← Back to Home

LLM Multi-Agent Systems: Problem Scenarios, Benchmarks, Core Capabilities, and Capability Construction

Date: August 20, 2026  |  Estimated Reading Time: 17 min  |  Author: Meng Fang

Overview

This report organises LLM-based multi-agent systems along a single conceptual chain: problem scenario → benchmark/environment → required multi-agent capabilities → capability construction. A complementary game-theoretic lens cuts across this chain by characterising the interaction structures underlying the representative scenarios.

The problem taxonomy focuses on settings where multiple agents are intrinsic to the task or provide a clear structural advantage beyond simply increasing inference-time computation. The game-theoretic taxonomy groups these settings into broad game families, while the capability taxonomy abstracts the recurring coordination requirements that arise across scenarios and interaction structures.

Problem Scenarios

Table 1. Representative problem scenarios for LLM-based multi-agent systems.

Problem Scenario Problem Description Why Multiple Agents?
Collaborative Formal Proof Search and Verification Formal proof decomposition, parallel lemma and proof search, checking of intermediate steps, and repair of failed proof attempts. Parallel exploration can cover a large proof-search space, while independent prover, critic, and verifier roles can reduce correlated reasoning and formalisation errors.
Large-Scale Software Development Architecture design, multi-module implementation, testing, debugging, code review, integration, and repository-level maintenance. Large software systems contain modular but interdependent artefacts. Specialised agents can work concurrently on different modules and roles while providing independent testing, review, and integration.
Parallel Deep Research and Evidence Synthesis Parallel information search, literature analysis, evidence extraction, fact verification, and synthesis of findings into coherent reports or decisions. Information gathering is naturally parallelisable, while independent search paths improve coverage and source diversity. Multiple agents can also cross-check conflicting evidence before synthesis.
End-to-End Scientific Discovery Literature exploration, hypothesis generation, experimental design, implementation, data analysis, interpretation, and scientific critique. Scientific discovery combines heterogeneous stages and expertise. Multiple agents enable parallel hypothesis exploration, specialised scientific roles, independent validation, and iterative feedback between experimentation and reasoning.
Multidisciplinary Clinical Diagnosis and Treatment Planning Multiple specialists analyse heterogeneous clinical evidence, formulate diagnoses, assess risks, and jointly develop or evaluate treatment plans for complex cases. Complex clinical decisions require complementary expertise and evidence from different modalities or specialties. Independent assessments can be compared and reconciled before reaching a joint decision.
Multi-Party Business Negotiation and Contracting Multiple organisations or stakeholders negotiate prices, quantities, delivery conditions, service levels, contractual terms, and other business agreements. Multiple decision makers are intrinsic to the problem: parties have distinct objectives, constraints, preferences, and potentially private information.
Supply-Chain Planning and Logistics Coordination Suppliers, manufacturers, warehouses, carriers, and retailers coordinate procurement, production, inventory, transportation, and delivery decisions. Information, resources, and control are distributed across different actors, while local plans are coupled through shared resources, temporal dependencies, and operational constraints.
Embodied Team Tasks with Complementary Capabilities Multiple embodied agents or robots cooperate on navigation, exploration, construction, manipulation, transportation, rescue, or related team tasks. Agents may differ in location, perception, tools, embodiment, or action capabilities. Successful execution therefore requires distributed sensing, complementary actions, and spatial or temporal coordination.
Long-Horizon Enterprise Workflow Automation Multi-stage business processes spanning heterogeneous tools, applications, databases, documents, and organisational functions. Enterprise workflows contain heterogeneous subtasks, long-range dependencies, specialised tool access, and failure-recovery requirements that favour specialised agents and dynamic hand-offs.
Population-Scale Social and Institutional Simulation Large populations of language agents represent individuals, organisations, or institutions and interact repeatedly to study norms, markets, information diffusion, policy effects, and other collective phenomena. Multiple interacting autonomous actors are intrinsic to the phenomenon being studied. Population-level outcomes emerge from heterogeneous local behaviours and repeated interactions.

