Yufan Zhuang

Hey, I’m Yufan Zhuang 庄宇凡

I am a Research Scientist at Google Cloud AI Research working on making language models reason better and act more agentically. I received my PhD from UC San Diego, advised by Jingbo Shang.

📢 I’m looking for motivated interns! If you’re interested in working on LLM reasoning and agents, feel free to email me.

🔍 LLM Reasoning and Agentic Learning

RRSI (arXiv’26) - Regularized Recursive Self-Improvement of Agent Harnesses
Regularizing how agent harnesses evolve to improve generalization beyond the tasks used during self-improvement, while reducing token costs.

Register Tokens (arXiv’26) - Bounded-State Reasoning in Diffusion Language Models
Carrying reasoning progress across generation chunks in a fixed-size set of continuous register states, improving math and code reasoning without retaining all earlier generated text.

EnvHarness (arXiv’26) - Awakening Static Worlds for Agent Learning
Adapting existing environments to agents’ weaknesses through programmable components that preserve the original verifier, supporting targeted agent learning and reinforcement learning.

Test-time Recursive Thinking (arXiv’26) - Self-Improvement without External Feedback Test-time Recursive Thinking improves model reasoning without external feedback. +10.4-14.8 pp on the most difficult problems of LiveCodeBench (accuracy), with pure test-time scaling.

Knowledge Flow (Blog) - Scaling Reasoning beyond Context Limit
Knowledge Flow scales reasoning beyond context limits, enabling open-source models to reach 100% on AIME-25 without training or tools or external feedbacks.

Mixture of Inputs (NeurIPS’25) - Beyond Discrete Sampling
Exploring continuous mixture approaches that fundamentally improve how language models reason through complex problems by operating in continuous vector space rather than discrete tokens.

Vector-ICL (ICLR’25) - Cross-Modal In-Context Learning
Enabling models to learn from any data modality in-context through continuous vector representations, breaking down traditional barriers between text, images, and other data types.

AgenticLU (ACL’25) - Agentic Long Context Learning
Improving long-context capabilities through intelligent agentic approaches that can dynamically reason through complex, extended content with strategic planning.

MetaTree (TMLR’24) - Tabular Data Processing with Transformers
Training transformers via meta-learning to directly produce strong decision trees, bridging classical ML algorithms with modern neural approaches for superior generalization.

Deep Contamination (EMNLP’24) - Cross-lingual Data Contamination Study
Uncovering critical issues in how we evaluate contemporary LLMs, particularly around cross-lingual contamination that affects model reliability.

🔬 AI for Software Engineering & Sociology

Signal-Aware Code Understanding (ACM TOSEM’23) - Trustworthy AI for Source Code
Developing signal-aware AI models for software vulnerability detection that focus on learning task-relevant source code features rather than spurious correlations, improving model robustness and reliability.

Prediction-Preserving Input Minimization (ESEC/FSE’21) - Probing AI Model Understanding
Introducing P2IM approach to systematically evaluate AI models’ signal awareness by reducing source code to minimal snippets needed to maintain predictions, revealing models’ reliance on incorrect signals.

Computational Social Science (Annals AAG’22) - Machine Learning for Sociological Analysis
Applying machine learning approaches to decipher heterogeneous representations and images of Chinese communities in North America, bridging computational methods with sociological research.

Prior to my PhD study, I worked at IBM T. J. Watson Research Center as a research engineer helping to enhance software engineering with the power of AI and vice versa. I received my MS in Data Science from Columbia, my BSc in Applied Math Minor in CS (with First Class Honors) from Hong Kong Polytechnic University.