Research

My long-term goal is to build AI systems that remain faithful and useful under uncertainty: systems that can detect unsupported claims, localize what should be changed, allocate computation to uncertain parts of a reasoning trace, and act safely in long-horizon environments. I approach this through a bridge between diffusion-based generation/editing and agent evaluation/training.

In diffusion language modeling, I study how non-autoregressive models can use their ability to infill and revise arbitrary spans for faithful summarization, reasoning templates, and efficient inference. In agents, I study how realistic environments, verifiers, and harnesses can make training and evaluation match the way agents are actually deployed.

News

• Jun 2026: Detect, Remask, Repair is on arXiv and under review at EMNLP 2026.
• Jun 2026: Continuing at Columbia CS as a Research Staff Associate in Prof. McKeown's group on faithful summarization.
• 2026: OSWorld 2.0, a benchmark for long-horizon real-world computer-use agents, is under review at NeurIPS 2026.
• 2025: No Compute Left Behind studies reasoning, sampling, and inference-time compute allocation for masked diffusion language models.
• 2024: Our question rewriting work was accepted to EMNLP 2024.

Selected Papers & Projects

	Detect, Remask, Repair: Diffusion Editing for Faithful Summarization of Evolving Contexts Hao Zou, Zachary Horvitz, Chandhru Karthick, Zhou Yu, Kathleen McKeown Under review, EMNLP 2026 arXiv / bibtex A diffusion-based framework for localized faithfulness repair: detect unsupported spans in an existing summary, remask them, and repair only the unreliable content instead of fully regenerating the summary.
	OSWorld 2.0: Benchmarking Computer Use Agents on Long-Horizon Real-World Tasks Large collaborative project Under review, NeurIPS 2026 project page / arXiv coming soon A benchmark of 102 realistic computer-use tasks and 30 self-hosted mock websites for evaluating long-horizon GUI agents, including dynamic workflows, tutorial following, simulated-user interaction, multimodal editing, and safety-oriented checks.
	Scalable Environments and Training Infrastructure for Harness-Based Agents Ongoing collaboration with Xiao Yu In preparation, 2026 Building agent-curated task environments and training infrastructure that connects real inference harnesses to SFT/RL pipelines, reducing the mismatch between how agents are trained and how they are deployed.
	No Compute Left Behind: Rethinking Reasoning and Sampling with Masked Diffusion Models Zachary Horvitz, Raghav Singhal, Hao Zou, Carles Domingo-Enrich, Zhou Yu, Rajesh Ranganath, Kathleen McKeown Under review, COLM 2026 arXiv / bibtex Studies how masked diffusion language models use inference-time compute, including reasoning-as-infilling, answer-conditioned posterior sampling, uncertainty-aware early exit, and reasoning validation.
	You Make me Feel like a Natural Question: Training QA Systems on Transformed Trivia Questions Tasnim Kabir, Yoo Yeon Sung, Saptarashmi Bandyopadhyay, Hao Zou, Abhranil Chandra, Jordan Boyd-Graber EMNLP 2024 paper / bibtex Transforms trivia-style questions into more natural information-seeking questions to improve retrieval alignment and cross-domain question answering.
	A Survey of Diffusion Models in Natural Language Processing Hao Zou, Zae Myung Kim, Dongyeop Kang arXiv, 2023 arXiv / bibtex Surveys diffusion formulations for NLP, including non-autoregressive generation, text editing, token-level control, robustness, and scaling directions for diffusion language models.
	Debiasing Language Models for In-Context Learning Using a Causal Inference-Inspired Method Hao Zou, Karin de Langis, Dongyeop Kang, Yohan Jo Manuscript, 2023 bibtex Uses intervention-inspired estimates of the causal effect of input text to reduce spurious prompt-label bias in in-context learning.

Experience