Mary-Anne Hartley

Zeming Chen, Alejandro Hernández Cano, Angelika Romanou et al. · allen-ai

Large language models (LLMs) can potentially democratize access to medical knowledge. While many efforts have been made to harness and improve LLMs' medical knowledge and reasoning capacities, the resulting models are either closed-source (e.g., PaLM, GPT-4) or limited in scale (<= 13B parameters), which restricts their abilities. In this work, we improve access to large-scale medical LLMs by releasing MEDITRON: a suite of open-source LLMs with 7B and 70B parameters adapted to the medical domain. MEDITRON builds on Llama-2 (through our adaptation of Nvidia's Megatron-LM distributed trainer), and extends pretraining on a comprehensively curated medical corpus, including selected PubMed articles, abstracts, and internationally-recognized medical guidelines. Evaluations using four major medical benchmarks show significant performance gains over several state-of-the-art baselines before and after task-specific finetuning. Overall, MEDITRON achieves a 6% absolute performance gain over the best public baseline in its parameter class and 3% over the strongest baseline we finetuned from Llama-2. Compared to closed-source LLMs, MEDITRON-70B outperforms GPT-3.5 and Med-PaLM and is within 5% of GPT-4 and 10% of Med-PaLM-2. We release our code for curating the medical pretraining corpus and the MEDITRON model weights to drive open-source development of more capable medical LLMs.

Vinitra Swamy, Malika Satayeva, Jibril Frej et al.

Predicting multiple real-world tasks in a single model often requires a particularly diverse feature space. Multimodal (MM) models aim to extract the synergistic predictive potential of multiple data types to create a shared feature space with aligned semantic meaning across inputs of drastically varying sizes (i.e. images, text, sound). Most current MM architectures fuse these representations in parallel, which not only limits their interpretability but also creates a dependency on modality availability. We present MultiModN, a multimodal, modular network that fuses latent representations in a sequence of any number, combination, or type of modality while providing granular real-time predictive feedback on any number or combination of predictive tasks. MultiModN's composable pipeline is interpretable-by-design, as well as innately multi-task and robust to the fundamental issue of biased missingness. We perform four experiments on several benchmark MM datasets across 10 real-world tasks (predicting medical diagnoses, academic performance, and weather), and show that MultiModN's sequential MM fusion does not compromise performance compared with a baseline of parallel fusion. By simulating the challenging bias of missing not-at-random (MNAR), this work shows that, contrary to MultiModN, parallel fusion baselines erroneously learn MNAR and suffer catastrophic failure when faced with different patterns of MNAR at inference. To the best of our knowledge, this is the first inherently MNAR-resistant approach to MM modeling. In conclusion, MultiModN provides granular insights, robustness, and flexibility without compromising performance.

Yusuf Kesmen, Fay Elhassan, Jiayi Ma et al.

Large language models are increasingly deployed as autonomous diagnostic agents, yet they conflate two fundamentally different capabilities: natural-language communication and probabilistic reasoning. We argue that this conflation is an architectural flaw, not an engineering shortcoming. We introduce BMBE (Bayesian Medical Belief Engine), a modular diagnostic dialogue framework that enforces a strict separation between language and reasoning: an LLM serves only as a sensor, parsing patient utterances into structured evidence and verbalising questions, while all diagnostic inference resides in a deterministic, auditable Bayesian engine. Because patient data never enters the LLM, the architecture is private by construction; because the statistical backend is a standalone module, it can be replaced per target population without retraining. This separation yields three properties no autonomous LLM can offer: calibrated selective diagnosis with a continuously adjustable accuracy-coverage tradeoff, a statistical separation gap where even a cheap sensor paired with the engine outperforms a frontier standalone model from the same family at a fraction of the cost, and robustness to adversarial patient communication styles that cause standalone doctors to collapse. We validate across empirical and LLM-generated knowledge bases against frontier LLMs, confirming the advantage is architectural, not informational.

Cécile Trottet, Thijs Vogels, Martin Jaggi et al.

