Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SAUCE: Synchronous and Asynchronous User-Customizable Environment for Multi-Agent LLM Interaction

Shlomo Neuberger, Niv Eckhaus, Uri Berger, Amir Taubenfeld, Gabriel Stanovsky, Ariel Goldstein

TL;DR

SAUCE: a customizable Python platform, allowing researchers to plug-and-play various LLMs participating in discussions on any topic chosen by the user, and a novel feature of SAUCE is its asynchronous communication feature, where models decide when to speak in addition to what to say, thus modeling an important facet of human communication.

Abstract

Many human interactions, such as political debates, are carried out in group settings, where there are arbitrarily many participants, each with different views and agendas. To explore such complex social settings, we present SAUCE: a customizable Python platform, allowing researchers to plug-and-play various LLMs participating in discussions on any topic chosen by the user. Our platform takes care of instantiating the models, scheduling their responses, managing the discussion history, and producing a comprehensive output log, all customizable through configuration files, requiring little to no coding skills. A novel feature of SAUCE is our asynchronous communication feature, where models decide when to speak in addition to what to say, thus modeling an important facet of human communication. We show SAUCE's attractiveness in two initial experiments, and invite the community to use it in simulating various group simulations.

SAUCE: Synchronous and Asynchronous User-Customizable Environment for Multi-Agent LLM Interaction

TL;DR

SAUCE: a customizable Python platform, allowing researchers to plug-and-play various LLMs participating in discussions on any topic chosen by the user, and a novel feature of SAUCE is its asynchronous communication feature, where models decide when to speak in addition to what to say, thus modeling an important facet of human communication.

Abstract

Many human interactions, such as political debates, are carried out in group settings, where there are arbitrarily many participants, each with different views and agendas. To explore such complex social settings, we present SAUCE: a customizable Python platform, allowing researchers to plug-and-play various LLMs participating in discussions on any topic chosen by the user. Our platform takes care of instantiating the models, scheduling their responses, managing the discussion history, and producing a comprehensive output log, all customizable through configuration files, requiring little to no coding skills. A novel feature of SAUCE is our asynchronous communication feature, where models decide when to speak in addition to what to say, thus modeling an important facet of human communication. We show SAUCE's attractiveness in two initial experiments, and invite the community to use it in simulating various group simulations.

Paper Structure

This paper contains 19 sections, 4 figures.

Table of Contents

  1. Introduction
  2. SAUCE*X: Highlights
  3. SAUCE*X: Architecture
  4. Experiment
  5. Session Room
  6. Host
  7. HostRoundRobin
  8. HostRandom
  9. Person
  10. Implemented Person Classes
  11. Asynchronous Communication
  12. Implemented Asynchronous Persons.
  13. EndType
  14. EndTypeNumMsgs
  15. Evaluation: Case Studies
  16. ...and 4 more sections

Figures (4)

  • Figure 1: Illustration of a discussion between different agents, run on SAUCE*X. Our framework allows setting up a discussion topic, and then manages the group discussion by instantiating models and scheduling their responses.
  • Figure 2: Example JSON configuration file setting up all the required objects for a multi-agent discussion.
  • Figure 3: Hierarchy of classes in SAUCE*X, including section numbers with detailed descriptions. All classes inheriting from AsynchronousPerson are described under its section.
  • Figure 4: Illustration of several messages that were sent as part of an asynchronous group discussion, run on SAUCE*X. This experiment was run with acknowledgment of the current time and a time limit for the discussion. In bold: two participants (Joseph and Jennifer) chose not to speak, driving a third participant (Amanda) to request their opinion (in bold); a fourth participant (Robert) chose to speak again, after noticing the others kept silent.