A Game-Theoretic Lens on Multi-Agent Interaction

The application scenarios above are not themselves uniquely defined games. A formal game additionally requires a specification of the players, actions, utilities, information, timing, and state transitions. Nevertheless, their dominant interaction structures can be organised into three broad families: common-payoff team games, general-sum strategic games, and repeated population and institutional games. This lens complements rather than replaces the application taxonomy, and connects LLM-based multi-agent systems to the broader Cooperative AI agenda spanning multi-agent systems, game theory, and social choice [34].

Table 2. Core game families underlying representative LLM multi-agent scenarios.

Core Game Family Problem Scenario Canonical Formulation Important Variants Defining Interaction Structure
Common-Payoff
Team Games		 Collaborative Formal Proof Search and Verification Common-payoff stochastic team problem; under decentralised partial observability, a Dec-POMDP[35] Designed proposer–critic–verifier subgame Agents share the objective of producing a valid proof while distributing search, checking, and repair.
Large-Scale Software Development Common-payoff stochastic team game Assignment, matching, or coalition-formation game for task allocation Agents share a repository-level objective but control interdependent modules, tools, and development stages.
Parallel Deep Research and Evidence Synthesis Common-payoff Bayesian team game[36] Information-acquisition and aggregation game; Dec-POMDP when search histories remain local Agents share the research objective but acquire different evidence, signals, and source coverage.
End-to-End Scientific Discovery Common-payoff Bayesian stochastic game[36, 37] Designed adversarial subgame among hypothesis generator, experimenter, critic, and verifier Agents share a discovery objective while reasoning under uncertainty about hypotheses, evidence, and experimental outcomes.
Multidisciplinary Clinical Diagnosis and Treatment Planning Bayesian team game under a shared patient-level objective[36] Voting or collective-choice game; general-sum formulation when stakeholder objectives diverge Specialists hold heterogeneous evidence and uncertain beliefs that must be aggregated into a collective decision.
Embodied Team Tasks with Complementary Capabilities Cooperative Markov game; under decentralised partial observability, a Dec-POMDP[35, 37] Coordination, assignment, or potential game Agents share a team reward but possess local observations and complementary action capabilities.
Long-Horizon Enterprise Workflow Automation Common-payoff stochastic team game or extensive-form coordination game Assignment or matching game for routing and hand-offs Agents share a workflow-level objective but operate through different tools, permissions, and dependent stages.
General-Sum
Strategic Games		 Multi-Party Business Negotiation and Contracting General-sum Bayesian extensive-form bargaining game[36, 38] Repeated bargaining, auctions, coalition formation, or multilateral bargaining Parties benefit from agreement but have private valuations, outside options, constraints, and preferred outcomes.
Supply-Chain Planning and Logistics Coordination General-sum stochastic game across independent organisations[37] Stackelberg, auction[39], inventory, congestion, or coalitional game Firms’ decisions are dynamically coupled, while costs, capacities, inventories, information, and incentives may differ.
Repeated Population
and Institutional Games		 Population-Scale Social and Institutional Simulation Repeated network game or evolutionary game[40] Public-goods, coordination, congestion, voting, market, diffusion, or collective-action games Heterogeneous agents interact repeatedly through social, economic, or institutional networks, producing emergent population-level outcomes.

Note: The table assigns each scenario to its dominant game family rather than claiming a unique formalisation. Classification depends on the system boundary and the definition of utilities. Supply-chain coordination, for example, may be common-payoff within a vertically integrated organisation but general-sum across independent firms. Clinical decision making may similarly move from a Bayesian team problem to a general-sum collective-choice problem when stakeholder objectives diverge. A Dec-POMDP is more precisely a decentralised common-payoff sequential decision model than a conventional non-cooperative game.

