Source-linked AI summary
LLM as a Mastermind: A Survey of Strategic Reasoning with Large Language Models
Yadong Zhang, Shaoguang Mao, Tao Ge, Xun Wang, Adrian de Wynter, Yan Xia, Wenshan Wu, Ting Song, Man Lan, Furu Wei
TL;DR
Strategic reasoning with LLMs concerns anticipating others’ actions and adapting decisions in dynamic multi-agent settings, but its growing and varied literature lacks a systematic review. This survey organizes applications and methods and concludes that LLMs are being applied across diverse settings. It identifies the lack of unified benchmarks as a barrier to direct method comparison.
Problem
LLM strategic-reasoning applications and methods are proliferating, but the literature lacks a systematic review that organizes their differences and connections.
Source: §1 · pp. 2–3Method
Results
The survey categorizes applications across societal simulation, economic simulation, game theory, and gaming, and describes LLMs as enhancing strategic reasoning across diverse simulations.
Source: §3 · pp. 3, 5Takeaways & Limitations
The conclusion calls for interdisciplinary collaboration to bridge theoretical advances and practical applications in strategic reasoning.
Source: §7 · p. 9The absence of unified benchmarks hinders direct comparisons among methods across the broad range of strategic-reasoning applications.
Source: §6.2 · p. 9
Abstract
from arXiv · showhide
This paper presents a comprehensive survey of the current status and opportunities for Large Language Models (LLMs) in strategic reasoning, a sophisticated form of reasoning that necessitates understanding and predicting adversary actions in multi-agent settings while adjusting strategies accordingly. Strategic reasoning is distinguished by its focus on the dynamic and uncertain nature of interactions among multi-agents, where comprehending the environment and anticipating the behavior of others is crucial. We explore the scopes, applications, methodologies, and evaluation metrics related to strategic reasoning with LLMs, highlighting the burgeoning development in this area and the interdisciplinary approaches enhancing their decision-making performance. It aims to systematize and clarify the scattered literature on this subject, providing a systematic review that underscores the importance of strategic reasoning as a critical cognitive capability and offers insights into future research directions and potential improvements.
1 INTRODUCTION
Strategic reasoning requires agents to understand dynamic multi-agent settings, anticipate others’ actions, and adapt their own decisions. The survey addresses the scattered literature on LLM strategic reasoning, whose language and contextual abilities have broadened the settings it can address.
1 INTRODUCTION
- Strategic reasoning selects actions in multi-agent settings by anticipating others’ behavior and considering how one’s own decisions influence them.Its dynamic environment and uncertain adversary actions distinguish it from other reasoning tasks; Table 1 selects representative cognitive skills rather than listing them exhaustively.Source: §1 · p. 2, Table 1 · p. 2
- LLMs extend strategic reasoning beyond digital games through dialogue-based agents, rapid contextual adaptation, and text-based reasoning that can make decisions more transparent.Earlier work relied heavily on extensive reinforcement-learning simulation in spatial action, board, and competitive video games, limiting application scope and transferability.Source: §1 · p. 2
- Despite growing applications and interdisciplinary methods, no systematic review organizes the differences and connections among LLM strategic-reasoning studies.Existing LLM reviews address agents, simulation, and game-playing, while this survey focuses on strategic reasoning as a cognitive capability.Source: §1 · pp. 2–3
- The survey covers definitions, applications, enhancement methods, evaluation, and future challenges to clarify LLMs’ strategic-reasoning capabilities and research directions.It notes that methods and application scopes diverge substantially across the literature.Source: §1 · p. 3
2 DEFINITION: WHAT IS STRATEGIC REASONING WITH LLMS
The survey defines strategic reasoning as dynamic, interactive decision-making in which agents pursue goals while anticipating and influencing others. It excludes settings without meaningful strategic complexity or adjustment to other agents.
2 DEFINITION: WHAT IS STRATEGIC REASONING WITH LLMS
- Strategic reasoning is the ability to anticipate and influence others’ actions in competitive or cooperative multi-agent settings while understanding motives and environmental causal relationships.Unlike static or single-agent problem-solving, it requires ongoing assessment of changing situations and other agents’ intentions.Source: §2 · p. 3
- Its core characteristics are goal-directed action, reciprocal interaction, prediction under limited information, and adaptation to agents’ actions and environmental change.These characteristics describe the core features listed by the survey.Source: §2 · p. 3
- The survey excludes simulations and collaboration tasks that lack obvious strategic reasoning, dynamic environmental adjustment, or feedback from collaborators.The stated scope centers on contexts requiring understanding of goals, competition, and environmental dynamics.Source: §2 · p. 3
3 SCENARIOS: WHERE TO APPLY STRATEGIC REASONING WITH LLMS
The survey groups LLM strategic-reasoning applications into societal simulation, economic simulation, game theory, and gaming. Across these settings, it describes models forecasting and adapting to multi-agent dynamics.
