Skip to content

Active research

LLM-driven Social Simulation

Designing executable environments that turn generated social actions into consistent and replayable world-state transitions.

LLM AgentsSocial SimulationEvaluation

Overview

This line of work asks how concurrent agent actions should settle when several individually valid proposals interact with the same world state.

Motivation

Natural-language plausibility is not enough for a simulation. The environment must define what happened, preserve constraints, make progress measurable, and support exact replay.

Method

The current implementation uses typed snapshots and explicit settlement policies. Audits separate order sensitivity, useful progress, and replay consistency so that different failure modes remain visible.

Results

The first public audit evaluates five settlement policies through exhaustive permutation trials and scripted multistep episodes. See the linked paper for the complete experimental protocol and reported results.