LLM Strategic Reasoning: Agentic Study through Behavioral Game Theory

🕒 Published (v1): 2025-01-01 · Source: NeurIPS · Venue: NeurIPS 2025 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

This paper introduces a behavioral game-theoretic evaluation framework—grounded in Truncated Quantal Response Equilibrium (TQRE)—that quantifies LLM strategic reasoning depth (τ) beyond Nash Equilibrium benchmarks. Applied to 22 LLMs across 13 matrix games, it finds that reasoning-focused models (GPT-o3-mini, GPT-o1, DeepSeek-R1) lead on depth, while demographic persona injection reveals systematic bias even in top performers.

Problem

Existing LLM strategic reasoning benchmarks predominantly use Nash Equilibrium (NE) as the evaluation criterion, which assumes perfect rationality, focuses on outcomes rather than decision processes, and cannot explain why models deviate from optimal play. NE also fails to capture bounded rationality and stochastic behavior intrinsic to LLMs, making it an incomplete and circular metric for assessing their reasoning quality.

Method

The framework models LLM behavior using TQRE, which combines Quantal Response Equilibrium (probabilistic choice under utility noise) with Cognitive Hierarchy (bounded recursive reasoning levels). Each agent draws a reasoning level k ~ Poisson(τ), forms beliefs over lower-level opponents' mixed strategies, computes expected utilities, and selects actions via a logit rule scaled by λₖ = γᵏ. The parameters τ (reasoning depth) and γ (decision precision) are estimated via maximum likelihood over 30 independent trials per model-game pair. The game library covers 13 matrix games across 7 types: competitive, cooperative, mixed-motive, sequential, Bayesian coordination, and signaling games, spanning complete and incomplete information. Demographic persona experiments embed 10 socio-demographic attribute groups into prompts and measure τ shifts via OLS regression.

Key Contributions

• A TQRE-based evaluation framework that measures bounded strategic reasoning depth (τ) empirically without assuming NE play, applied to 22 SOTA LLMs across 13 games
• Qualitative reasoning-chain analysis linking each top model's dominant strategy style (maximin for GPT-o1, belief-guided for GPT-o3-mini, opponent-oriented for DeepSeek-R1) to game-type performance
• Empirical finding that longer CoT does not consistently improve strategic decisions; concise chains correlate with higher confidence and better outcomes
• Demographic persona analysis showing that identity-induced τ shifts expose implicit biases (e.g., female persona boosts reasoning depth in GPT-4o and Claude-3-Opus; minority sexuality personas reduce τ in Gemini 2.0 and Granite MoE)

Results

• GPT-o3-mini leads in 7 games, GPT-o1 in 6, DeepSeek-R1 in 5; these three substantially outperform all other evaluated models on τ
• Notable τ values (from Table 3): GPT-o3-mini reaches τ = 4.226 on cooperative and Bayesian games; GPT-o1 reaches τ = 4.741 on competitive (high-stake); DeepSeek-R1 reaches τ = 4.226 on Bayesian and τ = 3.944 on mixed-motive
• GPT-4o, DeepSeek-V3, and DeepSeek-V2.5 perform substantially below the top tier on most games
• Smaller models (Gemma-2-27B, LLaMA-3.1-8B) occasionally match larger models on specific game types
• Six LLMs exceed human-level τ on the SW10 benchmark game
• CoT prompting shows mixed effects: Claude-3-Opus improves via self-correction; GPT-4o degrades due to a flawed initial premise that CoT amplifies
• Female persona framing significantly increases τ in GPT-4o, InternLM V2, and Claude-3-Opus; homosexual/bisexual personas decrease τ in Gemini 2.0 and Granite MoE

Limitations

• Causal link between prompt content and decision outcomes remains underexplored; only correlational OLS regression is used for demographic effects
• Game library uses abstract normal-form (matrix) games; ecological validity to real multi-turn agent interactions is unverified
• Framework is semi-dynamic: does not capture real-time sequential multi-agent interaction where strategies evolve over repeated play
• Reasoning-chain analysis is qualitative and limited to three top models in three game types
• Model selection is broad but not exhaustive; some newer models post-dating the study are absent
• TQRE parameter estimation assumes the model's strategy distribution is stationary across 30 trials, which may not hold

Relevance to Agentic AI / LLM Agents

This work directly addresses a core gap for agentic LLM deployment: how to rigorously measure and interpret strategic decision-making in interactive, multi-agent environments where outcomes depend on anticipating others' behavior. The TQRE-based τ metric provides a principled, model-internal measure of recursive belief reasoning—directly relevant to designing agents for negotiation, bidding, coordination, and adversarial tasks. The finding that reasoning style (maximin vs. belief-based) determines game-type performance has practical implications for agent selection and prompt engineering in specific deployment contexts. The demographic bias analysis is critical for multi-agent systems serving diverse populations, where identity-conditioned strategy shifts could propagate inequity at scale.