[2I1-OS-49] OS-49

9:45 AM - 10:00 AM JST(12:45 AM - 1:00 AM UTC)

[2I1-OS-49-04] Context-Dependent Parameter Modulation and Phase Transition-like Behavior in Social Learning among Large Language Model Agents

〇Kazuya Horibe¹, Masto S Abe^2,1, Takahiro Ezaki³, Peter Romero^4,5, Wataru Toyokawa¹ (1. RIKEN, 2. Doshisha University, 3. The University of Tokyo, 4. Universitat Politècnica de València, 5. University of Cambridge,)

Keywords:

Multi LLM Agents,Social Learning,Critical Dynamics

The potential impact of autonomous AI agents on information ecosystems and social processes has become a topic of increasing concern. To characterize collective dynamics among large language model (LLM) agents, we deployed GPT-5.2 and Claude-Sonnet-4.5 in a social learning task (multi-armed bandit with environmental change). RL parameter estimation revealed distinct profiles: GPT exhibited rapid annealing and strong conformity, leading to collective lock-in, while Claude maintained exploration. Mixed-agent experiments showed a sharp threshold effect at the 2:3 GPT:Claude ratio. Parameter re-estimation revealed asymmetric context-dependent modulation: GPT's shifts were extreme and systematic across all ratios, while Claude remained near baseline at the cascade threshold. A replay analysis showed that absorbing these context-dependent responses into individual RL parameters worsened predictions, suggesting dynamic strategy adjustment beyond fixed-parameter RL. These findings highlight group composition as a critical design variable for multi-agent LLM systems.