The Big Picture
Agents that share a learned 'consensus' of what they see and only accept advice from peers with similar views learn faster, make better team decisions, and avoid copying bad guidance.
ON THIS PAGE
The Evidence
Learning a shared consensus representation from each agent’s local view helps agents judge when another agent’s advice is relevant. Agents dynamically pick who acts as a teacher, weigh suggested actions by how confident and similar the teacher is, and resample when advice looks unlikely. Across strategy and simulated-soccer benchmarks, this approach speeds up early learning, raises team win rates and scores, and reduces training instability compared with common decentralized baselines.
Data Highlights
1Evaluated on two standard benchmarks: StarCraft II Multi-Agent Challenge and Google Research Football (multiple maps/scenarios).
2All experiments report mean and standard deviation across 5 random seeds.
3Consensus model is a compact 3-layer neural network trained with contrastive learning as an auxiliary input to each agent.
What This Means
Engineers building cooperative multi-agent systems where centralized training is impractical — because this provides a lightweight way to share useful guidance without forcing blind imitation. Technical leaders evaluating agent trust or agent-to-agent evaluation can use consensus signals to measure when advice is likely to help. Researchers exploring decentralized training will find a practical plug-and-play component to improve coordination.
Not sure where to start?Get personalized recommendations
Key Figures
Fig 1: Figure 1
Fig 2: Figure 2
Fig 3: Figure 3
Fig 4: Figure 4
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreConsiderations
Results are demonstrated by adding the method to one base algorithm (independent Q-learning); benefits may vary with other learning algorithms. The approach depends on a communication channel and on agents having enough overlap in observations so a meaningful consensus can form. Hyperparameters (how consensus is trained, teacher selection criteria, resampling thresholds) and scaling to many heterogeneous agents were not fully explored and may require tuning.
Methodology & More
CCKS creates a shared, learned representation called a consensus from each agent’s local observation so agents can judge whether another agent is facing a similar situation. The pipeline has four stages: train a consensus model using contrastive learning (a compact three-layer network), broadcast an observation request when a student needs help, let potential teachers respond (teacher roles chosen dynamically using value estimates and entropy to favor competent, confident agents), and then filter and score suggested actions based on consensus similarity and uncertainty. A ‘think twice before you leap’ step makes agents resample actions when advice has low probability, preventing premature adoption of poor suggestions.
When plugged into a decentralized training and execution setup and combined with independent Q-learning, consensus-based advice led to faster early learning and more stable, higher final performance on both StarCraft micromanagement tasks and cooperative football simulations. Ablation tests show both the consensus representation and the resampling mechanism contribute to gains. The method is designed to be plug-and-play for decentralized systems where central training is impossible or costly, offering a practical way to improve multi-agent coordination while reducing blind imitation. Future work should validate the approach with other base learners, larger agent counts, and in settings with constrained or noisy communications.
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
All authors have low h-index or unspecified affiliations, published only on arXiv, and no citations — limited identifiable reputation or venue signals.