Research map / Multi-agent systems and games / LLM agents
Great Minds Think Alike: Contextual Tacit Communication for Decentralized LLM-Agent Cooperation
Yue Pei, Hongming Zhang, Jiarui Guan, Jusheng Zhang, Liang Lin, Haogang Zhu, Ziliang Chen
Read the paperCheck your idea against 137,217 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 a contextual tacit communication approach that lets decentralized LLM agents coordinate for embodied cooperation without online message passing.
Problem
LLM planners for cooperative embodied agents suffer inconsistency and miscoordination under partial observability, and explicit communication is costly or unavailable.
Approach
Uses contextual tacit communication, coordinating through shared context rather than online message passing, to avoid the independent-planning fallback when communication is removed.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add TRM rule context at test time (vs numeric rectification only) (ablation) | Tact, Qwen2.5-VL-7B-SFT · VIKI-L2 (ID) | ACC_ID: 77.4 -> 89.2 | Improved |
| add TRM rule context at test time (ablation) | Tact, Qwen2.5-VL-7B-SFT · VIKI-L2 (OOD) | ACC_OOD: 3.6 -> 30.2 | Improved |
| add residual banding for rule synthesis (ablation) | Tact · VIKI-L2 (ID) | ACC_ID: 10.4 -> 89.2 | Improved |
| add bias supervision for rule generation (ablation) | Tact · VIKI-L2 (ID) | ACC_ID: 8.5 -> 89.2 | Improved |
| add the full Tact protocol to the base planner (ablation) | Qwen2.5-VL-7B-SFT planner · VIKI-L2 (ID) | ACC_ID: 5.2 -> 89.2 | Improved |
| add rule constraints on top of numeric rectification (Tact vs Tact-b) (ablation) | GPT-4o/Qwen backbones · VIKI-L2 (ID) | ACC_ID: 5.4 -> 16.4 | Improved |
| use Tact instead of decentralized planner | Qwen2.5-VL-7B-SFT · VIKI-L2 (OOD) | ACC_OOD: 2.0 -> 30.2 | Improved |
| use Tact instead of SFT centralized planner under global observation | Qwen2.5-VL-7B-SFT · VIKI-L2 (ID) | ACC_ID: 96.5 -> 89.2 | Hurt |
| use residual band 0.5 instead of 2.0 (ablation) | Tact · VIKI-L2 (ID) | ACC_ID: 82.6 -> 89.2 | Improved |
| use residual band 0.5 instead of 0.2 (ablation) | Tact · VIKI-L2 (ID) | ACC_ID: 88.4 -> 89.2 | Improved |
| loosen residual band from 0.2 to 2.0 (ablation) | Tact · VIKI-L2 | value regret: 0.55 -> 0.92 | Hurt |
| use Tact instead of Dec-RL-ZERO with zero SFT | Qwen2.5-VL-3B · VIKI-L2 (zero-SFT, ID) | ACC_ID: 0.72 -> 8.45 | Improved |
Closest papers by meaning
- Building Cooperative Embodied Agents Modularly with Large Language Models (ICLR 2024 · similarity 0.90)
- LLawCo: Learning Laws of Cooperation for Modeling Embodied Multi-Agent Behavior (ICML 2026 · similarity 0.89)
- CaPo: Cooperative Plan Optimization for Efficient Embodied Multi-Agent Cooperation (ICLR 2025 · similarity 0.89)
- From Assumptions to Actions: Turning LLM Reasoning into Uncertainty-Aware Planning for Embodied Agents (ICLR 2026 · similarity 0.89)
- COOP$^2$: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.89)
- Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration (CVPR 2025 · similarity 0.88)
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.
