Tianyi Zhang

Research

Personalized language intelligence

My research asks how language models can better understand, adapt to, and simulate individuals or particular people rather than treating all users, speakers, and contexts with the same generic assumptions.

Topics

Current directions

Individualized Cognitive Simulation

Modeling the unique, deeper thought processes and decision behaviors of a specific human being.

Dialogue

Building interactive systems that reason over user goals, conversational history, preferences, and social context.

Token Compression

Studying compact representations that preserve the information models need while reducing context and computation costs.

Selected Work

Projects and papers

In Progress

Individualized Cognitive Simulation

This project studies whether language models can simulate the deeper thought processes and decision behaviors of specific individuals rather than only imitating surface-level style.

In Progress

Visual and token-efficient representations for NLP tasks

This project studies Vision-Text Augmentation (VTA), a complementary direction to vision-text compression where text is rendered into image space and varied through factors such as font, spacing, layout, and highlighting.

ACL 2026

GameplayQA: A Benchmarking Framework for Decision-Dense POV-Synced Multi-Video Understanding of 3D Virtual Agents

Yunzhe Wang, Runhui Xu, Kexin Zheng, Tianyi Zhang, Jayavibhav Niranjan Kogundi, Soham Hans, Volkan Ustun

GameplayQA evaluates multimodal models in fast-paced 3D virtual environments where agents must track self state, other agents, world events, and temporally synchronized perspectives. The benchmark is designed to expose failures in temporal grounding, role attribution, and cross-video reasoning. Project website.

Framework overview for GameplayQA

SIGDial 2026

DiPS: Dialogue Policy Selection for High-Stakes Persuasion Agents

Tianyi Zhang*, Mousumi Das*, Abrar Anwar, Jesse Thomason, David Traum (* Equal contribution)

DiPS studies adaptive policy selection for persuasion agents in high-stakes dialogue settings. The system selects a persuasion policy from recent conversation state, conditions generation on strategy descriptions and retrieved examples, and evaluates whether the interaction reaches a successful outcome.

Overview figure for DiPS dialogue policy selection

For the latest publication list, see my Google Scholar profile.