Zhen Wang

Zhen Wang 王震

Postdoctoral Researcher

Hi! I'm Zhen, currently a Gordon and Betty Moore Foundation Postdoctoral Fellow hosted at UC San Diego. My postdoc mentors are Profs. Eric Xing (CMU & MBZUAI) and Zhiting Hu (HDSI & CSE, UCSD).

I am interested in building the foundations of general-purpose knowledge and reasoning systems that learn to acquire knowledge and reason to discover at scientist-level efficiency and extrapolation, with reliability, and that collaborate with humans across interdisciplinary and scientific domains.

To this end, my research seeks to answer fundamental questions: How do models represent knowledge for discovery? How do they reason and plan to extrapolate? How do they reliably collaborate and co-discover with humans? My philosophy centers on bringing new perspectives to shift existing paradigms or unifying seemingly different ones—exemplified by LLM Reasoners (reasoning as planning), PromptAgent (prompting as search), ToolkenGPT (tools as tokens), and the LAW 2025 workshop at NeurIPS (bridging language, agent, and world models).

I earned my PhD from the Computer Science department at The Ohio State University, advised by Prof. Huan Sun, where I developed foundational frameworks for knowledge-centric AI systems. My work has been supported by and recognized with the Gordon and Betty Moore Foundation Fellowship, OpenAI Research Grant, Rising Star in Data Science (UChicago), Best Paper Award at SoCal NLP, Amazon Alexa Prize, and more. I collaborate with MIT-IBM Lab, Microsoft Research, NEC Labs, and academic institutions including CMU, MBZUAI, and multiple UC campuses and national labs through the UC-LEAP project.

Research Overview View Details →

Research Overview: Scientist AI that Reason to Discover - Accelerating the Lifecycle of Scientific Discovery

My research agenda, Scientist AI, sits at the intersection of advanced AI reasoning and accelerated discovery across science and society. This space extends traditional data-driven discovery, calling for human-centered AI systems that can acquire structured knowledge from data, actively explore hypotheses, simulate possibilities, and generate new, verifiable knowledge across disciplines.

Selected Honors

News

Jan 2026
We released FIRE-Bench, benchmarking full-cycle scientific discovery automation with verfiable evaluation.
Jan 2026
Nabla-Reasoner was accepted to ICLR 2026: latent test-time reasoning with first-order optimization in logit space.
Dec 2025
Honored to organize LAW 2025 workshop (Bridging Language, Agent, and World Models) at NeurIPS 2025. Thanks to our amazing co-organizers, Ziqiao, Jessy, Melanie, Jianwen, Kelsey, Alane, Jacob, Tianmin, and Zhiting; and speakers, Sherry, Chelsea, Francois, Eric, Danijar, Philop, Stephen, Ying Nian, and Keyon. Check the recording of their talks here (with a NeurIPS registration); public videos will be released 30 days after the workshop.
Dec 2025
Presenting two papers at NeurIPS 2025: scPilot (first omics-native reasoning agents) and DeepPersona (generative engine for real-human personas).
Nov 2025
Invited talk at UCSD CSE AI Seminar: "Building AI That Discovers."
Oct 2025
Received the Gordon and Betty Moore Foundation Fellowship for the Natural Sciences to support research on AI-driven scientific reasoning and discovery.
Sep 2025
Part of newly accepted Nature paper on multi-omics discovery; contributed TaijiChat, a Paper Copilot.
Sep 2025
scPilot accepted to NeurIPS 2025: omics-native reasoning for grounding LLMs in raw omics data.
Sep 2025
Released preprint: MutationProjector: cancer genomics foundation model for tumor mutation profiles.
Aug 2025
Attended UC-LEAP Kick-off at UCSB: exploring AI agents for Materials Science.
Jul 2025
Hosting LAW 2025 workshop at NeurIPS 2025 in San Diego this December!
Jun 2025
Attended DL4SCI 2025 at Berkeley: diving into AI for automated scientific discovery.
Jan 2025
Self-MoE accepted to ICLR 2025: transforming monolithic LLMs into modular self-specialized experts.
Oct 2024
Released Decentralized Arena: democratic LLM benchmarking where models judge each other. HF Leaderboard
Oct 2024
Released TxT360: first dataset to globally deduplicate 99 CommonCrawl snapshots. Dataset
Sep 2024
DRPO accepted to EMNLP 2024: first tuning-free method for self-aligning LLMs with human preferences.
Jul 2024
Selected for 2024 AI+Science Summer School at University of Chicago.
Jul 2024
LLM Reasoners accepted to COLM 2024. Check the 1k+ star GitHub package!
Feb 2024
Received OpenAI research grant to support agentic systems research.
Jan 2024
PromptAgent accepted to ICLR 2024: first principled framework for API-based prompt optimization.
Nov 2023
Received Top Reviewer Award at NeurIPS 2023.
Nov 2023
Best Paper Award at SoCal NLP 2023.
Oct 2023
RAP accepted to EMNLP 2023 Main: augmenting LLM reasoning with world models and planning.
Sep 2023
ToolkenGPT accepted as oral presentation at NeurIPS 2023: augmenting LLMs with efficient tool learning.
Feb 2023
Started postdoc at UCSD, collaborating with teams from CMU and MBZUAI on large language models research.
Jan 2023
Multitask Prompt Tuning accepted to ICLR 2023.
Nov 2022
Passed PhD defense: "Toward Knowledge-centric Natural Language Processing: Acquisition, Representation, Transfer, and Reasoning."
Jun 2022
TacoBot earned 3rd place in inaugural Alexa Prize TaskBot Challenge.
May 2022
Joined MIT-IBM Watson AI Lab as research intern on efficient LLM adaptation.
Mar 2022
Received 2022 Graduate Research Award from the CSE department.
Dec 2020
Selected as Rising Star in Data Science at the University of Chicago CDAC.