Do large language models use one reasoning style or many?

The "LLM Strategic Reasoning" paper moves beyond standard NE-based evaluation to apply behavioral game theory across 22 LLMs in diverse strategic scenarios. The core finding: strategic reasoning is not a single capability but a set of distinct reasoning styles, and different models excel through different styles.

Three dominant profiles emerge from thinking chain analysis:

• GPT-o1: minimax reasoning. Consistently evaluates options by worst-case outcome. Explicitly states "I will now use minimax" in nearly every chain. Strong in competitive games where minimizing losses aligns with optimal strategy. But becomes overly cautious in cooperative or mixed-motive settings, sometimes assuming the opponent intends to minimize o1's payoff — interpreting cooperation as adversarial.

• DeepSeek-R1: trust-based reasoning. Begins with assumptions about opponent's likely action based on self-interest alignment. Works well in cooperative games where incentives are aligned. Exhibits "strategic trust" — assumes opponents won't deviate just to cause harm. But lacks adversarial caution for competitive settings.

• GPT-o3-mini: belief-based anticipation. Attempts to infer the opponent's likely move and respond accordingly. Performs well across cooperative and mixed-motive settings but falls back to worst-case logic under uncertainty. The most balanced profile.

Token length inversely correlates with performance. Leaders produce the shortest CoT within their strongest games. Longer reasoning chains signal hesitation and uncertainty, not deeper insight. DeepSeek-R1 in competitive games exhibits "repeated self-doubt in its CoT" that creates redundant reasoning loops inflating tokens without improvement. This independently confirms Why do correct reasoning traces contain fewer tokens? in a completely different domain.

Persona framing shifts reasoning depth. When prompted with demographic personas, some models show measurable changes: female personas increase reasoning depth in GPT-4o, Claude-3-Opus, and InternLM V2, while minority sexuality personas diminish reasoning in Gemini 2.0. The mechanism likely operates through training-corpus statistical associations modulated by RLHF.

The game-type dependence of reasoning profiles extends When does explicit reasoning actually help model performance? by adding strategic interaction as a third domain where task structure determines reasoning effectiveness.

Enrichment (2026-02-22, from Arxiv/Personas Personality): The MBTI-in-Thoughts framework adds personality priming as a strong behavioral variable in strategic games. Thinking-primed agents defect in ~90% of Prisoner's Dilemma rounds vs ~50% for Feeling types. Introverted agents show higher truthfulness (0.54 vs 0.33 for Extraverts) and produce longer, more deliberate rationales. Thinking types switch strategies infrequently (0.07) while Feeling types switch nearly twice as often (0.16). These personality-induced behavioral divergences are statistically significant and align with established MBTI theory, suggesting that game-specific reasoning profiles interact with personality-priming effects — both the game structure AND the agent's personality conditioning shape strategic behavior.

Source: Reasoning Logic Internal Rules