Societal Simulation
- Societal simulation models interactions shaped by norms and collective behavior, while frameworks assess social understanding and studies identify persistent social biases in political simulations.WarAgent illustrates using LLM-powered systems to recreate and analyze historical conflicts.Source: §3 · p. 4
Economic Simulation
- Economic simulations examine market, business, and financial decisions, including competition, hiring, auctions, and predictions of stock movements.CompeteAI models restaurant-customer competition, while AucArena studies LLM participation in auctions.Source: §3 · p. 5
Game Theory
- Game-theoretic applications study strategic interaction among rational decision-makers, including fairness and cooperation in matrix and repeated games and multi-round reasoning.Alympics and k-level reasoning are cited as approaches to multi-round strategic thinking.Source: §3 · p. 5
Gaming
- Gaming applications span board, card, conversational, and video games, where LLMs interpret mechanics, develop strategies, and adapt to opponents.The survey cites reported poker performance exceeding traditional algorithms and human-parity tactical battles in video games.Source: §3 · p. 5
4 METHODS: HOW TO IMPROVE STRATEGIC REASONING WITH LLMS
The survey organizes methods for improving LLM strategic reasoning into prompt engineering, modular agent enhancement, Theory of Mind, and imitation or reinforcement learning. The categories can overlap and complement one another.
Prompt Engineering
- Prompt engineering uses learning-based or task-specific prompts to clarify complex strategic scenarios and can draw on game histories, feedback, or reflection.The survey describes these techniques as supporting LLMs’ understanding and adaptation in strategic games and systems.Source: §4 · pp. 5–6
Module Enhancement
- Modular agents add capabilities such as memory, external knowledge retrieval, summarization, analysis, planning, and action selection.In bargaining, OG-Narrator combines a deterministic quote generator with an LLM narrator and reports tenfold profitability relative to baselines.Source: §4 · p. 6
Theory of Mind
- Theory of Mind methods use others’ mental states to guide strategic anticipation, belief tracking, and opponent-specific behavior prediction.The survey reports more precise opponent predictions with opponent-specific sessions and improved prediction accuracy from broader action histories.Source: §4 · p. 7
Imitation Learning and Reinforcement Learning
- Imitation and reinforcement learning incorporate expert chess play or use LLM reasoning as a value function, and also refine performance in dialogue interactions.The cited dialogue settings include meeting scheduling and public debates.Source: §4 · p. 7
Methods overview
- The four method categories are not fully orthogonal: Theory of Mind can use prompt engineering while drawing on game-theoretic principles.The survey presents the categories as approaches that may be integrated and complement one another.Source: Figure 3 · p. 6, §4 · p. 7
5 EVALUATIONS: HOW TO ASSESS STRATEGIC REASONING WITH LLMS
The survey frames strategic-reasoning evaluation as a combination of outcome measures, process analysis, and qualitative assessment. It argues that reasoning traces and strategic behaviors provide information that game results alone cannot capture.
5 EVALUATIONS: HOW TO ASSESS STRATEGIC REASONING WITH LLMS
- Outcome measures such as win rates, survival rates, and rewards quantify performance, while GTBench and LLMArena support comparisons across game types and difficulty levels.The survey names Normalized Relative Advantage (NRA) and TrueSkill as scoring systems used by these frameworks.Source: §5 · p. 8
- Process-oriented metrics assess how models perceive, predict, and adapt, including opponent-move analysis under incomplete information and behavior prediction from public information.MAgIC evaluates opponent-move analysis accuracy; K-Level Reasoning assesses prediction precision.Source: §5 · p. 8
- Because LLMs generate reasoning processes, their inferential steps can be analyzed to understand decision-making beyond outcome-focused evaluation.The survey contrasts this explainability opportunity with reinforcement learning methods that focus merely on outcomes.Source: §5 · p. 8
- Qualitative evaluation examines strategic behaviors such as deception, cooperation, and discernment, which can matter independently of a game's ultimate outcome.The survey gives bluffing and cooperation in Werewolf and Poker as examples.Source: §5 · p. 8
- The survey presents quantitative benchmarks and qualitative insights as complementary parts of a holistic evaluation of strategic depth and adaptability.It states that this dual approach enhances evaluation robustness and mitigates challenges in measuring cognitive processes.Source: §5 · p. 8
6 DISCUSSIONS: AN OUTLOOK ON STRATEGIC REASONING WITH LLMS
The discussion identifies open questions about whether LLM agents simulate human strategic reasoning and how their capabilities scale across task complexity. It also highlights fragmented benchmarks and the limits of relying on static-text learning or scale alone.
