Main picture

Ph.D. Candidate

Department of Statistics
University of California, Los Angeles
Advisor: Ying Nian Wu and Tao Gao

Email: minglu.zhao@ucla.edu

Google Scholar   GitHub

Bio

I am a fourth-year Ph.D. student at UCLA, advised by Ying Nian Wu and Tao Gao. My research explores the intersections of language modeling, decision-making, representation learning, and human cognition, with a focus on developing generative models and latent variable approaches that enhance language understanding and improve decision-making processes in complex environments.

I obtained my B.S. in Statistics and B.S. in Cognitive Science also at UCLA. Go Bruins!! 🐻

 

News

 

Selected Publications

* denotes equal contribution.

Scalable Language Models with Posterior Inference of Latent Thought Vectors
, , , , , , , , , ,
Available in Preprint
We introduce Latent-Thought Language Models (LTMs), a novel language model family that incorporates explicit latent thought vectors. LTMs leverage dual-rate optimization, rapidly updating local latent vectors while gradually refining global decoder parameters. This approach unlocks new scaling dimensions, achieving superior efficiency, perplexity, and zero-shot performance over traditional models. They also exhibit emergent few-shot reasoning, highlighting their potential for advanced language tasks.
paper
Inverse Attention Agent in Multi-Agent System
, , , , ,
ICLR 2025
We introduce Inverse Attention Agents, leveraging Theory of Mind concepts through an attention mechanism to enable adaptability in dynamic multi-agent environments. These agents infer the goals and attentional states of other agents, refining their attention weights for improved decision-making. Tested across cooperative, competitive, and mixed tasks, our approach enhances performance and human-like cooperation compared to conventional models.
paper
A Minimalistic Representation Model for Head Direction System
, , , ,
In NeurIPS Workshop on Symmetry and Geometry in Neural Representations, 2024.
We present a model for the head direction (HD) system that captures essential HD cell properties through a high-dimensional U(1) representation. This model reveals Gaussian-like tuning and 2D circular geometry, accurately supporting path integration in both fully connected and convolutional forms.
paper
Latent Plan Transformer: Planning as Latent Variable Inference
, , , , , , , ,
NeurIPS 2024
We introduce the Latent Plan Transformer (LPT), a novel model that leverages a latent space to connect a Transformer-based trajectory generator and the final return. This architecture enables planning without step-wise rewards, addressing temporal consistency challenges in long-term tasks. LPT uses maximum likelihood estimation on trajectory-return pairs, with posterior sampling of latent variables for consistent sub-trajectory abstraction. During inference, LPT deduces the latent variable based on expected returns, realizing a planning-as-inference approach.
paper website
Intention beyond desire: Spontaneous intentional commitment regulates conflicting desires
, , , , , , ,
In Cognition, 2023.
We explore how coherent actions emerge from conflicting desires, contrasting classical desire-driven behavior with intention-driven action. Through 2D navigation games, we identify three unique markers of human intentional commitment—goal perseverance, self-binding, and temporal leap—that distinguish human actions from purely desire-driven agents. Our findings suggest that humans form committed intentions to manage conflicting desires, enhancing predictability and reducing computational load in action planning.
paper
Sharing rewards undermines coordinated hunting
, , , , , ,
In Journal of Computational Biology, 2022.
We investigate the impact of reward sharing on coordinated hunting using Multi-agent Reinforcement Learning (MARL) and reveal surprising findings: rather than facilitating coordination, sharing rewards undermines it due to issues like the free-rider problem and coordination limits at larger group sizes. Individually rewarded agents outperform those sharing rewards, particularly in challenging scenarios. Our results suggest that reward sharing may not be crucial for animal coordination, challenging assumptions in AI models that rely on shared rewards to motivate group coordination.
paper website
Exploring an imagined “we” in human collective hunting: Joint commitment within shared intentionality
, , , , , , ,
In CogSci, 2022.
We examine human collaboration in goal selection, demonstrating that shared intentionality allows humans to form robust commitments to collective goals without communication. In a real-time cooperative hunting game, humans maintained high-quality cooperation, even with many targets. We develop a Bayesian "Imagined We" (IW) model which mirrored this behavior, outperforming a Reward Sharing (RS) model, which struggled with coordination as target numbers rose. These findings highlight shared intentionality as central to human cooperation, offering insights into its computational basis.
paper
Modeling communication to coordinate perspectives in cooperation
, , , , , , ,
In CogSci, 2021.
We introduce the Imagined We for Communication framework, a model where agents leverage shared agency to interpret overloaded signals in ambiguous contexts. By simulating rational cooperators, our model demonstrates strong performance in high-ambiguity settings, even with minimal reasoning depth, underscoring how shared knowledge and cooperative logic support effective communication.
paper
Bootstrapping an Imagined We for Cooperation
, , , , ,
In CogSci, 2020.
We develop a Bayesian-Theory-of-mind-based framework named the Imagined We (IW), showing how agents can reliably converge on a joint intention in uncertain, multi-choice settings through bootstrapping. In a real-time cooperative hunting task, our model proves resilient to challenges like numerous choices, approximate partner models, and noisy perceptions, highlighting its robustness in maintaining joint commitment under imperfect conditions.
paper

 

Teaching