The first family accounts for most current LLM multi-agent applications. In these systems, agents usually share a global objective, and the central difficulty is coordinating distributed information, expertise, tools, and actions rather than resolving fundamentally opposed interests. Negotiation and inter-organisational supply chains instead introduce general-sum strategic interaction, where cooperation coexists with private objectives and incentives. Population-scale simulations form a third family because their primary object of study is the long-run collective dynamics generated by repeated interaction among heterogeneous agents.

A cross-cutting pattern is the use of designed internal games within otherwise common-payoff systems. Proposer–critic, solver–verifier, advocate–judge, and implementer–tester structures deliberately introduce local opposition to improve a shared global outcome. These are best understood not as a fourth application family, but as interaction mechanisms embedded within one of the three broader game families. In this sense, current multi-agent scaffolding often designs not only agent roles and communication, but also the local game in which the agents interact.

Benchmarks and Evaluation Environments

Table 3. Representative benchmarks and environments for the problem scenarios above. “Yes” denotes a dedicated MAS benchmark, “Partial” denotes partial scenario coverage, and “No” indicates reliance primarily on general-agent benchmarks or application-specific testbeds.

Problem Scenario Dedicated MAS Benchmark? Benchmark / Environment Note
Collaborative Formal Proof Search and VerificationNominiF2F[1]; ProofNet[2]Strong formal-proving benchmarks exist, but they are not multi-agent benchmarks and do not explicitly evaluate collaborative proof search, communication, or independent verification.
Large-Scale Software DevelopmentPartialMultiAgentBench--Coding[3]; SWE-bench[4]MultiAgentBench explicitly evaluates multi-agent coding coordination, while SWE-bench provides realistic repository-level tasks. Neither fully captures large-scale multi-module development and integration.
Parallel Deep Research and Evidence SynthesisYesMultiAgentBench--Research[3]; IDRBench[5]; WideSearch[32]; BrowseComp[33]MultiAgentBench directly evaluates multi-agent research collaboration, while IDRBench targets interactive deep research. WideSearch evaluates broad, large-scale information collection and completeness, making it particularly suitable for parallel multi-agent search. BrowseComp evaluates persistent and strategic deep web search for hard-to-find information. WideSearch and BrowseComp are general agentic-search benchmarks rather than dedicated MAS benchmarks, but provide useful substrates for evaluating multi-agent research systems.
End-to-End Scientific DiscoveryNoPhySciBench[6]PhySciBench[6] evaluates physical-science research tasks and is used with a multi-agent research framework, but does not benchmark a complete multi-agent hypothesis--experiment--analysis--critique cycle.
Multidisciplinary Clinical Diagnosis and Treatment PlanningPartialAI Hospital / MVME[7]A native multi-agent clinical interaction benchmark covers symptom collection, examination, and diagnosis, but not yet a full multidisciplinary specialist and treatment-planning setting.
Multi-Party Business Negotiation and ContractingYesAgenticPay[8]Explicit multi-agent buyer--seller negotiation benchmark with private constraints and valuations, from bilateral bargaining to many-to-many markets.
Supply-Chain Planning and Logistics CoordinationPartialMarket-Bench[9]Native multi-agent supply-chain market environment covering procurement auctions and retail competition, but not full production--inventory--transport logistics coordination.
Embodied Team Tasks with Complementary CapabilitiesYesTeamCraft[10]; MECoBench[11]Both explicitly benchmark embodied multi-agent collaboration, including communication, heterogeneous agents, and joint task completion.
Long-Horizon Enterprise Workflow AutomationNoWorkArena[12]; OSWorld[13]Realistic enterprise and computer-use environments exist, but they are primarily designed for individual agents and do not directly assess inter-agent workflow coordination.
Population-Scale Social and Institutional SimulationNo*AgentSociety[14]Large-scale multi-agent simulation platforms exist, but there is not yet a broadly adopted standardised benchmark for population-scale institutional simulation and coordination.

