without real reasoning?. A static-belief test is exactly the kind of thing surface tricks can fake; a shifting belief is exactly what they can't.
The deeper diagnosis several notes converge on is *behaviorism*: LLM agents produce plausible outputs without modeling the reasoning underneath, which is fine for snapshots but breaks down once you need belief *change* over time. Faithful social simulation, the argument goes, requires modeling thought — belief networks and reasoning traces — not just behavior, because only an internal model supports counterfactual adaptation as circumstances move Can language models simulate belief change in people?. And even causal belief networks, the obvious tool for representing how one belief updates another, capture only part of the picture: they miss the associative, analogical, and emotion-driven shifts that actually drive how human beliefs move Can causal models alone capture how humans actually reason?.
There's a hopeful counter-thread worth knowing about. When models are *finetuned* on psychology-experiment data rather than asked to reason cold, they predict human decisions better than purpose-built cognitive models and even capture individual differences Can language models learn to model human decision making?. And dynamic *signals* are readable in principle — confidence variance can be used live to steer reasoning Can confidence patterns reveal overthinking versus underthinking?, and behavioral cues like hesitation and gaze can be instrumented as a continuous read on someone's cognitive state mid-interaction Can AI systems read cognitive state from interaction patterns alone?. So the limitation isn't that dynamic state is unreadable — it's that off-the-shelf LLMs don't maintain the updatable internal model that tracking it requires.
The thing you might not have expected: the honest answer to whether modest mental states can even be attributed to LLMs is a qualified yes — a graded view that ascribes undemanding states like beliefs and desires (while withholding consciousness) survives the deflationist objections Can we defend modest mental attributions to large language models?. The bottleneck for dynamic tracking, then, isn't philosophical permission to talk about belief — it's the missing machinery to keep a belief *updated* as the conversation moves.
Sources 9 notes
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.
Supervised fine-tuning matches reinforcement learning performance on ToM tasks, suggesting models exploit structural vulnerabilities rather than develop genuine reasoning. Distribution biases and templated artifacts allow surface-level pattern recognition to achieve competitive generalization.
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 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.
ReBalance uses confidence variance and overconfidence as diagnostic signals to apply training-free steering vectors that reduce overthinking redundancy while promoting exploration during underthinking, improving accuracy across models from 0.5B to 32B parameters.
Research shows AI systems can instrument multimodal behavioral signals (gaze, hesitation, speed) to read cognitive state during interaction, preserving flow by avoiding disruptive explicit probes. However, the same substrate enables both helpful timing and manipulative profiling.
Both robustness and etiological deflationist arguments beg the question against inflationism. A graded approach ascribing metaphysically undemanding states like beliefs and desires—while withholding consciousness claims—mirrors how we treat non-human animals.