Mingkai Deng
mingkaid [at] cs [dot] cmu [dot] edu

I am a PhD student at the Language Technologies Institute at Carnegie Mellon University, working with Prof. Eric P. Xing. I also worked closely with Prof. Zhiting Hu.

My research centers on world models and agent models as complementary foundations for general-purpose intelligence. World models are generative simulators of all actionable possibilities across physical, social, and conceptual domains, letting agents run thought experiments and reason about counterfactual futures. Agent models internalize agency (e.g., hierarchical goals, an evolving identity, world-model-based planning, self-regulation, and self-directed learning) rather than relying on externally defined structure, with the world model as the substrate enabling the rest.

I completed my MS at CMU's Machine Learning Department. Before that, I worked as a data scientist in industry following an internship at Baidu Research, where I was advised by Prof. Hui Xiong. I earned my undergraduate degree from Columbia University, with a double major in Mathematics-Statistics and Computer Science.

CV (very outdated)  /  LinkedIn  /  Twitter  /  GitHub

• 2026-05 Our work "Efficient Agentic Reasoning Through Self-Regulated Simulative Planning " has been released!
• 2026-05 Excited to return to MBZUAI IFM for internship, training better world models
• 2025-11 Honored to have been part of the team project "PAN: A World Model for General, Interactable, and Long-Horizon World Simulation"
• 2025-07 Our position paper "Critiques of World Models" is released on Arxiv.
• 2025-05 Started internship at MBZUAI IFM working on training agent models with RL
• 2025-02 Our work "General Agentic Planning Through Simulative Reasoning with World Models" won 2nd place in the Fundamental Track at Berkeley's LLM Agents Hackathon.

This position paper critiques several schools of thoughts on world modeling, and proposes a new architecture for a general-purpose world model, with an outlook of a Physical, Agentic, and Nested (PAN) AGI system enabled by such a model.

A model of efficient agentic reasoning as simulative reasoning via a world model (System II), self-regulation of planning (System III), and reactive execution (System I). Resulting 30B model SR²AM-v1.0 competes with 685B–1T LLMs while using 25.8–95.3% fewer reasoning tokens than comparable agentic LLMs.

A general reasoning paradigm for agentic planning via world model (System II), rather than just a reactive policy (System I). Resulting agent architecture (SiRA) leads to 124% higher success rate than reactive baseline. Success rate on constrained navigation: 0% → 32.2% vs a representative open-web agent.

A general world model that predicts future world states through high-quality, action-conditioned video simulation. Built on the Generative Latent Prediction (GLP) architecture that combines autoregressive latent dynamics with language grounding, PAN excels at long-horizon forecasting and simulative reasoning.

A 65-billion parameter, 360-degree open-source LLM that surpasses LLaMA-65B and rivals LLaMA2-70B, while requiring fewer FLOPs and tokens.

An efficient and flexible framework for using RL to optimize prompts of discrete text that enable pre-trained LMs (e.g., BERT, GPT-2) to perform diverse NLP tasks. Experiments on few-shot classification and unsupervised text style transfer show superior performance to a wide range of existing methods.

The first Embodied Question Answering benchmark that methodically evaluates how well Vision-Language Models bridge the semantic–physical gap—translating high-level instructions into low-level physical actions. ActionEQA reveals a large gap between VLM and human performance (58.4% vs. 95.6%).

A multimodal conversational system that blends natural language and visual interaction for debugging hierarchical object-classification models, letting users express high-level debugging goals while adaptive explanations surface only the most task-relevant visual or textual details.

A large-scale webpage-to-code dataset which enhances the webpage understanding and HTML code translation abilities of MLLMs without compromising general visual capabilities.

A general framework that helps solve the difficulty of evaluating natural language generation (NLG) with a single unified operation. Inspired evaluation metrics improve over SOTA metrics for diverse NLG tasks. Our metrics are available as library on PyPI and GitHub.

An open-source web agent that operates in a Chromium-based browser and can perform a broad range of realistic tasks that require complex, long-range, and goal-achieving behavior, such as searching for flights, compiling online shopping options, and researching news coverage.

A data-driven narrative of NYC gentrification patterns over time and their second-order impact on the city's inhabitants. Construct significant leading predictors using resources from NYC OpenData.

Combine structured point-of-interest (POI, e.g., parking lots and buildings) data with unstructured satellite image data to automatically identify and draw the boundaries of urban areas-of-interest (e.g., residential areas and campuses).

Formulate the Transformer model as an ordinary differential equation (ODE) and perform forward-backward operations using an ODE solver. The resulting continuous-depth model has fewer parameters, converges faster, and shows promising performance in empirical experiments.

The source code of this website is adapted from here