Samuel Schmidgall

Hello there. My name is Samuel Schmidgall. I am a research scientist at Google Deepmind working on biomedical research as well as post-training for Gemini. I will recieve my PhD from Johns Hopkins University in Spring 2026. I was advised by Rama Chellappa and worked closely with Axel Krieger and Michael Moor. I also received support from the NSF Graduate Research fellowship (NSF GRFP).

Email  /  LinkedIn  /  Scholar  /  Twitter | X  /  Github

Research

Co-Scientist is a multi-agent system designed to accelerate end-to-end scientific research across hypothesis generation, experimentation, and manuscript generation, demonstrated on real-world problems in materials science, biology, and computer science, including designing precursor routes for MXenes, predicting bacterial swarming behavior, and discovering inference-time scaling architectures. A double-blind study of end-to-end generated papers with 30 domain experts across 450 reviews evaluates reliability and safety.

The first demonstration of expert-level AI in real-time clinical video consultations, using AMIE (Video), a Gemini-based multi-agent system integrating low-latency dialogue, clinical reasoning, and real-time audio-visual perception. In a randomized OSCE study with 30 primary care physicians, 15 patient actors, and 100 clinical scenarios, clinical evaluators rated AMIE (Video) on par with or better than physicians in history-taking, diagnosis, management, and physical observation.

ResidencyRL trains clinical AI agents through simulated patient dialogues of up to 60 turns with multiple tool interactions. The resulting agent improves diagnostic accuracy by 7.0% under adversarial conditions (88.0% vs. 81.0%), reduces missed red-flag rates by 31%, and is preferred by clinicians in 87.6% of evaluations, with skills that transfer across multiple clinical benchmarks.

MatrAIx is a population-scale simulated-user evaluation infrastructure built on Persona 8B, 8.3 billion persona records spanning 1,290 dimensions, with a quality-filtered coreset of roughly 1 million personas. Its Playground offers four evaluation environments (Survey, AI Chatbot, Web, App) and was tested across more than 1,000 application tasks spanning 25+ domains.

Gemma 4 is a new generation of open-weight, natively multimodal language models ranging from 2.3B to 31B parameters, featuring both dense and Mixture-of-Experts architectures, improved vision and audio encoders, a unified encoder-free architecture for the 12B model, and a thinking mode for generating reasoning traces prior to responding.

Engineered E. coli strains generate centimeter-scale swarming patterns that spatially record chemical and optical environmental inputs in analog or binary-like fashion. Scalable computational methods, including feature extraction, regression, and deep-learning models, decode the resulting bacterial patterns across space and time.

MedGemma 1.5 4B expands the MedGemma collection to high-dimensional medical imaging (CT/MRI volumes and histopathology whole slide images), anatomical localization via bounding boxes, multi-timepoint chest X-ray analysis, and improved medical document understanding, with substantial performance gains over its predecessor while remaining an open foundation for medical AI development.

CoDaS is an AI co-data-scientist that combines multi-agent hypothesis generation with statistical analysis to turn continuous wearable sensor signals into clinically reviewable biomarker hypotheses across nearly 9,300 participant observations, identifying circadian-instability patterns linked to depression and a cardiovascular-fitness metric associated with insulin resistance.

SymptomAI is a conversational AI agent for everyday symptom assessment, deployed through the Fitbit app in a randomized study with 13,917 participants across five AI agents. Against clinician-provided diagnoses, SymptomAI differentials were significantly more accurate (OR = 2.56) than those from independent clinicians, and dedicated symptom interviews substantially outperformed user-guided conversations with consumer LLM defaults.

Patients With Personality (PWP) is a framework for simulating realistic and diverse virtual patients by grounding each patient in the six-dimensional HEXACO personality space, enabling fine-grained control over conversational style, cooperativeness, and information disclosure. In clinician evaluations, PWP patients achieve near-parity with human actors and substantially outperform prior simulators, which are frequently flagged as unrealistically informative.

TeamBench evaluates how agent teams coordinate under system-enforced role separation, splitting specification access, workspace modification, and answer verification across Planner, Executor, and Verifier roles over 851 task templates. Prompt-based and sandbox-enforced teams achieve comparable success rates, but prompt-only configurations produce 3.6 times more role violations, and team structures help only when individual agents perform poorly.

