Research map / Multi-agent systems and games / LLM agents

REVECA: Adaptive Planning and Trajectory-Based Validation in Cooperative Language Agents Using Information Relevance and Relative Proximity

AAAI 2025 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 1 citations (OpenAlex)

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 REVECA, a GPT-4o-mini-driven cognitive architecture for two-agent cooperation that uses LLM relevance scoring, proximity-aware top-K adaptive planning, and trajectory-based plan validation to reduce communication and false plans.

Problem

Cooperative LLM embodied agents under decentralized, partially observable multi-room household tasks struggle with accumulating information, globally suboptimal planning ignoring collaborators, and false plans caused by collaborators' unseen environment changes, while constant communication is costl

Approach

Score object relevance by the LLM at observation time (Strong/Medium/Low/None), retrieve top-K observations ranked by relevance and natural-language relative proximity to agents, and validate plans by having the LLM predict collaborators' trajectories since the information was acquired, querying collaborators only when an interaction is likely.

Key results

• In C-WAH, REVECA with GPT-4o-mini reached 44.20 simulation steps and 38.44 m travel distance versus 57.00 steps for GPT-4-driven CoELA.
• In TDW-MAT, REVECA (4o-mini) achieved 0.87 TOTAL success rate versus 0.71 for GPT-4-driven CoELA and 0.79 for RHP.
• In Overcooked-AI, REVECA (4o-mini) performed comparably to ProAgent (4o-mini) across five layouts, without environment-specific modifications.

Measured effects reported in the paper

Surprising findings

• Disabling chain-of-thought cut per-component inference time from several seconds to about 1 second while only modestly increasing simulation steps (44.2 to 48.9).
• GPT-3.5-driven CoELA performed worse than the MCTS-based MHP planner in C-WAH.

Limitations

• Validated only in household and cooperative game settings; open-world outdoor environments not tested.
• Relies on a hand-built low-level action skill book provided by the benchmark.
• Scalability to three or more agents is discussed but not experimentally validated, and validation cost may grow linearly with agent count.

Experimental setup

Datasets: C-WAH TDW-MAT Noisy-C-WAH Overcooked-AI

Benchmarks: C-WAH TDW-MAT Noisy-C-WAH Overcooked-AI

Metrics: simulation steps travel distance success rate reward user study Likert score

Baselines: MHP RHP CoELA SP PBT FCP MEP COLE ProAgent

Builds on: CoELA ProAgent zero-shot chain-of-thought Watch-And-Help Overcooked-AI

Closest papers by meaning

• COOP$^2$: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.89)
• CaPo: Cooperative Plan Optimization for Efficient Embodied Multi-Agent Cooperation (ICLR 2025 · similarity 0.89)
• Building Cooperative Embodied Agents Modularly with Large Language Models (ICLR 2024 · similarity 0.88)
• Adaptive Collaboration with Humans: Metacognitive Policy Optimization for Multi-Agent LLMs with Continual Learning (ICLR 2026 · similarity 0.87)
• From Assumptions to Actions: Turning LLM Reasoning into Uncertainty-Aware Planning for Embodied Agents (ICLR 2026 · similarity 0.87)
• Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration (CVPR 2025 · similarity 0.86)