Research map / Multi-agent systems and games / LLM agents
LLawCo: Learning Laws of Cooperation for Modeling Embodied Multi-Agent Behavior
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 LLawCo, which induces natural-language cooperation laws from failed multi-agent episodes, uses them to regenerate law-referencing reasoning on successful traces, fine-tunes LLM agents via SFT, and introduces the PARTNR-Dialog communicative benchmark.
Problem
Communicative embodied LLM agents often act misaligned with partners and environment state, leading to inefficient cooperation and low task success; existing communicative benchmarks lack training splits.
Approach
Agents summarize failure traces into a small set of high-level laws, filter successful law-aligned traces, regenerate law-referencing thoughts, and SFT the backbone so the laws are internalized in its chain of thought and can be edited at inference time.
Key results
- On PARTNR-Dialog with Qwen-3-14B, LLawCo reaches 0.84 completion and 0.74 success, above all communicative baselines (best baseline 0.70 success with CommPARTNR).
- On TDW-MAT, LLawCo with LLaMA-3.1-8B reaches 0.67 total success vs 0.61 for CoELA; with Qwen-3-14B it reaches 0.80 vs 0.73 for CoELA.
- Removing SFT drops Qwen-3-14B success from 0.74 to 0.69; removing laws drops it to 0.67.
Surprising findings
- Replacing a law at inference time with a contradictory one changed send-message rate on TDW-MAT from 92% to 19%, showing controllability via laws.
- On LLaMA-3.1-8B, LLawCo underperforms ReAct on PARTNR-Dialog (0.57 vs 0.62 success), attributed to weak Talk usage.
Limitations
- Experiments restricted to two-agent collaboration.
- Extra training cost for trace collection, law induction, and SFT.
- Induced laws may encode undesirable behaviors and do not guarantee safety.
Experimental setup
Datasets: PARTNR TDW-MAT
Benchmarks: PARTNR-Dialog TDW-MAT
Metrics: percentage complete success rate steps replan rounds talk count food/stuff success rate
Baselines: ReAct RoCo CoELA CommPARTNR MHP
Builds on: ReAct PARTNR CoELA Constitutional AI
Closest papers by meaning
- Building Cooperative Embodied Agents Modularly with Large Language Models (ICLR 2024 · similarity 0.90)
- Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration (CVPR 2025 · similarity 0.90)
- CaPo: Cooperative Plan Optimization for Efficient Embodied Multi-Agent Cooperation (ICLR 2025 · similarity 0.90)
- PARTNR: A Benchmark for Planning and Reasoning in Embodied Multi-agent Tasks (ICLR 2025 · similarity 0.90)
- 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)
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.