Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration
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 Cooperative Tree Search (CoTS), which runs a modified MCTS over multi-agent LLM plan discussions with LLM-based rewards and a plan evaluation module, improving embodied collaboration on TDW-MAT and C-WAH.
Problem
Multi-agent embodied LLM agents struggle with long-term collaborative planning and coordination; single-chain dialogue (RoCo) and independent decisions (CoELA) lead to suboptimal or erroneous plans and disruptive frequent replanning.
Approach
Adapts Monte Carlo tree search so that agents discuss and search over collaborative plans, scoring nodes with a non-simulated LLM-based reward (allocation and movement cost) instead of costly environment rollouts, plus a plan evaluation module that updates plans only when they become unsuitable.
Key results
- On TDW-MAT with oracle perception and GPT-4, CoTS averages 93% transport rate, a 9.41% relative improvement over CoELA (85%).
- Without oracle perception, CoTS reaches 83% average transport rate versus 71% for CoELA (16.90% relative gain).
- On C-WAH with GPT-4, CoTS needs 81 visual-observation steps vs 92 for CoELA (11.96% efficiency gain).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CoTS instead of CoELA | GPT-4 multi-agent collaboration · TDW-MAT w/ Oracle Perception | Average transport rate (%): 85 -> 93+9.41% | Helped |
| use CoTS instead of CoELA | GPT-4 multi-agent collaboration · TDW-MAT w/o Oracle Perception | Average transport rate (%): 71 -> 83+16.90% | Helped |
| use CoTS instead of CoELA | GPT-4 multi-agent collaboration · C-WAH Visual Obs | Average step: 92 -> 81+11.96% | Helped |
| use CoTS instead of CoELA | GPT-4 multi-agent collaboration · C-WAH Symbolic Obs | Average step: 57 -> 49+14.04% | Helped |
| remove the LLM-based reward function (ablation) | CoTS · TDW-MAT w/ Oracle Perception | Average transport rate (%): 93 -> 89−4.30% | Hurt |
| replace collaborative tree search with a single-branch structure (ablation) | CoTS · TDW-MAT w/ Oracle Perception | Average transport rate (%): 93 -> 88−5.38% | Hurt |
| remove the plan evaluation and update module (ablation) | CoTS · TDW-MAT w/ Oracle Perception | Average transport rate (%): 93 -> 90−3.23% | Hurt |
| increase the branching factor (ablation) | CoTS tree search · TDW-MAT w/ Oracle Perception | Average transport rate (%): null | Helped |
Surprising findings
- With GPT-3.5, CaPo slightly outperforms CoTS on some metrics, which the authors attribute to GPT-3.5's weaker evaluation and scoring ability.
Limitations
- Relies heavily on the reasoning ability of the underlying LLM; GPT-3.5 is verbose and sometimes inaccurate.
- Performance on some metrics with GPT-3.5 trails CaPo.
Experimental setup
Datasets: TDW-MAT C-WAH
Benchmarks: TDW-MAT C-WAH
Metrics: transport rate average steps to completion
Baselines: RHP MHP MAT CoELA RoCo CaPo
Builds on: CoELA Monte Carlo tree search chain-of-thought
Closest papers by meaning
- CaPo: Cooperative Plan Optimization for Efficient Embodied Multi-Agent Cooperation (ICLR 2025 · similarity 0.92)
- LLawCo: Learning Laws of Cooperation for Modeling Embodied Multi-Agent Behavior (ICML 2026 · similarity 0.90)
- From Assumptions to Actions: Turning LLM Reasoning into Uncertainty-Aware Planning for Embodied Agents (ICLR 2026 · similarity 0.90)
- PARTNR: A Benchmark for Planning and Reasoning in Embodied Multi-agent Tasks (ICLR 2025 · similarity 0.89)
- Building Cooperative Embodied Agents Modularly with Large Language Models (ICLR 2024 · 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.