Min Woo Sun

Incoming Research Scientist @GoogleDeepMind | PhD @Stanford

[Min Woo Sun @minwsun]

Grateful to everyone who made this work possible. It was a privilege to contribute alongside such an amazing group of collaborators.

@SRSchmidgall , @_shaw_marth, Lin Yang, @valentinlievin, Jingyun Yang , Yuchen Zhuang, @timstro, Alex Bijamov, @minwsun, @apalepu13, Justin Chen, David Steiner, @JShreibati, @ckbjimmy, Yilin Zhao, Xingjian Hu, Nicholas Zahn, Sadhya Garg, Julia Kirby, Yuxiang Gan, Jiaoli Li, @divy93t, @AziziShekoofeh, David Racz, @Mysiak, @vivnat, @ChenglinWu3, @tdanino, @KeranRong, @harryhaozhewang, @benoitschilling, Yong Cheng, @quocleix, and @taotu831

[Min Woo Sun @minwsun]

1/ Excited to share our new work: “Accelerating Scientific Research with Gemini in the Real-World.”

We extend the AI Co-Scientist into an execution-grounded research system spanning ideation → experimentation → paper writing, and validate it across materials science, biology, and computer science. 🧵

[Min Woo Sun @minwsun]

8/ What I find most exciting is the direction this points toward: AI systems that don’t just reason about science, but can iteratively interact with experiments, code, measurements, and human experts.

[Min Woo Sun @minwsun]

7/ We evaluated these mechanisms in a double-blind study with 30 domain experts and 450 independent reviews.

The reliability modules reduced severe result hallucinations and plagiarism relative to the ablated baseline, while the safety system refused 98.7% of hazardous prompts.

[Min Woo Sun @minwsun]

6/ An important part of autonomous science is not just capability, but scientific reliability.

We therefore added mechanisms that explicitly penalize hallucination/plagiarism and verify manuscript claims against raw experimental execution logs before finalization.

[Min Woo Sun @minwsun]

5/ In computer science, we pushed autonomy further.
Given only a research directive, Co-Scientist autonomously designed and iterated an inference-time scaling architecture, Agent_H, which outperformed six frontier models on length-adjusted HealthBench Hard & Professional.

[Min Woo Sun @minwsun]

4/ In biology, Co-Scientist built a vision-based system to predict emergent E. coli swarming morphologies at unseen inducer concentrations from sparse experimental images.

The predictions showed quantitative concordance with subsequent wet-lab measurements across most evaluated morphology metrics.

[Min Woo Sun @minwsun]

3/ In materials science, Co-Scientist helped design CVD synthesis protocols tailored to a custom lab setup.
This included a safer precursor route for a 2D layered material with characteristics consistent with Ti₃C₂Tₓ MXene, as well as single-attempt growth of monolayer MoS₂, MoSe₂, and WS₂.

[Min Woo Sun @minwsun]

2/ A central question for AI-for-science is:
Can an AI system move beyond hypothesis generation to interface with real-world experiments and empirical feedback?
Here, we explore that spectrum, from human-executed wet-lab experiments to fully autonomous computational research.

[Min Woo Sun @minwsun]

Incredible work by @valentinlievin, @SRSchmidgall, @timstro, Alex Bijamov and the @GoogleDeepMind @GoogleResearch team! So proud to be part of this project before officially starting my RS role in a few weeks!

[Valentin Liévin @valentinlievin]

Medical training begins in textbooks, but clinical competence is forged through practice. We introduce ResidencyRL: a multi-turn online RL method to train health agents in simulation, and demonstrate significant gains on top of frontier systems. https://t.co/CGw4EO7qh7

[James Burgess @jmhb0]

Introducing MicroVQA: A Multimodal Reasoning Benchmark for Microscopy-Based Scientific Research #CVPR2025

✅ 1k multimodal reasoning VQAs testing MLLMs for science
🧑‍🔬 Biology researchers manually created the questions
🤖 RefineBot: a method for fixing QA language shortcuts
🧵

[James Burgess @jmhb0]

🚨Large video-language models LLaVA-Video can do single-video tasks.

But can they compare videos?

Imagine you’re learning a sports skill like kicking: can an AI tell how your kick differs from an expert video?

🚀 Introducing "Video Action Differencing" (VidDiff), ICLR 2025
🧵

[Alejandro Lozano @Ale9806_]

Biomedical datasets are often confined to specific domains, missing valuable insights from adjacent fields. To bridge this gap, we present BIOMEDICA: an open-source framework to extract and serialize PMC-OA.

📄Paper: https://t.co/zkb2m5yeal
🌐Website: https://t.co/atIKEfOpkv

[Thomas Wolf @Thom_Wolf]

Among the most impressive aspect of the Llama 3.1 release is the accompanying research paper! Close to 100 pages of deep knowledge-sharing on LLMs like we havn't seen very often recently

What a treat!

It covers everything, pretrainining data, filtering, annealing, synthetic data, scaling laws, infrastructures, parallelism, training recipees, post-training adaptation, tool-use, benchmarking, inference strategies, quantization, vision, speech, videos...

Mind-blown! Maybe the single paper you can read today to join the field of LLM from zero right to the frontier

Read it here and feel the open-science https://t.co/nANpZtiP0s

[Freddy A Boulton @freddy_alfonso_]

The @Gradio ecosystem can bridge the gap between AI and life sciences 🧬

Very nice custom component @simonduerr

cc @katieelink @minwsun

[rob tibshirani @robtibshirani]

Do Public Health Efforts Matter? Explaining Cross-Country Heterogeneity in Excess Death During the COVID-19 Pandemic https://t.co/5DsIm1H67D