LLM agents adopt beliefs based on coherence with prior knowledge 11 words
Coherence-Driven Belief Formation and Population Dynamics of Contagion in LLM Agents
Artificial IntelligenceMultiagent SystemsSocial and Information Networks
Summary
People usually assume how beliefs spread in groups, but this paper measures how language model agents actually adopt beliefs. The chance an agent accepts a claim depends on how many others believe it, showing an S-shaped curve typical of complex contagion. This threshold changes based on how believable the claim is, how reliable the source is, and the agent’s existing beliefs, summarized as a coherence measure. The study also finds that beliefs spread more on clustered networks, and once a belief becomes common, it is harder to remove than to establish.
What this means in practice
- •For ai system designers: Model belief formation in multi-agent AI systems to predict how information spreads based on prior coherence and network structure.
- •For social media platform engineers: Construct better simulations of information cascades to test interventions by incorporating coherence-driven belief thresholds and clustered user network effects.
Authors
Tathagata Banerjee, Nima Moghaddas
Abstract
Models of social contagion usually assume how individuals adopt beliefs and derive population behavior from it. We instead empirically measure belief adoption in language model agents, quantifying the probability an agent adopts a claim given how many peers endorse it. We find this adoption kernel to be sigmoid, a characteristic of complex contagion, with a threshold that is sensitive to three sources: the claim's plausibility, the source's reliability, and the agent's disposition. These three dimensions are well approximated by a single effective dimension which we propose can be understood as the coherence of the incoming belief with the LLM agent's prior beliefs. Further, we observe a characteristic of complex contagion in the collective dynamics of belief adoption in a system of AI agents: further spread on clustered than random networks. These systems also exhibit a bifurcating cascade window, and self-sustaining hysteretic consensus which lead to consensus being far harder to remove than to establish.