Core Multi-Agent Capabilities

Table 4. Core coordination capabilities abstracted from the problem scenarios.

IDCore CapabilityCore Question
C1Task Decomposition and Dependency StructuringWhat needs to be done, and how are subtasks related?
C2Role, Task, and Resource AllocationWho should do what, with which model, tool, or resource?
C3Communication and Information SharingWhat information should be exchanged, with whom, and when?
C4State, Memory, and Belief ManagementWhat does the system or each agent know, remember, and believe?
C5Joint Planning, Coordination, and SynchronisationHow should interdependent actions fit together?
C6Monitoring, Adaptation, and ReplanningIs execution proceeding correctly, and what should change when it is not?
C7Aggregation, Verification, and Conflict ResolutionHow are outputs checked, reconciled, and combined?
C8Partner, Intent, and Strategic ModellingWhat are other agents capable of, trying to do, or likely to do next?

Mapping Problem Scenarios to Capabilities

Table 5. Capability matrix linking problem scenarios to their primary multi-agent capability requirements.

Problem Scenario C1 C2 C3 C4 C5 C6 C7 C8
Collaborative Formal Proof Search and Verification●●●●●
Large-Scale Software Development●●●●●●
Parallel Deep Research and Evidence Synthesis●●●●●
End-to-End Scientific Discovery●●●●●●
Multidisciplinary Clinical Diagnosis and Treatment Planning●●●●
Multi-Party Business Negotiation and Contracting●●●●●
Supply-Chain Planning and Logistics Coordination●●●●●
Embodied Team Tasks with Complementary Capabilities●●●●●
Long-Horizon Enterprise Workflow Automation●●●●●●
Population-Scale Social and Institutional Simulation●●●●
C1 — Task Decomposition and Dependency Structuring C2 — Role, Task, and Resource Allocation C3 — Communication and Information Sharing C4 — State, Memory, and Belief Management C5 — Joint Planning, Coordination, and Synchronisation C6 — Monitoring, Adaptation, and Replanning C7 — Aggregation, Verification, and Conflict Resolution C8 — Partner, Intent, and Strategic Modelling

Note: ● indicates a primary capability requirement. An empty cell does not imply that the capability is irrelevant to the scenario.

Across these scenarios, allocation, communication, memory, adaptation, and verification recur most widely, whereas partner and intent modelling is especially central to settings involving autonomous actors with distinct objectives or persistent social identities. The matrix captures primary rather than exhaustive requirements: unmarked capabilities may still contribute to system performance, but are not treated here as defining coordination requirements.

Building Multi-Agent Capabilities

We use capability construction broadly to include mechanisms that instantiate or improve a coordination capability. These mechanisms need not follow a single data → environment → training pipeline. Depending on the capability, construction may rely on architectural scaffolding, prompting and protocol design, synthetic-data post-training, learned controllers, reinforcement learning or self-play, external memory or solver modules, and test-time search or verification.

Table 6. Representative approaches for building core multi-agent capabilities.

