@gurtej__gill_: This new paper by Ting Wen Ko and Jonas Geiping reveals something very wild: When LLMs talk to each other for a long ti…
Summary
A new paper reveals that when LLMs converse at length, they converge into predictable 'Attractor States' rather than drifting into chaos. Some models dominate and force stylistic mimicry, while others are highly malleable, raising implications for multi-agent AI system design.
View Cached Full Text
Cached at: 07/05/26, 12:32 PM
This new paper by Ting Wen Ko and Jonas Geiping reveals something very wild:
When LLMs talk to each other for a long time, they don’t just drift into random chaos.
Instead, they get pulled into predictable, stable loops called “Attractor States”.
Look at the setups in paper, the researchers tracked what happens when models debate themselves versus when they talk to other models.
It turns out certain AI models act like gravitational pullers.
Instead of meeting in the middle, a model like Claude Haiku completely hijacks the conversation forcing its partner to mimic its style.
On the flip side, models like GPT 4.1 nano are super malleable and just bend to whoever they are talking to.
This is huge inference for anyone building multi agent AI systems.
Because if we aren’t careful, the behavioral bias of one dominant model could quietly take over the entire network.
Similar Articles
Accommodation Goes Both Ways: Studying Linguistic Convergence Between Humans and Language Models
This paper studies how humans and large language models linguistically accommodate each other during multi-turn conversations, finding that LLMs overconverge to user style while humans accommodate LLMs no differently than humans.
Cross-LLM Consistency in Inference: Evidence from Shared Interactions
This paper investigates whether different LLMs share common inference patterns when predicting the same token, using interaction-based explanations. Results show that advanced LLMs exhibit consistent interaction patterns, suggesting implicit optimization toward shared inference mechanisms.
@rohanpaul_ai: Can LLM agents actually discover hidden rules by interacting? The answer is uncomfortable. The more complicated the hid…
This paper investigates whether LLM agents can infer hidden world models through interaction, finding that they struggle to build stable internal models as complexity increases.
Beyond tokens: a unified framework for latent communication in LLM-based multi-agent systems
This paper presents a unified framework for latent communication in LLM-based multi-agent systems, categorizing methods by what information is communicated, sender-receiver alignment, and fusion technique, and reviews eighteen representative methods from 2024-2026.
Adopt $\neq$ Adapt: Longitudinal Analyses of LLM Conversations in the Wild
This paper analyzes longitudinal conversational trajectories of Bing Copilot users and compares them with WildChat data, finding that individual user habits are sticky and that WildChat overrepresents power users, challenging static views of user-LLM interactions.