This repository accompanies the paper
"A Dynamic Approach to Collaborative Document Writing", accepted for publication at ECAI 2025 (European Conference on Artificial Intelligence, Bologna, Italy).
It provides the extended version and supplementary materials for the paper, including detailed proofs, additional experiments, and the full simulation framework code.
-
agents/
Contains implementations of the threeAgenttypes used in the study: unstructured, Euclidean, and LLM-based agents. -
datasets/
Input data for population profiles, including demographic distributions and a synthetic dataset on climate actions. -
events/
Defines theEventandEvent Listclasses and handles the generation of event lists used to simulate sequential agent actions. Includes all the functions discussed in the paper. -
figures/
Output directory for all figures produced during simulation, including Pareto frontiers, coverage plots, and comparative analyses. -
paragraphs/
ImplementsParagraphobjects used as candidate proposals in the document. Includes basic and interval paragraph types. -
results/
Stores all simulation results, intermediate outputs, and serialized objects generated during runs. Due to space limitations, only community information and figures are included. -
rules/
Contains rule configuration files (static and dynamic CSFs). -
simulation/
Core logic forSimulation Configuration,Scheduler, metrics evaluation and aggregation. -
tests/
Unit tests and validation scripts for major components in the pipeline.
-Event_List_Example_5/ The scheduled event list presented in the paper as example 5 with JSON files describing its agents, events, and paragraphs. Note that there is a log file of API calls and responses.
To run the main simulation pipeline, we used parallel computation with two scripts, one for LLM agents and another for the rest.
If you use this code or data, please cite:
Avital Finanser and Nimrod Talmon.
A Dynamic Approach to Collaborative Document Writing.
Proceedings of ECAI 2025.