Capability Approaches for Building the Capability Representative Methods / Papers Current Status / Gap
C1. Task Decomposition and Dependency Structuring Planner/orchestrator scaffolding that explicitly produces subtasks or dependency graphs; synthetic compositional task generation + SFT; process supervision on decomposition; search or RL using downstream execution quality. TaskCraft[15]; BUTTON[16]; MagicAgent[17]; SMART-LLM[18]; Magentic-One[19] Synthetic task construction is emerging, but much current MAS decomposition is still obtained through prompting or orchestration rather than learned as a dedicated capability.
C2. Role, Task, and Resource Allocation Rule- or solver-based assignment; capability profiling; auction or market mechanisms; learned routers/controllers; cost-aware RL; dynamic reassignment from execution feedback. MasRouter[20]; SMART-LLM[18]; COHERENT[21] Allocation is increasingly learnable, but most systems still assume manually specified roles or agent capability profiles.
C3. Communication and Information Sharing Fixed communication protocols and topologies; shared workspace or blackboard; selective message generation; SFT/DPO on interaction traces; communication-cost regularisation; topology optimisation; emergent communication via RL/MARL. Optima[22]; communication-structure optimisation methods; language-guided emergent communication[23] Architectural and learned approaches both exist, but scalable selective communication—deciding what to send, to whom, and when—remains difficult.
C4. State, Memory, and Belief Management External or shared memory; episodic retrieval and summarisation; structured world/belief state; Theory-of-Mind prompting; supervised hidden-state or mental-state prediction; recurrent and long-horizon learning. SOTOPIA-π[24]; ToMAgent[25]; LLM-Coordination[26]; shared-memory and blackboard architectures Mostly implemented through external memory and prompting. Dedicated training for persistent distributed beliefs remains relatively underdeveloped.
C5. Joint Planning, Coordination, and Synchronisation Centralised planning; distributed plan proposal and merging; hierarchical planning; classical planner/LLM hybrids; execution-feedback loops; expert-plan SFT; team-reward RL/MARL. SMART-LLM[18]; COHERENT[21]; LLM-Coordination[26]; hierarchical LLM + symbolic-planner approaches Planning is well represented architecturally, but relatively little work directly post-trains LLMs for robust joint planning and synchronisation.
C6. Monitoring, Adaptation, and Replanning Progress ledgers or watchdogs; critic or failure detectors; execution-feedback loops; failure and perturbation injection; recovery demonstrations; adaptive orchestration; RL under dynamic perturbations. Magentic-One[19]; COHERENT[21]; ToolMaze[27]; Gaia2[28] Strong system-level mechanisms and new failure-focused environments exist, but capability-specific post-training remains sparse.
C7. Aggregation, Verification, and Conflict Resolution Majority voting and ensembling; learned judges and verifiers; critic–revision loops; multi-agent debate; formal or executable verification; synthetic critique data + SFT; preference learning; test-time compute scaling. DebateCV / Debate-SFT[29]; Multi-Agent Verification (MAV)[30]; generator–critic–judge systems Relatively mature at test time. Recent work is beginning to convert synthetic disagreement, critique, and debate into post-training data.
C8. Partner, Intent, and Strategic Modelling Explicit partner profiles and trust scores; belief and Theory-of-Mind models; next-action or intent prediction; trajectory-based opponent modelling; self-play; population-based training; game-theoretic strategic adaptation. CICERO[31]; SOTOPIA-π[24]; ToMAgent[25]; self-play language games There is a strong tradition in games and social agents, but general partner modelling across heterogeneous, open-ended LLM agents remains immature.

The key observation is that there is currently no single recipe for constructing multi-agent capabilities. Different capabilities are instantiated through different combinations of system design, data generation, learning, interaction, external modules, and test-time computation. Current LLM-MAS research remains strongly scaffolding-heavy; capability-specific post-training that internalises coordination behaviour into model parameters is still comparatively limited.

