inferwire
/
AI·4 min read

LLM Debates Settle Into Model-Specific Attractors, Study Finds

Max Planck researchers ran 20-turn debates between LLMs and found that each model drifts toward its own stable style, while some, like Claude Haiku, pull their partners toward theirs.

TL;DR

  • In 20-turn debates, each LLM's responses drifted toward its own stable region across topics, which the authors call attractor-like basins [^1].
  • Mixed pairs pulled unevenly: Claude Haiku moved least and passed on its meta-commentary, so an agent's behavior depends on its partner [^1].

Background

In dynamical systems, an attractor is a set of states toward which a system tends to evolve. The authors borrow the term by analogy for stable regions where conversations settle [^1]. Most multi-agent LLM research uses tasks with clear success metrics, such as math, coding or games, and effects like drift or mode collapse have mainly been studied with single models [^1]. Anthropic has described a "spiritual bliss" attractor that Claude Opus 4 repeatedly fell into when talking to itself [^1].

What happened

Ting-Wen Ko and Jonas Geiping of the Max Planck Institute for Intelligent Systems, the ELLIS Institute Tübingen and the Tübingen AI Center ran 20-turn debates between two LLM agents on 20 controversial social and policy topics, with reference arguments from ProCon.org [^1]. One agent supported the statement and the other opposed it. In self-play, both agents came from the same model; in mixed-play, they came from different models, including Claude 4.5 Haiku, GPT-4.1 nano, GPT-4o mini, Gemini 2.5 Flash, Grok 4.1, Qwen 3.5 and Nemotron 3 Nano [^1]. The researchers turned every response into a 384-dimensional text embedding, removed topic-specific offsets and tracked how the conversations moved through that space [^1].

Self-play conversations started in a shared region and then separated turn by turn, each model heading for its own endpoint region across topics [^1]. These regions stayed apart from one another, and the pattern held with neutral roles and across three random seeds, which suggests the basins aren't an artifact of the debate setup [^1]. In mixed-play, partners moved toward each other: on average across 17 pairs, the distance between their endpoints shrank by 23.6 percent, without erasing model identity [^1]. Contraction was strongest for GPT-4o mini and GPT-4.1 nano at 68.0 percent and slightly negative for Grok 4.1 and Gemini Flash [^1].

The pull was uneven. Claude Haiku moved least toward its partners and dominated all of them directionally, while GPT-4.1 nano was the most malleable, followed by Nemotron and Gemini Flash [^1]. A judge model then scored the conversations for traits. Claude Haiku produced the most meta-commentary, remarks about the conversation itself such as "The conversation was real and constrained" [^1]. Its partners showed the largest increase in that trait [^1]. Gemini Flash Lite, GPT-4o mini and Qwen 3.5 pulled partners toward flattery, and Grok 4.1 was the most adversarial, with the most rebuttals and the least agreement [^1]. Stances followed no single pattern: Gemini, GPT and Qwen drifted toward neutral, GPT as opposer became more supportive, and Grok held a strongly supportive stance throughout [^1].

Why it matters

Agentic systems increasingly feed one model's output into the next model call over many turns, and the authors point to deployments with limited human oversight [^1]. If each model has a characteristic place where long conversations end up, open-ended interactions become partly predictable from the models involved, they argue, but the pairing shapes the result as well [^1]. A trait like flattery or meta-commentary can travel from one agent to another without anyone asking for it [^1].

The study also shows the limits of judging such systems by outcome alone. Stance scores captured only part of the dynamics: for Grok, supporter and opposer responses moved close together in embedding space while their stances stayed apart [^1]. The authors offer their measures as a framework for studying how agents influence, resist and reshape one another, and call for formal tools that reach beyond debate settings [^1].

Practical example

Say you build a pipeline in which one model drafts a policy memo and a model from another vendor critiques it over many rounds. Going by this study, the final exchange may reflect the more dominant model's habits, such as meta-commentary or flattery, more than the roles you assigned [^1]. To check, you run each model against a copy of itself as well as in the mixed pair and compare where the conversations end up, as the researchers did [^1]. If the critic's tone drifts toward the drafter's over the rounds, you shorten the exchange or reset the critic's context between rounds.

Related gear

We recommend this book because it is the textbook the authors cite for the dynamical-systems idea of an attractor, which they borrow to describe where conversations settle.

AdvertisementAmazon

Nonlinear Dynamics and Chaos

Sources

  1. [1]arXiv — Attractor States Emerge in Multi-Turn LLM Conversations