Xinran Zhao

PhD Candidate · Carnegie Mellon University

Xinran Zhao

I am a fourth-year PhD candidate at CMU LTI, advised by Sherry Wu. My PhD is supported by the Amazon AI PhD Fellowship.

At Google DeepMind (2024, 2026), I worked as a student researcher on LLM planning, long-horizon code generation, and continual learning with Azade Nova and Hanie Sedghi. At Ai2 (2025), I worked on deep research agents with Varsha, Aakanksha, Joseph, Jay, and Jena.

I received my MSCS from Stanford, where I worked with Christopher Manning, and my bachelor's degree from HKUST. At HKUST and during my exchange at Cornell, I was fortunate to work with Yangqiu Song, Dit-Yan Yeung, and Claire Cardie.

I always feel lucky to meet great mentors along my journey, including but not limited to, Shikhar Murty, Hongming Zhang, and Esin Durmus.

I am on the 2026–27 industry job market, with a flexible graduation date!

I build highly autonomous language model agents that seek information, learn from experience, plan, and collaborate with people. Specifically, I work on:

Building agents that seek information and deliver it proactively.

  • Knowing When to Stop: when a coding agent is blocked, proactively telling the user cuts wasted effort and unsafe actions.
  • Intent-aware Deep Research (ICLR'26): making the intents behind what to write and cite explicit improves both the performance and readability of deep research reports.
  • MoR (EMNLP'25): a zero-shot mixture of sparse, dense, and human retrievers, weighted per query.
  • Revela (ICLR'26 Oral): trains dense retrievers through language modeling, without annotated query–document pairs.

Building agents that plan, code, and forecast their own outcomes.

  • Action Sequence Similarity (ICLR'25): choosing in-context exemplars by plan similarity, not problem similarity, improves LLM planning.
  • 3DCodeEdit: a benchmark of instruction-guided 3D program editing, where models succeed on only 37% of tasks on average.
  • Tea Leaves: can LLMs read the start of a training run and predict how it ends?
  • ReSHAPE (BiAlign@NeurIPS'26 Oral): agents recover plans users would accept by showing concrete options instead of asking about preferences.

Building agents that debate, persuade, and stay honest about what they know.

  • TreeDebater (ICLR'26): tree-structured planning that rehearses attacks and tracks the debate flow, making LLM debaters more persuasive to human judges.
  • Real-Time Debate: debating in live speech, where the agent plans while it listens and speaks.
  • FaR (ACL'24): eliciting known facts and then reflecting on them improves confidence calibration.

Selected Honors and Awards

Academic Events

  • Winter, 2027 General chair, FrontierIR, the Second Workshop on New Frontiers in Information Retrieval at AAAI 2027.
  • Fall, 2026 Invited panelist, Global Embodied Ecosystem Night at IROS 2026.
  • Summer, 2026 Invited talk at Google DeepMind: Continual Learning for Coding Agents.
  • Summer, 2025 Invited talk at UW H2 Group: Improving Large Language Model Planning with Action Sequence Similarity.
  • Summer, 2024 Invited talk at Samaya AI: Towards Building Retrieval Systems in Complex Realistic Scenarios.
  • Summer, 2022 Invited talk at the ICML Commonsense Tutorial: On Task Difficulty of Few-shot Learning.
  • Spring, 2022 Invited talk at NUS, Department of Marketing: NLP as a Tool for Scientific Discovery.

Service

Miscellaneous

Besides research, I also enjoy: