Profile picture

PhD student at MIT CSAIL


· AI for math and science
· Neurosymbolic reasoning
· PL for LLM agents

E-mail: zli11010@mit.edu     X (Twitter): @zli11010
Google Scholar       LinkedIn       GitHub

Welcome! I’m Zed, a first-year PhD student at MIT CSAIL advised by Prof. Armando Solar-Lezama. Previously, I completed my undergrad and MEng at MIT, majoring in computer science and physics. My main research interests are machine learning and formal methods for math and science. I am also interested in programming frameworks for LLM-based agents. I am generously supported by the MIT Presidential Fellowship.

Publications

, , ,
EnCompass: Enhancing Agent Programming with Search Over Program Execution Paths
In NeurIPS, 2025.
EnCompass is an inference-time strategy framework for LLM-based agents written in Python. In EnCompass, "branchpoint()" causes the program's execution to split into multiple parallel branches of execution, and EnCompass searches over the resulting tree of possible execution paths of the program. By providing a unifying framework for interence-time strategies for AI agents, EnCompass enables easy experimentation of different scaling strategies and facilitates the discovery of better scaling laws.
pdf
, ,
When Do Skills Help Reinforcement Learning? A Theoretical Analysis of Temporal Abstractions
In ICML, 2024.
Focusing on deterministic sparse-reward environments, we show theoretically and empirically that RL performance gain from skills (temporal abstractions) is worse in environments where solutions to states are less compressible (in the information-theoretic sense). Further theoretical results suggest that skills benefit exploration more than they benefit learning from existing experience, and that using unexpressive skills such as macroactions may worsen RL performance.
pdfarXivposterGitHub
, , , ,
Predictive Chemistry Augmented with Text Retrieval
In EMNLP, 2023.
TextReact augments predictive chemistry with texts retrieved from the literature. For a given chemistry input, the retrieval model retrieves relevant texts, and the predictor uses both the original input and retrieved texts to output predictions. On reaction condition recommendation and one-step retrosynthesis, TextReact outperforms state-of-the-art deep-learning chemistry models by 58.4% and 13.6–15.7%, respectively.
pdfarXivGitHub
, , , , ,
MolScribe: Robust Molecular Structure Recognition with Image-to-Graph Generation
In Journal of Chemical Information and Modeling, 2023.
MolScribe translates molecular diagrams in image format to a structured graph format. It explicitly predicts atoms and bonds along with their positions, and flexibly incorporates symbolic chemistry constraints to recognize chirality and expand abbreviated structures. MolScribe achieves 76–93% accuracy on public benchmarks.
pdfarXivGitHub
, , , ,
LEMMA: Bootstrapping High-Level Mathematical Reasoning with Learned Symbolic Abstractions
In MATH-AI Workshop at NeurIPS, 2022.
Learning Mathematical Abstractions (LEMMA) learns a hierarchy of abstractions to enhance RL for mathematical reasoning. It augments Expert Iteration with an abstraction step, where solutions found so far are rewritten in terms of new higher-level actions, which then become available to solve new problems. LEMMA increases the performance and generalization of an RL agent on equation solving and fraction simplification.
pdfarXivGitHub