Can extended reasoning training capture individual strategic thinking styles?
This explores whether training models to reason at length (RL, extended thinking, chain-of-thought) can make them adopt or track a *particular person's* strategic style — not just reason well in general, but reason like you.
This explores whether longer, trained reasoning can capture individual strategic thinking styles — and the corpus's clearest answer is that the two are almost orthogonal: training shapes how *the model* reasons, but does little to anchor reasoning to *a specific person's* evolving strategy. The most direct evidence is a study finding that models fail to track individualized reasoning styles over time — GPT-4o leans on surface lexical cues rather than the underlying strategy, and even the stronger reasoners can't adapt as a player's approach shifts across a game Can models recognize how individuals reason differently?. Extended reasoning doesn't fix this, because the gap isn't reasoning depth; it's the ability to model another agent's changing strategy.
There's a twist worth sitting with: models *do* have strategic styles — they just have their own, not yours. Across 22 LLMs in behavioral game theory, distinct fixed profiles emerge: one model defaults to minimax, another to trust-based reasoning, another to anticipating what the opponent believes — and performance tracks the game's structure, not raw reasoning depth Do large language models use one reasoning style or many?. So strategic style is real and measurable in these systems, but it seems to be a property baked in by pretraining and post-training, not a flexible costume the model puts on to match an individual.
Why would training select a style rather than acquire arbitrary new ones? Several notes converge on the idea that post-training *elicits* capability already latent in the base model rather than creating it — RL, critique tuning, decoding tricks, and feature steering all surface reasoning that was already there Do base models already contain hidden reasoning ability?. And what reasoning generalizes draws on broad, transferable procedural knowledge absorbed during pretraining, not narrow memorized facts Does procedural knowledge drive reasoning more than factual retrieval?. If training is selecting from a pretrained repertoire, capturing one person's idiosyncratic strategy — which lives outside that repertoire — is a poor fit for the mechanism.
Extended reasoning also turns out to be a blunt instrument even on its own terms. More thinking tokens don't monotonically help — accuracy can peak and then decline as models overthink easy problems Does more thinking time always improve reasoning accuracy? — and what training actually changes is the *character* of thinking, redirecting it from counterproductive self-doubt into productive gap analysis Does extended thinking help or hurt model reasoning?. That's training tuning the model's own cognitive habits, not importing someone else's. The reasoning style being shaped is the model's, full stop.
The interesting flip for a curious reader: the bottleneck to capturing individual style probably isn't on the training side at all. Reasoning quality degrades sharply with longer inputs well before the context window fills Does reasoning ability actually degrade with longer inputs?, and chain-of-thought breaks down predictably once it leaves its training distribution, producing fluent-but-hollow reasoning Does chain-of-thought reasoning actually generalize beyond training data?. An individual's strategy is exactly the kind of out-of-distribution, history-dependent signal that these failure modes punish. So the corpus suggests the live question isn't "train harder/longer" but "how do you represent and feed an evolving individual strategy into a system whose reasoning is both pretrained-bounded and length-fragile?"
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.
Analysis of 22 LLMs across behavioral game theory reveals three dominant profiles: GPT-o1 uses minimax reasoning, DeepSeek-R1 uses trust-based reasoning, and GPT-o3-mini uses belief-anticipation. Performance correlates with game structure, not raw reasoning depth.
Five independent mechanisms—RL steering, critique fine-tuning, decoding changes, SAE feature steering, and RLVR—all elicit reasoning already present in base model activations. Post-training selects rather than creates reasoning; the bottleneck is elicitation, not capability acquisition.
Analysis of 5 million pretraining documents shows reasoning relies on broad, transferable procedural knowledge from diverse sources, unlike factual recall which depends on narrow, document-specific memorization of target facts.
Increasing thinking tokens from ~1,100 to ~16K reduced benchmark accuracy from 87.3% to 70.3%, revealing a non-monotonic relationship where models overthink easy problems and underthink hard ones.
Vanilla models use thinking mode counterproductively, inducing self-doubt that degrades performance. RL training reverses this, transforming the same mechanism into beneficial gap analysis. Training mediates reasoning quality, not just quantity.
FLenQA shows reasoning accuracy drops from 92% to 68% at just 3000 tokens of padding, far below context window capacity. The degradation is task-agnostic, uncorrelated with language modeling performance, and persists even with chain-of-thought prompting.
DataAlchemy experiments show CoT fails systematically under distributional shifts in task, length, and format. Models produce fluent but logically inconsistent reasoning — imitating reasoning form without valid underlying logic.