Research map / Multi-agent systems and games / LLM agents
Shapley-Coop: Credit Assignment for Emergent Cooperation in Self-Interested LLM Agents
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 Shapley-Coop, a workflow where self-interested LLM agents use Shapley-inspired chain-of-thought reasoning and structured negotiation for task-time pricing and post-task reward redistribution, evaluated on Escape Room, Raid Battle, and ChatDev.
Problem
Self-interested LLM agents in open-ended multi-agent settings without predefined roles or coordination rules fall into social dilemmas; the paper targets aligning heterogeneous goals and fair credit assignment.
Approach
Combines Shapley value marginal-contribution reasoning, implemented as LLM prompts (short-term qualitative externality reasoning and long-term retrospective Shapley estimation), with a structured negotiation protocol so agents set prices for actions and redistribute rewards after the task, without weight updates.
Key results
- In Raid Battle Level 3, LLM+SC allocates rewards within about 1.6 percentage points of expected Shapley shares (e.g., Agent 1 26.33% vs 27.89%), while LLM+STS deviates by up to 3.70 points (Agent 4 at 20.00% vs 23.70%).
- In Raid Battle Level 2, LLM+STS underestimates Agent 1 by 6.26 points; LLM+SC reduces the largest deviation to 2.90 points.
- On ChatDev, WEV-based benchmark rewards are mostly within 6% of human-assigned rewards, with programmer and reviewer roles near-perfectly aligned.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the full Shapley-Coop workflow instead of using short-term Shapley reasoning alone (ablation) | LLM agents in Raid Battle Level 1 · Raid Battle Level 1 | A1 reward-allocation deviation from expected share: -4.52 -> +0.48 | Helped |
| add the full Shapley-Coop workflow instead of using short-term Shapley reasoning alone (ablation) | LLM agents in Raid Battle Level 2 · Raid Battle Level 2 | A1 reward-allocation deviation from expected share: -6.26 -> -1.93 | Helped |
| add the full Shapley-Coop workflow instead of using short-term Shapley reasoning alone (ablation) | LLM agents in Raid Battle Level 3 · Raid Battle Level 3 | A4 reward-allocation deviation from expected share: -3.70 -> +0.63 | Helped |
| add structured negotiation and Shapley reasoning rather than using no negotiation or cooperation (ablation) | LLM agents in Escape Room · Escape Room | successful cooperation | Helped |
| add short-term Shapley reasoning to standard negotiation (ablation) | LLM agents in Escape Room · Escape Room | cooperative task success | Helped |
| add long-term Shapley reasoning and post-task reward redistribution to short-term Shapley reasoning (ablation) | LLM agents in Escape Room · Escape Room | fairness of payoff allocation | Helped |
| add Shapley-Coop pricing and coordination instead of standard negotiation (ablation) | LLM agents in Raid Battle · Raid Battle | game performance | Helped |
| increase negotiation rounds from 1 to 3 (ablation) | Shapley-Coop agents · Raid Battle Level 3 | A1 reward share: 22% -> 27% | Helped |
Limitations
- Pricing cannot be dynamically adjusted during collaboration, as the authors acknowledge.
- Shapley values are approximated by LLM reasoning and negotiation rather than computed exactly, so convergence to the true Shapley allocation is only shown empirically.
- Evaluation uses small custom games and a single backbone for the Escape Room; ChatDev results compare WEV against human-assigned rewards rather than an outcome-based metric.
Experimental setup
Benchmarks: Escape Room Raid Battle ChatDev BMI Calculator ChatDev ArtCanvas
Metrics: agent payoff global reward reward allocation deviation from expected Shapley share weighted earned value adjustment
Baselines: LLM-only LLM+NEG LLM+STS
Builds on: Shapley value ChatDev Pigovian taxes Coase theorem
Closest papers by meaning
- More Capable, Less Cooperative? When LLMs Fail at Zero-Cost Collaboration (ICML 2026 · similarity 0.91)
- COOP$^2$: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.89)
- Cooperation, Competition, and Maliciousness: LLM-Stakeholders Interactive Negotiation (NeurIPS 2024 · similarity 0.89)
- Cooperate or Collapse: Emergence of Sustainable Cooperation in a Society of LLM Agents (NeurIPS 2024 · similarity 0.89)
- Evaluating LLMs in Open-Source Games (NeurIPS 2025 · similarity 0.88)
- MARSHAL: Incentivizing Multi-Agent Reasoning via Self-Play with Strategic LLMs (ICLR 2026 · 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.