Proof of Time (PoT) is a semi-verifiable benchmarking framework that links scientific idea judgments to downstream signals that only become observable later, such as citations and shifts in researchers' agendas. Evidence snapshots are frozen before a cutoff date and models must predict outcomes occurring afterward, evaluated across more than 30,000 instances spanning four domains.

A controlled experiment holding task prompts, tools, and compute budgets constant while varying only coordination structure and model capability, across 260 configurations spanning six benchmarks, five architectures, and three LLM families. Single-agent baseline performance is the most robust predictor of whether coordination helps, and an empirical capability-saturation threshold beyond which additional agents are unlikely to improve performance correctly predicts the effect of multi-agent coordination in 94% of validation configurations on SWE-bench Verified and Terminal-Bench.

We introduce the first comprehensive catalog of validation pitfalls in AI-based surgical video analysis that was derived from a multi-stage Delphi process with 91 international experts. The collected pitfalls span three categories: (1) data (e.g., incomplete annotation, spurious correlations), (2) metric selection and configuration (e.g., neglect of temporal stability, mismatch with clinical needs), and (3) aggregation and reporting (e.g., clinically uninformative aggregation, failure to account for frame dependencies in hierarchical data structures). .

MedGemma is a collection of medical vision-language foundation models based on Gemma 3 4B and 27B. MedGemma demonstrates advanced medical understanding and reasoning on images and text, significantly exceeding the performance of similar-sized generative models and approaching the performance of task-specific models, while maintaining the general capabilities of the Gemma 3 base models.

SRT-H utilizes a high-level policy for task planning and a low-level policy for generating low-level trajectories. The high-level planner plans in language space, generating task or corrective instructions to guide the robot through the long-horizon steps and correct for the low-level policy's errors. We validate our framework through ex vivo experiments on cholecystectomy, a commonly-practiced minimally invasive procedure, and conduct ablation studies to evaluate key components of the system. Our method achieves a 100% success rate across n=8 different ex-vivo gallbladders, operating fully autonomously without human intervention.

State-of-the-art surgery is performed robotically under direct surgeon control. However, surgical outcome is limited by the availability, skill, and day-to-day performance of the operating surgeon. What will it take to improve surgical outcomes independent of human limitations? In this review, we explore the technological evolution of robotic surgery and current trends in robotics and AI which could lead to a future generation of autonomous surgical robots that will outperform today’s tele-operated robots.

A feature article on next-generation autonomous surgical systems that can suture soft tissue with minimal human input, and what it will take to bring surgical autonomy from the lab into the operating room.

Here we introduce TxGemma, a suite of efficient, generalist large language models (LLMs) capable of therapeutic property prediction as well as interactive reasoning and explainability. Unlike task-specific models, TxGemma synthesizes information from diverse sources, enabling broad application across the therapeutic development pipeline. The suite includes 2B, 9B, and 27B parameter models, fine-tuned from Gemma-2 on a comprehensive dataset of small molecules, proteins, nucleic acids, diseases, and cell lines.

Here, we introduce AgentRxiv, a centralized preprint server designed specifically for autonomous research agents to overcome the limitations of isolated research outputs by enabling collaborative, cumulative knowledge sharing. We find that agents sharing research have substantial performance improvements when making discoveries.

Here, we introduce Agent Laboratory, an open-source large language model (LLM) agent framework for accelerating the individual’s ability to perform research. Agent Laboratory takes a human-produced research idea as input and outputs a code repository and a research report. This is accomplished through specialized agents driven by LLMs that collaborate to perform research based on the human’s preference for the agent’s involvement.

MedBrowseComp is the first benchmark that systematically tests an agent's ability to reliably retrieve and synthesize multi-hop medical facts from live knowledge bases. MedBrowseComp contains more than 1,000 human-curated questions that mirror clinical scenarios where practitioners must reconcile fragmented or conflicting information to reach an up-to-date conclusion.

Surgical Gaussian Surfels (SGS) transforms anisotropic point primitives into surface-aligned elliptical splats by constraining the scale component of the Gaussian covariance matrix along the view-aligned axis.

Neurobench is a collaborative effort of nearly 100 co-authors across over 50 institutions in industry and academia, aiming to provide a representative structure for standardizing the evaluation of neuromorphic approaches.

Here, we introduce the Surgical Robot Transformer (SRT). We explore whether surgical manipulation tasks can be learned on the da Vinci robot via imitation learning. We demonstrate our findings through successful execution of three surgical tasks, including tissue manipulation, needle handling, and knot-tying.

