Research map / Multi-agent systems and games
LLM agents: research map
1,294 papers on LLM agents in Multi-agent systems and games, from ICML, NeurIPS, ICLR, CVPR, AAAI, ICCV and ECCV, and the journals TPAMI, IJCV and JMLR (2016–2026), grouped into 2 clusters and 9 approaches. The busiest year so far is 2026.
What it is
Systems where one or several language-model agents plan, use tools and work together on tasks such as software engineering or operating a computer's interface (GUI agents). The topic covers orchestration, memory, benchmarks, failure analysis and safety.
Terms you will see on this page:
- Multi-agent system
- several agents, here usually LLMs, interacting or collaborating.
- GUI agent
- an agent that operates apps by reading the screen and clicking or typing.
- Orchestration
- how agents and tools are coordinated to complete a task.
Start reading with the most cited papers here:
- Reflexion: language agents with verbal reinforcement learning (NeurIPS 2023 · 559 citations)
- MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework (ICLR 2024 · 153 citations)
Within Multi-agent systems and games, its share grew from 10.3% in 2023–24 to 60.4% in 2025–26 (65 → 1,157 papers at ICML, NeurIPS, ICLR, CVPR and AAAI, the venues with data for all four years).
Explore LLM agents in the interactive map
Working on something in this topic? Describe your idea in scime atlas to see which approach it falls under, the closest papers by meaning and how crowded the spot has become.
What changed recently
Approaches whose share of the topic moved most between 2023–24 and 2025–26 (ICML, NeurIPS, ICLR, CVPR and AAAI):
- Robust and efficient agent execution rising: 5% → 15% of the topic (3 → 169 papers)
- Agent safety and attack benchmarks rising: 0% → 8% of the topic (0 → 95 papers)
- Software engineering agent benchmarks rising: 5% → 11% of the topic (3 → 126 papers)
- LLM multi-agent collaboration and safety falling: 23% → 6% of the topic (15 → 69 papers)
- LLM multi-agent decision making falling: 26% → 15% of the topic (17 → 176 papers)
Papers from 2025–2026 that are already cited most:
- VerilogCoder: Autonomous Verilog Coding Agents with Graph-based Planning and Abstract Syntax Tree (AST)-based Waveform Tracing Tool (AAAI 2025 · 46 citations)
- Why Do Multi-Agent LLM Systems Fail? (NeurIPS 2025 · 33 citations)
- BotSim: LLM-Powered Malicious Social Botnet Simulation (AAAI 2025 · 14 citations)
- TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks (NeurIPS 2025 · 13 citations)
Approaches and key papers
Agent benchmarks, learning and safety · 787 papers
Keywords: gui, bench, software
Approaches in this cluster:
- LLM multi-agent decision making (195 papers)
Designs collaboration, hierarchy, and credit assignment for LLM-based multi-agent teams. - Self-evolving language agent learning (193 papers)
Improves language agents via search, step-level value models, reinforcement learning, and self-distillation. - Robust and efficient agent execution (185 papers)
Benchmarks tool-using agent robustness and speeds execution with plan caching and compilation. - Agent data synthesis and behavior studies (112 papers)
Synthesizes agent training data and examines behavioral biases in interactive LLM agents. - Agent safety and attack benchmarks (102 papers)
Benchmarks multi-turn attacks, prompt injection, and malicious tool use against LLM agents.
Most cited and most cited since 2024:
- Reflexion: language agents with verbal reinforcement learning (NeurIPS 2023 · 559 citations)
- MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework (ICLR 2024 · 153 citations)
- AgentBench: Evaluating LLMs as Agents (ICLR 2024 · 53 citations)
- AgentPoison: Red-teaming LLM Agents via Poisoning Memory or Knowledge Bases (NeurIPS 2024 · 49 citations)
Multi-agent LLM systems · 507 papers
Keywords: llm agents, llms, agent systems
Approaches in this cluster:
- Agentic benchmarks and orchestration (198 papers)
Benchmarks long-horizon agent memory, reliability, and orchestration of sub-agents. - LLM multi-agent collaboration and safety (87 papers)
Studies collaboration, competition, evaluation, and monitoring of LLM-based multi-agent systems. - Software engineering agent benchmarks (135 papers)
Builds repository-level environments and benchmarks for coding and computer-use agents. - GUI agent memory and trajectories (87 papers)
Improves GUI agents through memory, synthesized interaction data, and desktop benchmarks.
Most cited and most cited since 2024:
- Mind2Web: Towards a Generalist Agent for the Web (NeurIPS 2023 · 69 citations)
- ChatEval: Towards Better LLM-based Evaluators through Multi-Agent Debate (ICLR 2024 · 55 citations)
- VerilogCoder: Autonomous Verilog Coding Agents with Graph-based Planning and Abstract Syntax Tree (AST)-based Waveform Tracing Tool (AAAI 2025 · 46 citations)
- OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments (NeurIPS 2024 · 39 citations)
Related topics in Multi-agent systems and games
- Cooperative multi-agent RL (635)
- Games and equilibria (537)
- Human-AI and goal-conditioned (1,320)
