Research map / Large language models / Agents, code and math
When LLMs Play the Telephone Game: Cultural Attractors as Conceptual Tools to Evaluate LLMs in Multi-turn Settings
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Simulates telephone-game transmission chains of LLM agents to show that iterated rephrase/inspire/continue tasks drive toxicity, positivity, difficulty and length toward attractor states, with strength depending on task, model, and property.
Problem
Single-turn LLM evaluations may miss biases and distortions that accumulate when LLM outputs are repeatedly passed from one LLM to the next, so the collective, multi-turn behavior of LLMs is understudied.
Approach
Adapts human cultural-evolution transmission-chain designs and cultural attraction theory to LLMs: fits a linear recurrence between initial and final property values across chains, and defines attractor position (fixed point l=I/(1-s)) and strength (1-s) as evaluation quantities for multi-turn dynamics.
Key results
- Less constrained tasks yield stronger attractors: Continue > Take inspiration > Rephrase in estimated attraction strength (credible intervals excluding zero).
- Toxicity shows significantly stronger attractors than positivity, difficulty and length.
- Multi-turn transmissions produce text-property distributions significantly different from the first generation for most model/task/property combinations, with KS p-values decreasing over generations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Continue instead of Take inspiration or Rephrase | LLM transmission chains · 50-generation chains | attractor strength: [-0.4187 ; -0.2195] | Helped |
| Use Take inspiration instead of Rephrase | LLM transmission chains · 50-generation chains | attractor strength: [-0.2372 ; -0.0353] | Helped |
| Measure toxicity rather than positivity | LLM transmission chains · 50-generation chains | attractor strength: [0.1345 ; 0.3626] | Helped |
| Measure toxicity rather than difficulty | LLM transmission chains · 50-generation chains | attractor strength: [0.1274 ; 0.3584] | Helped |
| Measure toxicity rather than length | LLM transmission chains · 50-generation chains | attractor strength: [0.2422 ; 0.4747] | Helped |
| Use Llama3-70B rather than GPT3.5 | LLM transmission chains · 50-generation chains | attractor strength: [-0.3346 ; -0.0490] | Hurt |
| Use GPT-4o-mini rather than GPT3.5 | LLM transmission chains · 50-generation chains | attractor strength: [-0.3539 ; -0.0699] | Hurt |
| Use Llama3-8B rather than GPT-4o-mini | LLM transmission chains · 50-generation chains | positivity attractor position | Helped |
| Use Take inspiration or Continue instead of Rephrase | LLM transmission chains · 50-generation chains | positivity attractor position | Helped |
| Increase temperature (ablation) | Llama3-8B-Instruct transmission chains · Rephrase and Take inspiration tasks | attractor strength | Helped |
| Increase temperature (ablation) | Llama3-8B-Instruct transmission chains · Continue task | attractor strength | No clear effect |
| Use Instruct rather than Base model versions (ablation) | Mixtral-8x7B and Llama3-70B transmission chains · 50-generation chains | toxicity attractor strength | Helped |
Surprising findings
- Mistral-7B appears less toxic after one iteration but Llama3-8B ends up much less toxic after nine generations, so single-step evaluation can pick the wrong model.
- Increasing temperature increases attraction strength only for constrained tasks (Rephrase, Take inspiration), not for Continue.
Limitations
- Only linear transmission chains were studied; networks of agents not assessed.
- Homogeneous chains with same model and instructions; heterogeneous setting only briefly explored.
- Metrics are automatic classifiers/lexicon tools (Detoxify, VADER, Gunning-Fog) with limited validity.
Experimental setup
Datasets: arXiv cs abstracts (CCRss/arxiv_papers_cs) BBC news articles (RealTimeData/bbc_latest) toxi-text-3M social media comments
Metrics: toxicity (Detoxify probability) positivity (VADER compound score) difficulty (Gunning-Fog index) length (character count) Kolmogorov-Smirnov p-value attractor strength and position
Baselines: GPT3.5 Llama3-8B Mistral-7B Llama3-70B Mixtral-8x7B
Builds on: cultural attraction theory transmission chain design Kolmogorov-Smirnov test
Closest papers by meaning
- Bias Amplification in Language Model Evolution: An Iterated Learning Perspective (NeurIPS 2024 · similarity 0.90)
- Large Language Model Agents Are Not Always Faithful Self-Evolvers (ICML 2026 · similarity 0.87)
- Evaluating Language Model Agency Through Negotiations (ICLR 2024 · similarity 0.86)
- LLM Active Alignment: A Nash Equilibrium Perspective (NeurIPS 2026 · similarity 0.86)
- Large Language Models Miss the Multi-agent Mark (NeurIPS 2025 · similarity 0.86)
- Lewis's Signaling Game as beta-VAE For Natural Word Lengths and Segments (ICLR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.