6.1 CAN THE LLM AGENTS REALLY SIMULATE HUMAN STRATEGIC REASONING?
- Systematic research is lacking on which LLM levels can simulate strategic tasks of varying complexity and cognitive difficulty.The discussion notes that claims of human-like capabilities have appeared in some simulations, but model sizes and configurations remain insufficiently mapped to decision-making and prediction abilities.Source: §6.1 · p. 8
6.2 BRIDGING THE DIVIDE: THE URGENT NEED FOR UNIFIED BENCHMARKS
- Unified benchmarks are absent, while customized scenarios hinder direct method comparisons and combined qualitative and quantitative evaluation complicates benchmark design.The survey calls for recognized benchmarks spanning diverse applications and suitable difficulty levels.Source: §6.2 · p. 9
6.3 STRATEGIC REASONING: CHALLENGING YET PROMISING FOR LARGE LANGUAGE MODELS
- LLMs learn patterns from static text but do not directly infer complex dynamic multi-agent interactions; strategic framing through prompts or algorithms can elicit responses reflecting strategic considerations.The discussion characterizes this as a limitation alongside an opportunity, not as evidence that such responses fully capture strategic reasoning.Source: §6.3 · p. 9
- Increasing model parameters and training data alone may not suffice for general-purpose LLMs to master intention understanding, future-action prediction, and strategy adjustment.The discussion speculates that even powerful general-purpose models may fall short of fully realizing strategic reasoning capabilities.Source: §6.3 · p. 9
7 CONCLUSION
The survey concludes that LLMs have developed significant advantages in complex decision-making across domains. It calls for interdisciplinary work and continued refinement to connect theoretical advances with practical applications.
7 CONCLUSION
- The review highlights the evolution of LLMs and their significant advantages in complex decision-making across various domains.This is the survey's concluding assessment of LLMs' role in strategic reasoning.Source: §7 · p. 9
- Future efforts should use interdisciplinary collaboration to bridge theoretical advances and practical applications in decision-making and strategy development.Source: §7 · p. 9
- The conclusion presents continued exploration and refinement of LLMs as promising for strategic decision-making and solving complex problems.It calls for researchers and practitioners to work toward realizing this potential.Source: §7 · p. 9
A.1 COGNITIVE SKILLS OF REASONING
The appendix distinguishes cognitive skills relevant to reasoning, including deduction, contextual interpretation, prediction, abstraction, and empathy. The supplied passages define each skill in terms of its constituent abilities.
A.1 COGNITIVE SKILLS OF REASONING
- Logical deduction derives conclusions from premises by applying explicit logical rules and identifying relations such as causality, equivalence, and contradiction.Source: §A.1 · p. 15
- Contextual intelligence interprets information using context, social norms, implied meanings, and subtle cues.Source: §A.1 · p. 15
- Predictive analytics forecasts future events or trends from existing information by integrating past and present data and reasoning about possible scenarios.Source: §A.1 · p. 15
- Abstract thinking generalizes, categorizes, and conceptualizes beyond direct experience to apply broad principles across situations.Source: §A.1 · p. 15
- Cognitive empathy involves understanding how others think and feel through perspective-taking, fantasy, tactical empathy, and emotion regulation.Source: §A.1 · p. 15
A.2.1 FORMULATION OF STRATEGIC REASONING ENVIRONMENT
The appendix models strategic reasoning as an environment of interacting agents, then frames action choice as maximizing expected utility given beliefs about other players.
A.2.1 FORMULATION OF STRATEGIC REASONING ENVIRONMENT
- Figure 5 situates the MAS environment that the appendix formalizes as the GAME tuple ⟨N, A, H, Z, u, I⟩.The tuple represents agents, actions, histories, terminal histories, utilities, and information states.Source: §A.2.1 · p. 16, Figure 5 · p. 16
- In LLM-based environments, actions are textual responses, while information states include observable dialogue history, own action history, and private information.For finite-action settings, an LLM response must be parsed into a legitimate action.Source: §A.2.1 · p. 16
A.2.2 TARGET
- Each agent pursues individual goals by assessing possible outcomes in light of its own objectives and other agents’ possible actions.Agents select actions considered most likely to achieve their goals.Source: §A.2.2 · p. 16
- A player’s target is to maximize expected utility under a strategy profile, with expectations based on beliefs about other players’ actions.The notation defines u_i(s) as player i’s expected utility; s_-i denotes the strategies of the other players.Source: §A.2.2 · p. 17
A.3 STRATEGIC REASONING WITH LARGE LANGUAGE MODELS VS. REINFORCEMENT LEARNING
- The text suggests future models could combine LLMs’ knowledge handling with RL’s adaptive decision-making for strategic reasoning.The proposed synergy is presented as a future possibility across diverse and complex scenarios.Source: §A.3 · p. 17
- Table 2 frames a comparison between LLMs and reinforcement learning, which the text describes as distinct but complementary approaches.LLMs are associated with coherent language and broad knowledge, while RL learns actions through trial-and-error interaction.Source: Table 2 · p. 17, §A.3 · p. 17