Shibo Hao

I am a Ph.D. candidate at UC San Diego, advised by Zhiting Hu. My research is funded by the Bloomberg Fellowship. I was a research scientist intern at Meta FAIR lab, mentored by Yuandong Tian and Jason Weston. I received my B.S. in Computer Science from Peking University. I'm currently looking for a full-time position in industry.

My research focuses on pushing the boundaries of machine reasoning in large language models. My work includes developing latent-space reasoning methods (Coconut, Coconut-Theory, Coconut-Dynamics), training LLMs to reason via reinforcement learning (Guru, OREO), building system-2 inference-time reasoning frameworks (RAP, LLM Reasoners), and finding better ways for LLM agents to interact with the world (ToolkenGPT, CocoaBench).

Bold indicates first or co-first authorship.

News

Dec 2025 Released CocoaBench, a new benchmark for general digital agents.
Sep 2025 Guru and Reasoning-by-Superposition both accepted to NeurIPS 2025.
Apr 2025 Coconut featured in Quanta Magazine.
Dec 2024 Honored to receive the Bloomberg Data Science Ph.D. Fellowship for 2024–2025.
Dec 2024 Introducing OREO (Offline REasoning Optimization) for multi-step LLM reasoning.
Jul 2024 LLM Reasoners accepted to the first Conference on Language Modeling (COLM 2024).
May 2024 Check out Pandora, our new work towards a general world model.
Nov 2023 ToolkenGPT accepted to NeurIPS 2023 as an oral presentation and received the Best Paper Award at SoCalNLP 2023.
Oct 2023 RAP featured in the State of AI Report 2023.