Human-Agent Interaction and Simulation

Simulated users, AI patients, speech interaction, financial agents, and micro-world environments for studying how agents behave with people and worlds.

Human-agent interaction
User simulation AI standardized patients Multi-agent markets Micro-world simulation Speech interaction
Human-agent and market simulation

Human-agent interaction is the connective tissue between model capability and real deployment. This project line studies agents that interact with people, learners, patients, investors, speech partners, and simulated worlds, where behavior unfolds over time rather than in a single static prompt.

Background and Motivation

Static benchmarks miss interaction

Many failures only appear after several turns: users change goals, agents adapt, environments respond, and earlier decisions constrain later ones.

Human-facing agents need social realism

Medical learners, investors, patients, and speech users bring beliefs, emotions, incentives, and incomplete information into the loop.

Simulation creates safe testbeds

Before deploying agents in clinics, markets, classrooms, or social settings, we can study them in controlled and instrumented environments.

Research Storyline

User
Simulate users for dialogue learning

LLM user simulators and Socratic questioning systems generate interactive training environments for multi-turn dialogue.

Patient
Turn AI into a medical learning partner

AI standardized patients support repeatable practice in history taking, communication, and clinical reasoning.

Society
Study multi-agent social and market behavior

TwinMarket and Economic World Models use agent societies to examine emergent markets, policy sandboxes, and collective risk.

World
Build environments where state evolves

MicroVerse extends interaction to scientific worlds, where agents must reason about hidden mechanisms and changing biological states.

Project Clusters

Dialogue and user simulation

Large Language Model as a User Simulator and PlatoLM study how simulated users can teach, probe, and improve dialogue models.

Medical education agents

EasyMED and AI standardized patients connect human-agent interaction with medical training and learner-centered evaluation.

Agentic markets and economies

TwinMarket and Economic World Models make multi-agent interaction observable at social, financial, and policy scale.

Speech and embodied interaction

S2S-Arena, EchoMind, and speech-to-speech Turing tests examine whether agents can interact naturally through voice and paralinguistic signals.

Typical Work

Dialogue
Large Language Model as a User Simulator

Uses LLMs to simulate users and generate interactive training or evaluation data for dialogue systems.

Paper
Learning
PlatoLM and Socratic user simulation

Studies Socratic questioning as a way to teach multi-turn dialogue behavior through simulated interaction.

Paper
Markets
TwinMarket

Builds a scalable behavioral and social simulation for financial markets with LLM-based investor agents.

Project page
Worlds
MicroVerse

Explores micro-world simulation where visible phenomena depend on hidden scientific mechanisms and evolving states.

Project page

Display Figures

Resource Map

Economic World Models

Umbrella project for agentic economies, policy sandboxes, and self-correcting economic twins.

Project page
Medical Education and AI Patients

AI standardized patient project for repeatable and benchmarked medical learning scenarios.

Project page
Speech, Audio, and Talking-Head AI

Voice, paralinguistic, empathy, and embodied interaction benchmarks and datasets.

Project page