Key Takeaways

References

  1. Zheng, K., Han, J. M., & Polu, S. (2021). miniF2F: A Cross-System Benchmark for Formal Olympiad-Level Mathematics. arXiv:2109.00110. arXiv
  2. Azerbayev, Z., Piotrowski, B., Schoelkopf, H., Ayers, E. W., Radev, D., & Avigad, J. (2023). ProofNet: Autoformalizing and Formally Proving Undergraduate-Level Mathematics. arXiv:2302.12433. arXiv
  3. Zhu, K., Du, H., Hong, Z., et al. (2025). MultiAgentBench: Evaluating the Collaboration and Competition of LLM Agents. Proceedings of ACL 2025, 8580–8622. ACL Anthology
  4. Jimenez, C. E., Yang, J., Wettig, A., Yao, S., Pei, K., Press, O., & Narasimhan, K. R. (2024). SWE-bench: Can Language Models Resolve Real-World GitHub Issues? ICLR 2024. arXiv
  5. Feng, Y., Huang, Q., Xie, X., Yang, Z., Yu, J., Chen, W., & Tung, A. K. H. (2026). IDRBench: Interactive Deep Research Benchmark. arXiv:2601.06676. arXiv
  6. Jiang, Y., Yang, T., Cui, T., et al. (2026). Deep Research in Physical Sciences: A Multi-Agent Framework and Comprehensive Benchmark. arXiv:2606.18648. arXiv
  7. Fan, Z., Wei, L., Tang, J., Chen, W., Wang, S., Wei, Z., & Huang, F. (2025). AI Hospital: Benchmarking Large Language Models in a Multi-Agent Medical Interaction Simulator. COLING 2025. ACL Anthology
  8. Liu, X., Gu, S., & Song, D. (2026). AgenticPay: A Multi-Agent LLM Negotiation System for Buyer-Seller Transactions. arXiv:2602.06008. arXiv
  9. Zheng, Y., Duan, H., Zhang, Z., Zhu, Y., Min, X., & Zhai, G. (2026). Market-Bench: Benchmarking Large Language Models on Economic and Trade Competition. arXiv:2604.05523. arXiv
  10. Long, Q., Li, Z., Gong, R., Wu, Y. N., Terzopoulos, D., & Gao, X. (2024). TeamCraft: A Benchmark for Multi-Modal Multi-Agent Systems in Minecraft. arXiv:2412.05255. arXiv
  11. Liu, Q., Zhang, J., Hu, J., Wang, S., & Wei, Z. (2026). MECoBench: A Systematic Study of Multimodal Agent Collaboration in Embodied Environments. arXiv:2606.31966. arXiv
  12. Drouin, A., Gasse, M., Caccia, M., et al. (2024). WorkArena: How Capable Are Web Agents at Solving Common Knowledge Work Tasks? arXiv:2403.07718. arXiv
  13. Xie, T., Zhang, D., Chen, J., et al. (2024). OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments. NeurIPS 2024. arXiv
  14. Piao, J., Yan, Y., Zhang, J., et al. (2025). AgentSociety: Large-Scale Simulation of LLM-Driven Generative Agents Advances Understanding of Human Behaviors and Society. arXiv:2502.08691. arXiv
  15. Shi, D., Cao, J., Chen, Q., et al. (2025). TaskCraft: Automated Generation of Agentic Tasks. arXiv:2506.10055. arXiv
  16. Chen, M., Sun, H., Li, T., et al. (2024). Facilitating Multi-turn Function Calling for LLMs via Compositional Instruction Tuning. arXiv:2410.12952. arXiv
  17. Ren, X., Dong, S., Yang, C., et al. (2026). MagicAgent: Towards Generalized Agent Planning. arXiv:2602.19000. arXiv
  18. Kannan, S. S., Venkatesh, V. L. N., & Min, B.-C. (2023). SMART-LLM: Smart Multi-Agent Robot Task Planning using Large Language Models. arXiv:2309.10062. arXiv
  19. Fourney, A., Bansal, G., Mozannar, H., et al. (2024). Magentic-One: A Generalist Multi-Agent System for Solving Complex Tasks. arXiv:2411.04468. arXiv
  20. Yue, Y., Zhang, G., Liu, B., Wan, G., Wang, K., Cheng, D., & Qi, Y. (2025). MasRouter: Learning to Route LLMs for Multi-Agent Systems. ACL 2025. ACL Anthology
