Which game type reveals minimax reasoning in language models?
This explores how different kinds of strategic games surface distinct reasoning styles in LLMs — and which game structure is the one where minimax (worst-case-minimizing) reasoning shows up.
This explores how the *type* of game an LLM plays reveals what kind of reasoning it leans on, and specifically where minimax reasoning emerges. The most direct answer in the corpus comes from a large behavioral-game-theory study of 22 models, which found that strategic style isn't uniform — it splits into three profiles tied to game structure rather than to how 'deep' a model reasons. In that work, GPT-o1 defaulted to minimax reasoning (minimizing its worst-case loss), DeepSeek-R1 leaned on trust-based reasoning, and GPT-o3-mini used belief-anticipation about the opponent's next move. The headline finding is the relationship, not the label: minimax shows up in competitive, adversarial game types where guarding against the worst opponent move pays off, while cooperative or coordination games pull models toward trust and belief-tracking. Performance tracked the match between style and game structure — not raw reasoning depth Do large language models use one reasoning style or many?.
What makes that more interesting is a question the corpus raises about whether any of this is 'real' reasoning at all. One thread argues LLMs are semantic reasoners, not symbolic ones: strip the familiar meaning out of a task and performance collapses even when the logical rules are handed to the model directly Do large language models reason symbolically or semantically?. Minimax is, classically, a formal symbolic procedure — so a model 'doing minimax' in a game may actually be pattern-matching to adversarial framings it saw in training rather than computing a game tree. That reframes the original finding: game type might reveal which *learned association* a model reaches for, not which algorithm it runs.
That skepticism gets sharper from the work on reasoning traces, which finds that the visible chain-of-thought is closer to persuasive mimicry than a faithful record of computation — invalid logical steps perform almost as well as valid ones Do reasoning traces show how models actually think?. And the corpus has a pointed cautionary case: on constraint problems, most models look like they're reasoning correctly but are really exploiting a conservative default, and they get *worse* when the constraints are removed Are models actually reasoning about constraints or just defaulting conservatively?. Minimax — which is itself a kind of cautious, defend-the-downside posture — is exactly the style most easily faked by a conservative bias. So 'GPT-o1 uses minimax' could partly be 'GPT-o1 plays it safe,' which only looks like minimax in adversarial games.
The doorway worth walking through here: the answer to 'which game type reveals minimax' is 'competitive/adversarial games' — but the more useful takeaway is that game type is a *diagnostic instrument*. Because different structures elicit different styles, you can use a battery of game types to fingerprint a model's reasoning profile, the same way fine-tuning LLMs on psychology-experiment data turns them into general predictors of human decision-making Can language models learn to model human decision making?. The game becomes a probe for the mind behind it — for both humans and models.
Sources 5 notes
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.
When semantic content is decoupled from reasoning tasks, LLM performance collapses even with correct rules in context. Models rely on parametric commonsense and token associations rather than formal logical manipulation, constraining reasoning to training distribution semantics.
LLM reasoning traces perform as persuasive appearances rather than reliable explanations of computation. Invalid logical steps perform nearly as well as valid ones, and corrupted traces generalize comparably, showing that semantic correctness is not what produces the performance gains.
Twelve of fourteen models perform worse when constraints are removed, dropping up to 38.5 percentage points. Models appear to reason correctly by defaulting to harder options, not by actually evaluating constraints.
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.