James Zou

James Zou leads the Stanford AI for Science Lab and is an associate professor of biomedical data science, with courtesy appointments in computer science and electrical engineering. His research connects statistically rigorous machine learning with applications in science and human health. Across data valuation, medical-model evaluation, and scientific agents, he examines how AI systems can learn from feedback that tests whether their results are useful.

From applied mathematics to scientific AI

Zou studied mathematics and physics at Duke, pursued advanced mathematics at Cambridge, and completed his doctorate at Harvard. After a Simons fellowship at UC Berkeley and work at Microsoft Research, he joined Stanford in 2016. His research there spans machine-learning theory, genomics, and medicine. He received Chan–Zuckerberg Biohub Investigator awards in 2017 and 2023.

With Amirata Ghorbani, Zou introduced Data Shapley in 2019 to estimate how individual training examples contribute to a model’s predictive performance. The framework makes data quality a question about a particular learning task: which examples help the predictor, which may be corrupted, and what additional data would improve it? This gives researchers a way to assess a dataset beyond its size and make more informed decisions about cleaning and acquisition.

His medical collaborations show why those distinctions matter. With Roxana Daneshjou and colleagues, Zou co-authored the Diverse Dermatology Images study, using clinically curated, biopsy-confirmed skin lesions spanning different skin tones. Their 2022 study found weaknesses in leading dermatology models, especially on darker skin and uncommon diseases. Fine-tuning with the diverse images reduced performance disparities. The result connected model reliability to the patients and conditions represented in training and evaluation data.

Zou also co-authored the original 2019 Gradio paper. Its browser interfaces let people test machine-learning models without setting up a specialist computing environment. For interdisciplinary work, that makes expert feedback easier to obtain: a physician or other domain specialist can supply realistic inputs, inspect predictions, and identify problems directly.

Feedback that changes the system

As language models became capable of using tools and carrying out extended analyses, Zou’s collaborative research expanded to systems that revise their work, coordinate expertise, and test proposed solutions. TextGrad, published in Nature in 2025, propagates written feedback through systems containing language models and other tools. A critique of a final result can guide revisions to the earlier components that produced it. Its applications include prompts, code, molecules, and radiotherapy treatment plans, extending optimization to objects scientists and engineers can inspect and change. The open-source framework makes this feedback process available for experimentation.

Virtual Lab applies agent collaboration to interdisciplinary research. Zou and John Pak co-supervised the work; Kyle Swanson built the framework and computational design pipeline, and experimental collaborators tested the resulting nanobodies. A principal-investigator agent coordinates scientist agents through research meetings, while a human supplies high-level feedback. For SARS-CoV-2 nanobody design, the agents assembled a pipeline combining ESM, AlphaFold-Multimer, and Rosetta. The team designed 92 candidates, and laboratory experiments identified functional binders, including two with improved binding to newer variants. Those experiments supplied a physical check on the computational proposals; the candidates remained subjects for further investigation.

Environments for collective intelligence

Zou’s scientific-agent work with Together AI, in collaboration with Stanford, develops an argument for environments for collective intelligence. His concern is that prescribed workflows can constrain increasingly capable agents. An environment instead supplies problems, resources, incentives, guardrails, and feedback, leaving agents more freedom to choose how to work. EinsteinArena and DSGym develop this idea through different kinds of executable evaluation.

• EinsteinArena makes intermediate progress reusable. With Federico Bianchi, Yongchan Kwon, and Aneesh Pappu, Zou developed EinsteinArena, where agents choose scientific problems, submit solutions to verifiers, inspect other agents’ work, and exchange ideas in discussion forums. The problems are selected for scientific interest and for having well-defined checks that can score submissions. Public solutions and continuously updated leaderboards create opportunities for both competition and collaboration: agents can pursue better scores while reusing and refining one another’s partial progress.

One result improved the known lower bound for the 11-dimensional kissing-number problem from 593 to 604. The problem asks how many equal spheres can touch a central equal sphere without overlapping one another. The new construction establishes that at least 604 can do so in 11 dimensions; it does not establish the maximum. Successive submissions and forum discussions allowed agents to share attempted approaches and improve one another’s solutions. The example gives Zou’s argument for collaboration a concrete mechanism: useful intermediate work remains available for other agents to build on.

• Kernel optimization gives collaboration a measurable engineering target. The same environment supports GPU-kernel optimization by changing what the evaluator checks. Submitted kernels are compiled, tested for correctness, and benchmarked for runtime, with results returned through the leaderboard. Agents with different priorities—profiling, memory consumption, and numerical precision—collaborate and compete to improve kernels. Zou reported speedups exceeding twofold for some tested kernel configurations and said improved kernels were already used in production at Together AI. These results concern particular kernels and configurations, rather than a general doubling of system performance.

• DSGym requires agents to work with the data. DSGym addresses an evaluation failure: data-science tasks that agents can answer without examining the underlying datasets. Zou reported that, on the existing benchmarks his team examined, roughly 20–50% of tasks could sometimes be solved through reasoning or other shortcuts without using the data. Such scores can overstate an agent’s ability to perform an analysis.

Zou and collaborators built DSGym around executable tasks, curated data, and checks for shortcuts and task quality. Scientific-analysis tasks draw on recently published papers and receive review from human scientists and experts. Predictive-modeling tasks draw on completed Kaggle competitions that still accept submissions for leaderboard evaluation. The competitions have ended, but their continuing submission infrastructure lets researchers evaluate new predictions. A unified execution interface lets agents work with the data and run analyses in Docker containers, including parallel experiments.

• Execution-verified trajectories turn tested work into training data. DSGym also produces execution-verified trajectories: records of agent work checked by actually running the generated code. Successful analyses can then become training examples for smaller open-source models. Evaluation and training share the same practical foundation—an agent’s proposed analysis must survive execution against the data. This extends Zou’s earlier work on data quality and model evaluation into scientific systems that can use tested results to improve their next attempt.