At a Glance
Multi-agent systems behave like a “pick-the-expert” team: which agent wins is mostly driven by relative confidence, so success depends on having diverse, well-calibrated agents whose confidence reflects real competence.
ON THIS PAGE
What They Found
Deliberation among multiple language model agents can be modeled as a mixture-of-experts: the system implicitly routes weight to specific agents depending on the input. The strongest single predictor of who becomes influential is an agent’s mixture-of-experts confidence relative to peers, with initial alignment to the group and communication style also affecting influence. Influence often concentrates on one or a few agents (rather than averaging everyone), so the MAS gains only when those chosen agents are locally competent and complementary.
By the Numbers
1Predicting influence from observable signals: regression R² = 0.7 and classification accuracy = 0.9 (models trained to predict which agent becomes most influential).
2Experimental setup used 5 agents communicating for 5 rounds across datasets: MMLU-Pro (300 questions), BBQ (300 questions), and CommonsenseQA (100 questions).
3Influence typically concentrates on 1–2 agents out of 5, showing strong local specialization rather than broad averaging.
What This Means
Engineers building multi-agent systems and orchestration layers — so you can prioritize agent diversity and agent-to-agent evaluation evaluation rather than adding more agents. Technical leads and evaluators running continuous agent monitoring — to know which signals (relative confidence, alignment, communication style) to log and trust. Researchers studying agent governance and trust can use the Friedkin–Johnsen view to analyze routing and failure modes.
Not sure where to start?Get personalized recommendations
Key Figures
Fig 1: (a) Parameter variability. W: weighted in-degrees (avg. over senders).
Fig 2: Figure 3 : Example: Initial beliefs of agents (left). FJ weight matrix, γ \gamma , and color coded belief in correct answer (right). The most confident agent convinces the majority to change their answer.
Fig 3: (a) MMLU-Pro
Fig 4: (a) MMLU-Pro
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreKeep in Mind
Findings assume the agents’ initial beliefs and communicated confidence calibration reflect latent competence; if confidence is miscalibrated the router can pick the wrong expert. Experiments used a small set of models, agent roles, and datasets, so generalization to other tasks or larger agent pools is not guaranteed. The model explains influence emergence but does not prove that a distilled single model couldn’t match or surpass a tuned multi-agent system in some settings.
Methodology & More
Modeling agent deliberation with a classical opinion-dynamics model shows that multi-agent language-model systems act like a mixture-of-experts: the final answer is a weighted average of agents’ initial beliefs, where the weights depend on input-specific interaction parameters. Those interaction parameters vary by question, so the system implicitly routes influence toward particular agents rather than holding fixed ensemble weights. The routing favors agents who display higher confidence relative to peers, but alignment with the group and an agent’s communication style also shift influence.
Empirical tests with two language models and three benchmarks (MMLU-Pro, BBQ, CommonsenseQA) used 5 agents over 5 rounds. Fitting the opinion-dynamics model to run traces showed input-dependent weights and frequent concentration of influence on only one or two agents. Predictive models using observable signals (confidence, alignment, peer-influence, prompts) explained most of the variance in influence (R² ≈ 0.7, classification accuracy ≈ 0.9). The practical takeaway: multi-agent gains come from adaptive routing and local specialization — but only when agents are diverse and their confidence is a reliable sign of competence. Otherwise, routing errors or misleading consensus can hurt performance. adaptive routing and local specialization can be important design choices.
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
Authors include researchers with mid-range h-indexes (10–20) but no top-tier venue or affiliations listed — recognized but not top.