Tracking the perspectives of interacting language models

By Hayden S. Helm, Brandon Duderstadt, Youngser Park, Carey E. Priebe. Conference on Empirical Methods in Natural Language Processing, 2024

17 citations recorded by the source corpus.

Large language models (LLMs) are capable of producing high quality information at unprecedented rates. As these models continue to entrench themselves in society, the content they produce will become increasingly pervasive in databases that are, in turn, incorporated into the pre-training data, fine-tuning data, retrieval data, etc. of other language models. In this paper we formalize the idea of a communication network of LLMs and introduce a method for representing the perspective of individual models within a collection of LLMs. Given these tools we systematically study information diffusion in the communication network of LLMs in various simulated settings.

Subfields: Large Language Models, LLM Agents & Tool Use

View Paper: https://arxiv.org/abs/2406.11938

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Large Language Models

In LLM Agents & Tool Use

Citations in the corpus

Jailbreak susceptibility prediction and mitigation via the behavioral geometry of models (2026, unknown): “The Data Kernel Perspective Space (DKPS) [10, 16, 1, 14] provides a low-dimensional Euclidean estimate of a behavioral geometry.”
Black-box model classification under the discriminative factorization (2026, unknown): “Recent work has shown that collections of black-box models can be represented in Euclidean space by comparing embedded responses to a set of queries [15, 1], and that these representations enable consistent model-level inference [16].”
Concentration bounds on response-based vector embeddings of black-box generative models (2025, unknown): “In particular, Helm et al. (2024b) discusses a novel technique for obtaining a vector representation for every generative model in a given collection of generative models, based on their responses to a set of queries, which can be used for further downstream inference tasks on the collection of…”
Gaussian mixture models as a proxy for interacting language models (2025, unknown): “In McGuinness et al., for example, [6] the time-varying output of interacting LLM agents is used to model two well-documented features of human social networks: first, human mirroring, in which individuals copy the behaviors of others, and second, social alignment, which is the emergence of a…”
LLM Web Dynamics: Tracing Model Collapse in a Network of LLMs (2025, unknown): “This can be challenging due to the high dimension of the embedding model, and one can adopt the framework of Data-Kernel Perspective Space (DKPS) (Helm et al., 2024), which reduces its dimension by Classical Multidimensional Scaling to get nice vector representations.”
Toward a digital twin of U.S. Congress (2025, unknown): “Further, simulations of interacting generative agents based on LLMs parameterized by tailored system prompts and data from previous interactions show resemblance to human social systems more generally (21; 12; 18).”
LLM as a Broken Telephone: Iterative Generation Distorts Information (2025, unknown): “These frameworks enable agents to iteratively refine outputs through debate-style interactions (Helm et al., 2024) or cooperative task decomposition (Pham et al., 2023), often improving accuracy in mathematical and logical tasks (Zhang et al., 2024).”
Investigating social alignment via mirroring in a system of interacting language models (2024, unknown): “We note that increasing the complexity of the agents will require using general black-box techniques (Duderstadt et al., 2024; Helm et al., 2024b; Acharyya et al., 2024; Helm et al., 2024a) to study agent, group, and system dynamics.”
Consistent estimation of generative model representations in the data kernel perspective space (2024, methodology): “For example, [12] uses the perspective space to demonstrate that the communication structure underlying a system of interacting language models has an impact on system-level and model-level evolution.”
When LLMs Play the Telephone Game: Cultural Attractors as Conceptual Tools to Evaluate LLMs in Multi-turn Settings (2024, unknown): “This idea of using LLM-generated content to fine-tune the next generation has also been applied to groups of LLMs with various communication structures (Helm et al., 2024).”
A Model of the Language Process (2026, unknown): “We call this curve the "temporal control curve," and conjecture that the model vectors constituting this curve can be applied to tasks including model comparison (Duderstadt et al., 2024), model level inference (Helm et al., 2025), and model perspective tracking (Helm et al., 2024).”
When LLMs Play the Telephone Game: Cumulative Changes and Attractors in Iterated Cultural Transmissions (2024, methodology): “This idea of using LLM-generated content to fine-tune the next generation has also been applied to groups of LLMs with various communication structures [33].”
Extracting information from fine-tuned weights (unknown): “For example, [11] introduces DKPS – a method that maps a collection of responses to a low-dimensional Euclidean subspace that can be used to track the dynamics of a population of interacting language models.”