Research map / Multi-agent systems and games / Human-AI and goal-conditioned
AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes AgentVerse, a multi-agent framework that simulates human group problem-solving (expert recruitment, collaborative decision-making, action execution, evaluation) with GPT-3.5/GPT-4 agents, and reports gains over single agents plus emergent collaborative and destructive behaviors in Minecraft.
Problem
Single LLM agents with fixed roles struggle with complex real-world tasks that benefit from cooperation; prior multi-agent work is narrow in tasks and uses static roles.
Approach
A four-stage loop in which a recruiter LLM dynamically generates expert descriptions, agents collaborate via horizontal or vertical communication structures, execute actions, and an evaluator's verbal feedback drives recomposition of the group in later rounds.
Key results
- On Humaneval with GPT-4, pass@1 rises from 83.5 (CoT) to 87.2 (Solo) and 89.0 (Group).
- On tool-use tasks, AgentVerse completes 9 of 10 complex tasks versus 3 of 10 for a single ReAct agent.
- For GPT-3.5-Turbo, the Group setup underperforms Solo on two of three tasks, attributed to swaying by incorrect feedback.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use multi-agent Group instead of Solo | AgentVerse with GPT-3.5-Turbo · FED | win rate: 81.1 -> 85.1 | Helped |
| Use multi-agent Group instead of Solo | AgentVerse with GPT-3.5-Turbo · Commongen-Challenge | concept coverage: 93.6 -> 92.3 | Hurt |
| Use multi-agent Group instead of Solo | AgentVerse with GPT-3.5-Turbo · MGSM | accuracy: 82.4 -> 80.8 | Hurt |
| Use multi-agent Group instead of Solo | AgentVerse with GPT-4 · FED | win rate: 95.8 -> 96.8 | Helped |
| Use multi-agent Group instead of Solo | AgentVerse with GPT-4 · Commongen-Challenge | concept coverage: 99.0 -> 99.1 | Helped |
| Use multi-agent Group instead of Solo | AgentVerse with GPT-4 · MGSM | accuracy: 96.0 -> 95.2 | Hurt |
| Use multi-agent Group instead of Solo | AgentVerse with GPT-4 · Logic Grid Puzzles | accuracy: 64.0 -> 66.5 | Helped |
| Use multi-agent Group instead of Solo | AgentVerse with GPT-3.5-Turbo · HumanEval | pass@1: 74.4 -> 75.6 | Helped |
| Use multi-agent Group instead of Solo | AgentVerse with GPT-4 · HumanEval | pass@1: 87.2 -> 89.0 | Helped |
| Use AgentVerse multi-agent collaboration instead of a standalone ReAct agent | Tool-use agents · 10 complex tool-use tasks | tasks accomplished: 3 -> 9 | Helped |
Surprising findings
- With GPT-3.5-Turbo, group discussion sometimes hurts performance, as agents are swayed by incorrect feedback from peers.
- Agents in Minecraft exhibit destructive behaviors, such as harming other agents or destroying a village library to obtain materials.
Limitations
- Number of experts per task is predefined manually rather than automated.
- Evaluation of tool use is manual on only 10 instructions, with no standard benchmark.
- Evaluator is an agent rather than a human; the paper notes human evaluation is future work.
Experimental setup
Datasets: FED Commongen-Challenge MGSM Logic Grid Puzzles Humaneval GSM-8k
Benchmarks: FED Commongen-Challenge MGSM Logic Grid Puzzles Humaneval 10 manually crafted tool-use instructions
Metrics: win rate concept coverage percentage accuracy pass@1 task success (manual)
Baselines: CoT ReAct agent Solo (single agent in AgentVerse)
Builds on: chain-of-thought ReAct
Closest papers by meaning
- Emergent Coordination in Multi-Agent Language Models (ICLR 2026 · similarity 0.90)
- Scaling Large Language Model-based Multi-Agent Collaboration (ICLR 2025 · similarity 0.90)
- Do as We Do, Not as You Think: the Conformity of Large Language Models (ICLR 2025 · similarity 0.89)
- Latent Collaboration in Multi-Agent Systems (ICML 2026 · similarity 0.89)
- G-Memory: Tracing Hierarchical Memory for Multi-Agent Systems (NeurIPS 2025 · similarity 0.89)
- COOP$^2$: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.