How does reasoning instability prevent models from modeling individuals?
This explores a surprising link: the same failures that make reasoning models unstable — wandering between ideas, switching paths too early, drifting in style — are what stop them from holding a single, evolving individual in focus over time.
This explores how reasoning instability (the tendency to wander, switch paths prematurely, and drift mid-thought) undermines a model's ability to track one specific person rather than a generic average. The corpus suggests the connection is tighter than it looks: modeling an individual is fundamentally a *stability* problem, and the very things that destabilize reasoning are the things that erase the individual.
The most direct evidence is that models simply can't follow how a particular person reasons over time. When asked to track individualized reasoning styles, LLMs lean on surface lexical cues and fail to adapt as a person's strategy evolves Can models recognize how individuals reason differently?. The counterintuitive twist comes from role-playing research: piling on *more* reasoning doesn't help an LLM stay in character — it actively hurts. Large reasoning models suffer "attention diversion" and "style drift," and extending the chain of thought without guardrails degrades persona consistency rather than sharpening it Why do reasoning models lose character consistency during role-playing?. So the reasoning process itself is the destabilizing force pulling the model off the individual it's supposed to embody.
Why would thinking harder make you worse at being someone? Two well-documented instabilities explain it. Reasoning models "wander" through invalid exploration and "underthink" by abandoning promising paths too soon — failures of structural organization, not raw capability Why do reasoning models abandon promising solution paths? Do reasoning models switch between ideas too frequently?. An individual is a stable trajectory you have to commit to and hold; a wandering, path-switching process keeps re-rolling that commitment. This dovetails with the view that an LLM holds a *superposition* of possible characters that only narrows as a conversation proceeds Does an LLM commit to a single character or maintain many?. Reasoning instability re-broadens that distribution at every step, so the model never collapses onto one consistent person.
There's a deeper reason individuals are uniquely hard. A real person carries private information and a history, and models break exactly there: they look socially competent when one model puppets all sides of an interaction, but fail systematically once agents hold private knowledge the model has to infer rather than narrate Why do LLMs fail when simulating agents with private information?. An individual is, almost by definition, a novel instance — and reasoning failures track instance-level *unfamiliarity* far more than task complexity Do language models fail at reasoning due to complexity or novelty?. The specific person in front of you is the unfamiliar instance the model was never trained on.
The hopeful thread worth knowing: these are stability problems, and stability is fixable without retraining. Penalizing thought-switching at decode time improves accuracy Do reasoning models switch between ideas too frequently?, role-aware constraints restore character fidelity Why do reasoning models lose character consistency during role-playing?, and making latent reasoning *deliberately* stochastic lets a model hold uncertainty as an explicit distribution instead of thrashing between guesses Can stochastic latent reasoning help models explore multiple solutions?. The lesson hiding here is that "model the individual" and "reason stably" may be the same engineering target viewed from two directions.
Sources 8 notes
LLMs struggle to anchor reasoning in temporal gameplay and adapt to evolving strategies. GPT-4o relies on surface lexical cues while DeepSeek-R1 shows early promise, but dynamic style adaptation remains largely insufficient across all models tested.
Large reasoning models exhibit attention diversion and style drift during role-playing, but the RAR method—using role-aware constraints and contrastive learning on reasoning style—recovers character fidelity across multiple benchmarks. Simply extending reasoning without guidance actively degrades persona consistency.
Reasoning LLMs exhibit two reinforcing failures: wandering (invalid exploration) and underthinking (premature path-switching). Decoding-level interventions like thought-switching penalties improve accuracy without fine-tuning, suggesting viable solutions exist but are abandoned prematurely.
o1-like models frequently abandon reasoning paths mid-exploration, wasting tokens on incomplete approaches. A decoding-only penalty on thought-transition tokens (TIP strategy) discourages switching, improving accuracy on challenging math without model fine-tuning.
Research shows LLMs don't commit to a single character but instead maintain a probability distribution over many consistent simulacra. Each response samples from this distribution, explaining why regenerations can yield different personalities while remaining consistent with prior context.
Research shows LLMs perform well when one model controls all interlocutors but fail systematically when agents possess private information. This reveals that apparent social competence relies on grounding work that models skip in omniscient settings.
LRMs don't break at complexity thresholds but at instance-novelty boundaries. Models fit instance-based patterns rather than generalizable algorithms, so any reasoning chain succeeds if trained on similar instances, regardless of length.
GRAM replaces deterministic latent updates with stochastic sampling, enabling models to represent distributions over solutions rather than single predictions. This allows handling of ambiguous problems and multiple valid strategies that deterministic designs cannot represent.