Sophont

NEW RELEASE: Today we're releasing CortexMAE: a family of fMRI foundation models trained on 2.1K hours of open fMRI data. We're also releasing Brainmarks: an open benchmark suite for evaluating fMRI foundation models. Full paper is on arXiv (accepted to ICML 2026) blog post: https://lnkd.in/gbXMjqcJ paper: https://lnkd.in/ghjBMNvs github: https://lnkd.in/gdAmRzUi model: https://lnkd.in/gZCRb_W5 benchmarks: https://lnkd.in/gA6v26WE A quick summary (see blog post/paper for more info): At Sophont, we are interested in developing foundation models for neuroimaging to advance the diagnosis and treatment of neurodegenerative and other mental disorders. The fMRI neuroimaging method measures brain activity by detecting changes associated with blood flow over time. Previous fMRI foundation models have taken two main approaches: 1. Parcellation: reduce each 3D fMRI volume to a list of ~400 numbers by averaging within a fixed set of “parcels” 2. Volume: model the native 3D data directly. We propose an intermediate approach: flat maps! In our flat map approach, we represent fMRI activity as 2D maps overlaid on a flattened cortical surface mesh. This maintains the full cortical fMRI signal (like volume approaches), while also explicitly injecting the inductive bias of local cortical neighborhoods (like parcellation approaches) To train ViTs on sequences of fMRI flat maps, we adopt the spatiotemporal masked autoencoder (MAE) framework. In order to evaluate fMRI foundation models we introduce Brainmarks: an open reproducible evaluation suite for this purpose! Brainmarks currently includes >30 benchmark tasks across 7 publicly available source datasets. In our paper, we focus on a core group of 8 benchmarks. CortexMAE is able to do denoising of fMRI data. A lot of interesting dynamics emerge in the reconstruction that are difficult to see in the original data (default mode network oscillation, traveling waves in visual cortex, large activation of the sensorimotor hand areas). The unique value of fMRI is its ability to capture dynamic patterns of brain activity. CortexMAE excels at representing these dynamic patterns. Our models are able to decode dynamic cognitive states with SOTA accuracy. The ultimate goal of an fMRI foundation model is to help understand, diagnose, and treat brain and mental health conditions. CortexMAE is better than existing models on clinical diagnosis tasks, but all foundation models struggle to beat simple baselines. The field of neuro foundation models is still in its early stages. We're currently working on building multimodal MRI foundation models. If you're interested in working with us on research or exploring commercial partnerships, contact us directly or join the MedARC Discord!

Hi! This week on The Information Bottleneck we will host Tanishq Abraham, PhD, Co-Founder and CEO of Sophont and founder of MedARC, to discuss open-source multimodal medical foundation models, applying AI to healthcare, and building the "DeepSeek of medical AI." Leave your questions in the comments and we'll try to ask him!

We're releasing Medmarks v0.1, the largest completely open-source automated evaluation suite for assessing the medical capabilities of LLMs! Developed in our MedARC community, with support from Prime Intellect So far we’ve explored 46 models to figure out the best! Why did we build this benchmark suite? Because there isn't yet a completely open and easy-to-run medical LLM benchmark, evaluated on various kinds of realistic tasks and updated regularly with new models. We aggregated a total of 28 tasks in 20 benchmarks in our suite. We divide them into two subsets: 1. Medmarks-Verifiable: 14 verifiable benchmarks, mostly multiple-choice question answering but also other tasks like medical calculations 2. Medmarks-OE: 6 open-ended benchmarks, e.g. answering patient questions On average, we find that GPT-5.1, GPT-5.2, and Qwen3-235B-A22B-Thinking are the best-performing models on the medical tasks evaluated. Lots of interesting observations: • medical-specific LLMs can be quite performant • open weight models are close to proprietary model perf but less token efficient • reasoning post-training improves performance • There's a few standout datasets that aren't saturated yet We built our evaluation suite on top of Prime Intellect's verifiers library. This provides us many advantages: 1. RL environments for free for datasets with training split 2. integration with Inference API enables easy benchmarking of model APIs 3. benchmarks available on Hub (under the MedARC organization) This is only just the beginning! We plan to add more models and benchmarks, and of course develop our own models and benchmarks! If you're interested in collaborating, join the MedARC Discord or contact us (contact@sophontai.com) To learn more about our benchmark suite, leaderboard, and interesting findings, read our blog post: https://lnkd.in/gNeZRVdw Check out the leaderboard: https://medmarks.ai code: https://lnkd.in/gNwmaXPZ environment hub: https://lnkd.in/gPCv3NUJ

We released our first paper, check out the article for a quick summary!

MedARC is relaunched and we've got 3 exciting research projects available for collaboration! fMRI foundation models, histopathology foundation models, and real-time brain-to-image decoding

Check out Paul Scotti's perspective piece on the winners of the Algonauts 2025 challenge where teams competed to best predict brain activity in response to movie viewing!