  21. Liu, K., Tang, Z., Wang, D., Wang, Z., Zhao, B., & Li, X. (2024). COHERENT: Collaboration of Heterogeneous Multi-Robot System with Large Language Models. arXiv:2409.15146. arXiv
  22. Chen, W., Yuan, J., Qian, C., Yang, C., Liu, Z., & Sun, M. (2024). Optima: Optimizing Effectiveness and Efficiency for LLM-Based Multi-Agent System. arXiv:2410.08115. arXiv
  23. Kim, Y., Seo, S., Park, J., Bennis, M., Kim, S.-L., & Choi, J. (2024). Knowledge Distillation from Language-Oriented to Emergent Communication for Multi-Agent Remote Control. arXiv:2401.12624. arXiv
  24. Wang, R., Yu, H., Zhang, W., et al. (2024). SOTOPIA-π: Interactive Learning of Socially Intelligent Language Agents. ACL 2024. ACL Anthology
  25. Hwang, E., Yin, Y., Carenini, G., West, P., & Shwartz, V. (2025). Infusing Theory of Mind into Socially Intelligent LLM Agents. arXiv:2509.22887. arXiv
  26. Agashe, S., Fan, Y., Reyna, A., & Wang, X. E. (2025). LLM-Coordination: Evaluating and Analyzing Multi-Agent Coordination Abilities in Large Language Models. Findings of NAACL 2025. ACL Anthology
  27. Zhu, D., Ma, X., Shen, Y., et al. (2026). When Tools Fail: Benchmarking Dynamic Replanning and Anomaly Recovery in LLM Agents. arXiv:2606.05806. arXiv
  28. Froger, R., Andrews, P., Bettini, M., et al. (2026). Gaia2: Benchmarking LLM Agents on Dynamic and Asynchronous Environments. ICLR 2026. arXiv
  29. He, H., Li, Y., Wen, D., Chen, Y., Cheng, R., Chen, D., & Lau, F. C. M. (2025). Debating Truth: Debate-Driven Claim Verification with Multiple Large Language Model Agents. arXiv:2507.19090. arXiv
  30. Lifshitz, S., McIlraith, S. A., & Du, Y. (2025). Multi-Agent Verification: Scaling Test-Time Compute with Multiple Verifiers. arXiv:2502.20379. arXiv
  31. Meta Fundamental AI Research Diplomacy Team (FAIR). (2022). Human-Level Play in the Game of Diplomacy by Combining Language Models with Strategic Reasoning. Science, 378(6624), 1067–1074. DOI
  32. Wong, R., Wang, J., Zhao, J., Chen, L., Gao, Y., Zhang, L., Zhou, X., Wang, Z., Xiang, K., Zhang, G., Huang, W., Wang, Y., & Wang, K. (2025). WideSearch: Benchmarking Agentic Broad Info-Seeking. arXiv:2508.07999. arXiv
  33. Wei, J., Sun, Z., Papay, S., McKinney, S., Han, J., Fulford, I., Chung, H. W., Passos, A. T., Fedus, W., & Glaese, A. (2025). BrowseComp: A Simple Yet Challenging Benchmark for Browsing Agents. arXiv:2504.12516. arXiv
  34. Dafoe, A., Hughes, E., Bachrach, Y., Collins, T., McKee, K. R., Leibo, J. Z., Larson, K., & Graepel, T. (2020). Open Problems in Cooperative AI. arXiv:2012.08630. arXiv
  35. Bernstein, D. S., Givan, R., Immerman, N., & Zilberstein, S. (2002). The Complexity of Decentralized Control of Markov Decision Processes. Mathematics of Operations Research, 27(4), 819–840. DOI
  36. Harsanyi, J. C. (1967). Games with Incomplete Information Played by “Bayesian” Players, I–III: Part I. The Basic Model. Management Science, 14(3), 159–182. DOI
  37. Shapley, L. S. (1953). Stochastic Games. Proceedings of the National Academy of Sciences, 39(10), 1095–1100. DOI
  38. Rubinstein, A. (1982). Perfect Equilibrium in a Bargaining Model. Econometrica, 50(1), 97–109. Author PDF
  39. Myerson, R. B. (1981). Optimal Auction Design. Mathematics of Operations Research, 6(1), 58–73. DOI
  40. Sandholm, W. H. (2010). Population Games and Evolutionary Dynamics. MIT Press. Publisher