Research map / Multi-agent systems and games / LLM agents
Robust and efficient agent execution
Benchmarks tool-using agent robustness and speeds execution with plan caching and compilation.
174 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2017–2026) · an approach in LLM agents, keyword cluster “gui · bench · software”
State History Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
3 papers in 2023–24 and 169 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 4.6% to 14.6%.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2017 · Parameterised Verification of Infinite State Multi-Agent Systems via Predicate Abstraction (AAAI 2017)
- 2020 · An Operational Semantics for True Concurrency in BDI Agent Systems (AAAI 2020 · 3 citations)
- 2024 · Multi-Stage Prompting for Next Best Agent Recommendations in Adaptive Workflows (AAAI 2024 · 3 citations)
- 2025 · WebPilot: A Versatile and Autonomous Multi-Agent System for Web Task Execution with Strategic Exploration (AAAI 2025 · 12 citations)
- 2026 · Measuring Agents in Production (ICML 2026 · 3 citations)
Earliest papers here
- Parameterised Verification of Infinite State Multi-Agent Systems via Predicate Abstraction (AAAI 2017)
- An Operational Semantics for True Concurrency in BDI Agent Systems (AAAI 2020 · 3 citations)
- Multi-Stage Prompting for Next Best Agent Recommendations in Adaptive Workflows (AAAI 2024 · 3 citations)
What has been tried
23 papers here report 237 measured changes (a component added, removed or swapped, compared with a baseline): helped 73%, no clear effect 7%, hurt 18%, mixed 2%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add Child Reflection | WebArena IS tasks · success rate | 74% -> 100% | Helped | WebPilot: A Versatile and Autonomous Multi-Agent System for Web Task Execution with Strategic Exploration (AAAI 2025 · 12 citations) |
| Represent workflows as code instead of JSON | ComfyBench vanilla · resolve rate | 33.0 -> 41.0 | Helped | ComfyBench: Benchmarking LLM-based Agents in ComfyUI for Autonomously Designing Collaborative AI Systems (CVPR 2025 · 9 citations) |
| SFT the planner on OWL trajectories | GAIA validation, average · accuracy | 36.36 -> 41.21 | Helped | OWL: Optimized Workforce Learning for General Multi-Agent Assistance in Real-World Task Automation (NeurIPS 2025 · 3 citations) |
| replace the Stage 1 mapping with the Ground Truth agent list | ABCD · NBA Acc.% | 0.5125 -> 0.6875 | Helped | Multi-Stage Prompting for Next Best Agent Recommendations in Adaptive Workflows (AAAI 2024 · 3 citations) |
| add chain-of-thought (use_thinking) | WorkArena · success rate | GPT-3.5 8.5 to 6.1 when removed; Llama3 20.0 to 8.5 when removed | Helped | WorkArena: How Capable are Web Agents at Solving Common Knowledge Work Tasks? (ICML 2024 · 2 citations) |
| Add both Reader and Proposer to the research team | ImageNet16-120 test · accuracy | 47.14±1.30 -> 48.72±0.14 | Helped | NADER: Neural Architecture Design via Multi-Agent Collaboration (CVPR 2025 · 1 citation) |
| Add objective environment feedback in place of LLM critique | rSDE-Bench Website Basic · accuracy (%) | 78.08 -> 90.75 | Helped | Self-Evolving Multi-Agent Collaboration Networks for Software Development (ICLR 2025 · 1 citation) |
| Allow Plan states to transition to any state type | SWE-bench Lite · resolve rate (Pass@1) | 24.3 -> 25.7 | Mixed | SWE-Search: Enhancing Software Agents with Monte Carlo Tree Search and Iterative Refinement (ICLR 2025 · 1 citation) |
| Predict three candidate tasks instead of one | ProactiveBench · F1-Score | 52.15% -> 78.80% | Helped | Proactive Agent: Shifting LLM Agents from Reactive Responses to Active Assistance (ICLR 2025 · 1 citation) |
| Add agent-centric semantic memory (ASM) to ReAct | AI2-THOR · Success Rate | +0.03 | Helped | Mosaic: Runtime-Efficient Multi-Agent Embodied Planning (ICML 2026 · 0 citations) |
| Add priority scheduling | SWE-bench Common · Function Match Rate | 73.12% -> 76.34% | Helped | OrcaLoca: An LLM Agent Framework for Software Issue Localization (ICML 2025 · 0 citations) |
| Add CI/CD configuration files to the setup context rather than use only text files | SWA-Bench · successfully extracted environments | 15 -> 44 | Helped | Automated Benchmark Generation for Repository-Level Coding Tasks (ICML 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 169 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- GAIA · 12 of 169 papers
- HotpotQA · 11 of 169 papers
- SWE-bench Verified · 6 of 169 papers
- GPQA · 6 of 169 papers
- HumanEval · 6 of 169 papers
- SWE-bench Lite · 5 of 169 papers
- MATH · 5 of 169 papers
- GSM8K · 5 of 169 papers
Baselines
- ReAct · 10 of 169 papers
- Gemini-2.5-Pro · 8 of 169 papers
- GPT-5 · 7 of 169 papers
- GPT-4o · 7 of 169 papers
- OpenHands · 5 of 169 papers
- CoT · 5 of 169 papers
- Reflexion · 5 of 169 papers
- Multi-Agent Debate · 4 of 169 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 172 papers; from OpenAlex and the papers' affiliations).
- Stanford University · 8 papers
- Tsinghua University · 7 papers
- The University of Hong Kong · 7 papers
- Carnegie Mellon University · 6 papers
- Google DeepMind · 5 papers
- Massachusetts Institute of Technology · 5 papers
- Amazon · 5 papers
- Hong Kong University of Science and Technology · 5 papers
- Shanghai Jiao Tong University · 4 papers
- Meta · 4 papers
Use it
Download BibTeX of 21 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- LLM multi-agent decision making (194 papers)
- Self-evolving language agent learning (188 papers)
- Agent data synthesis and behavior studies (108 papers)
- Agent safety and attack benchmarks (98 papers)
Elsewhere in the topic
- Agentic benchmarks and orchestration (185 papers)
- Software engineering agent benchmarks (129 papers)
- LLM multi-agent collaboration and safety (84 papers)
- GUI agent memory and trajectories (72 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).