Research map / Multi-agent systems and games / Games and equilibria
From Debate to Equilibrium: Belief‑Driven Multi‑Agent LLM Reasoning via Bayesian Nash Equilibrium
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.
Casts multi-LLM reasoning as an incomplete-information game and proposes ECON, a hierarchical coordinator–executor RL framework that seeks a Bayesian Nash equilibrium via belief networks instead of costly multi-round debate.
Problem
Multi-agent LLM debate improves reasoning but incurs heavy token/communication costs and lacks convergence guarantees.
Approach
Replace explicit inter-agent message passing with belief-driven coordination: each Execution LLM uses a belief network to map its local history to prompt-control parameters (temperature, repetition penalty), trained via TD losses and a centralized mixing network, with a Coordinator LLM aggregating final outputs; the setup is framed as seeking a BNE with a claimed sublinear regret bound.
Key results
- ECON outperforms existing multi-LLM approaches by 11.2% on average across six benchmarks, per the abstract and introduction.
- On TravelPlanner test set with GPT-4-Turbo, ECON reaches 9.3% final pass rate versus 3.7% for three-round debate.
- Compared with a 3-round MAD protocol, ECON reduces token usage by 21.4% on average.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ECON instead of three-round debate | GPT-4 TravelPlanner planning · TravelPlanner validation | final pass rate: 2.3 -> 7.2 | Helped |
| Use ECON instead of three-round debate | GPT-4 TravelPlanner planning · TravelPlanner test | final pass rate: 3.7 -> 9.3 | Helped |
| Use three LLaMA3.1 405B execution models instead of three LLaMA3.1 8B models | ECON · GSM-Hard | accuracy: 48.71 -> 61.29 | Helped |
| Use heterogeneous execution models rather than homogeneous three-model configurations | ECON · GSM-Hard and MATH | accuracy | Mixed |
| Use ECON instead of three-round debate | LLaMA3.1 70B reasoning · MATH, GSM8K, GSM-Hard | performance: 71.58 -> 81.47; 86.32 -> 92.70; 41.98 -> 51.43 | Helped |
| Use ECON instead of three-round debate | LLaMA3.1 70B reasoning · MATH, GSM8K, GSM-Hard | token usage: 2154.87 -> 1629.79; 1391.57 -> 1131.65; 3030.73 -> 1518.76 | Helped |
| Remove task-specific reward R2 (ablation) | ECON reward design · MATH — LLaMA3.1 70B | accuracy: 81.47 -> 77.55 | Hurt |
| Remove task-specific reward R2 and reward R3 (ablation) | ECON reward design · MATH — LLaMA3.1 70B | accuracy: 81.47 -> 74.32 | Hurt |
| Remove collaborative/self-evaluation reward R3 (ablation) | ECON reward design · MATH — LLaMA3.1 70B | accuracy: 81.47 -> 76.21 | Hurt |
| Remove coordinator strategy and direct answer (ablation) | ECON · MATH — LLaMA3.1 70B | accuracy: 81.47 -> 71.35 | Hurt |
| Provide detailed coordinator strategy (ablation) | ECON · MATH — LLaMA3.1 70B | accuracy: 81.47 -> 72.31 | Hurt |
| Remove concatenation (ablation) | ECON · MATH — LLaMA3.1 70B | accuracy: -4.4 | Hurt |
Surprising findings
- Heterogeneous Execution LLM ensembles performed worse than homogeneous ones, attributed to difficulty reaching equilibrium.
- Adding more than four Execution LLMs gave minimal gains and sometimes declined performance with a single coordinator.
Limitations
- Convergence and BNE proofs rely on strong assumptions (quasi-concavity, Lipschitz gradients, approximate posterior alignment) that are asserted rather than verified for LLM outputs.
- Scaling beyond four Execution LLMs with one coordinator yields diminishing or negative returns, requiring hierarchical coordinators.
Experimental setup
Datasets: GSM8K GSM-Hard MATH SVAMP StrategyQA TravelPlanner
Benchmarks: GSM8K GSM-Hard MATH SVAMP StrategyQA TravelPlanner
Metrics: accuracy pass rate token usage
Baselines: zero-shot CoT few-shot CoT self-consistency tree of thoughts RAP ReAct rStar multi-agent debate TS-LLM PPO-MCTS
Builds on: multi-agent debate QMIX chain-of-thought prompting self-consistency
Closest papers by meaning
- Strat-Reasoner: Reinforcing Strategic Reasoning of LLMs in Multi-Agent Games (ICML 2026 · similarity 0.89)
- Reinforce LLM Reasoning through Multi-Agent Reflection (ICML 2025 · similarity 0.89)
- Unlocking the Power of Multi-Agent LLM for Reasoning: From Lazy Agents to Deliberation (ICLR 2026 · similarity 0.88)
- GTBench: Uncovering the Strategic Reasoning Capabilities of LLMs via Game-Theoretic Evaluations (NeurIPS 2024 · similarity 0.88)
- LLM Active Alignment: A Nash Equilibrium Perspective (NeurIPS 2026 · similarity 0.88)
- $\texttt{Multi}^2$: Hierarchical Multi-Agent Decision-Making with LLM-Based Agents in Interactive Environments (ICML 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.