To improve on existing noncardiac surgery risk scores, this work proposes a novel approach which leverages features of the preoperative 12-lead electrocardiogram to predict major adverse postoperative outcomes.

This perspective aims to provide a path toward increasing robot autonomy in robot-assisted surgery through the development of a multi-modal, multi-task, vision-language-action model for surgical robots.

This paper introduces SurGen, a text-guided diffusion model tailored for surgical video synthesis, producing the highest resolution and longest duration videos among existing surgical video generation models.

AgentClinic turns static medical QA problems into agents in a clinical environment in order to present a more clinically relevant challenge for multimodal language models. Here, we introduce AgentClinic, a multimodal agent benchmark for evaluating LLMs in simulated clinical environments that include patient interactions, multimodal data collection under incomplete information, and the usage of various tools, resulting in an in-depth evaluation across nine medical specialties and seven languages.

This paper introduces GP-VLS, a general-purpose vision language model for surgery that integrates medical and surgical knowledge with visual scene understanding. For comprehensively evaluating general-purpose surgical models, we propose SurgiQual, which evaluates across medical and surgical knowledge benchmarks as well as surgical vision-language questions. To train GP-VLS, we develop six new datasets spanning medical knowledge, surgical textbooks, and vision-language pairs for tasks like phase recognition and tool identification.

The addition of simple cognitive bias prompts significantly degrades performance. We introduce BiasMedQA to evaluate bias robustness on medical QA problems, and demonstrate mitigation techniques.

Surgical Gym is an open-source high performance platform for surgical robot learning where both the physics simulation and reinforcement learning occur directly on the GPU.

This work introduces an occupancy network-based method for the localization of tumors within kidney phantoms undergoing deformations at interactive speeds.

Comprehensive review of current brain-inspired learning representations in artificial neural networks.

Autonomous surgical robots have the potential to transform surgery and increase access to quality health care. Advances in artificial intelligence have produced robots mimicking human demonstrations. This application might be feasible for surgical robots but is associated with obstacles in creating robots that emulate surgeon demonstrations.

We open-source the largest dataset of general surgery videos to-date, consisting of 680 hours of surgical videos, including data from robotic and laparoscopic techniques across 28 procedures; we propose a technique for video pre-training a general surgery vision transformer (GSViT) on surgical videos based on forward video prediction that can run in real-time for surgical applications, toward which we open-source the code and weights of GSViT.

This paper outline challenges and training needs of junior researchers working across AI and neuroscience. We also provide advice and resources to help trainees plan their NeuroAI careers.

Preliminary progress towards learning a library of surgical manipulation skills using the da Vinci Research Kit (dVRK).

We show that when a patient proposes incorrect bias-validating information, the diagnostic accuracy of LLMs drop dramatically, revealing a high susceptibility to errors in self-diagnosis.

This paper introduces the Synaptic Motor Adaptation (SMA) algorithm, a novel approach to achieving real-time online adaptation in quadruped robots through the utilization of neuroscience-derived rules of synaptic plasticity with three-factor learning.

We introduce a bi-level optimization framework that seeks to both solve online learning tasks and improve the ability to learn online using models of plasticity from neuroscience.

We translate the motor circuit of the C. Elegans nematode into artificial neural networks at varying levels of biophysical realism and evaluate the outcome of training these networks on motor and non-motor behavioral tasks.

We construct locked fronts for a particular piecewise linear reproduction function. These fronts are shown to be linear combinations of exponentially decaying solutions to the linear system near the unstable state.

This work demonstrates that utilizing plasticity together with ANNs leads to instability beyond the pre-specified lifespan used during training. This instability can lead to the dramatic decline of reward seeking behavior, or quickly lead to reaching environment terminal states.

We introduce a framework for simultaneously learning the underlying fixed-weights and the rules governing the dynamics of synaptic plasticity and neuromodulated synaptic plasticity in SNNs through gradient descent.

We present a trajectory planning method for multiple vehicles to navigate a crowded environment, such as a gridlocked intersection or a small parking area.

This paper presents a simple method for learning neural architecture through random mutation.

We show quadrupedal agents evolved using self-modifying plastic networks are more capable of adapting to complex meta-learning learning tasks, even outperforming the same network updated using gradient-based algorithms while taking less time to train.

Original source code.