Data-driven Clinical Decision Support Systems (CDSS) have the potential to improve and standardise care with personalised probabilistic guidance. However, the size of data required necessitates collaborative learning from analogous CDSS's, which are often unsharable or imperfectly interoperable (IIO), meaning their feature sets are not perfectly overlapping. We propose Modular Clinical Decision Support Networks (MoDN) which allow flexible, privacy-preserving learning across IIO datasets, while providing interpretable, continuous predictive feedback to the clinician. MoDN is a novel decision tree composed of feature-specific neural network modules. It creates dynamic personalised representations of patients, and can make multiple predictions of diagnoses, updatable at each step of a consultation. The modular design allows it to compartmentalise training updates to specific features and collaboratively learn between IIO datasets without sharing any data.

Mohammad Iqbal Nouyed, Mary-Anne Hartley, Gianfranco Doretto et al.

This work addresses how to efficiently classify challenging histopathology images, such as gigapixel whole-slide images for cancer diagnostics with image-level annotation. We use images with annotated tumor regions to identify a set of tumor patches and a set of benign patches in a cancerous slide. Due to the variable nature of region of interest the tumor positive regions may refer to an extreme minority of the pixels. This creates an important problem during patch-level classification, where the majority of patches from an image labeled as 'cancerous' are actually tumor-free. This problem is different from semantic segmentation which associates a label to every pixel in an image, because after patch extraction we are only dealing with patch-level labels.Most existing approaches address the data imbalance issue by mitigating the data shortage in minority classes in order to prevent the model from being dominated by the majority classes. These methods include data re-sampling, loss re-weighting, margin modification, and data augmentation. In this work, we mitigate the patch-level class imbalance problem by taking a divide-and-conquer approach. First, we partition the data into sub-groups, and define three separate classification sub-problems based on these data partitions. Then, using an information-theoretic cluster-based sampling of deep image patch features, we sample discriminative patches from the sub-groups. Using these sampled patches, we build corresponding deep models to solve the new classification sub-problems. Finally, we integrate information learned from the respective models to make a final decision on the patches. Our result shows that the proposed approach can perform competitively using a very low percentage of the available patches in a given whole-slide image.

Xavier Theimer-Lienhard, Mushtaha El-Amin, Fay Elhassan et al.

Clinical decision support systems (CDSS) require scrutable, auditable pipelines that enable rigorous, reproducible validation. Yet current LLM-based CDSS remain largely opaque. Most "open" models are open-weight only, releasing parameters while withholding the data provenance, curation procedures, and generation pipelines that determine model behavior. Fully Open (FO) models, which expose the complete training stack end-to-end, do not currently exist in medicine. We introduce Fully Open Meditron, the first fully open pipeline for building LLM-CDSS, comprising a clinician-audited training corpus, a reproducible data construction and training framework, and a use-aligned evaluation protocol. The corpus unifies eight public medical QA datasets into a normalized conversational format and expands coverage with three clinician-vetted synthetic extensions: exam-style QA, guideline-grounded QA derived from 46,469 clinical practice guidelines, and clinical vignettes. The pipeline enforces system-wide decontamination, gold-label resampling of teacher generations, and end-to-end validation by a four-physician panel. We evaluate using an LLM-as-a-judge protocol over expert-written clinical vignettes, calibrated against 204 human raters. We apply the recipe to five FO base models (Apertus-70B/8B-Instruct, OLMo-2-32B-SFT, EuroLLM-22B/9B-Instruct). All MeditronFO variants are preferred over their bases. Apertus-70B-MeditronFO improves +6.6 points over its base (47.2% to 53.8%) on aggregate medical benchmarks, establishing a new FO SoTA. Gemma-3-27B-MeditronFO is preferred over MedGemma in 58.6% of LLM-as-a-judge comparisons and outperforms it on HealthBench (58% vs 55.9%). These results show that fully open pipelines can achieve state-of-the-art domain-specific performance without sacrificing auditability or reproducibility.

Julien T. T. Vignoud, Valérian Rousset, Hugo El Guedj et al.

Data is often impractical to share for a range of well considered reasons, such as concerns over privacy, intellectual property, and legal constraints. This not only fragments the statistical power of predictive models, but creates an accessibility bias, where accuracy becomes inequitably distributed to those who have the resources to overcome these concerns. We present DISCO: an open-source DIStributed COllaborative learning platform accessible to non-technical users, offering a means to collaboratively build machine learning models without sharing any original data or requiring any programming knowledge. DISCO's web application trains models locally directly in the browser, making our tool cross-platform out-of-the-box, including smartphones. The modular design of \disco offers choices between federated and decentralized paradigms, various levels of privacy guarantees and several approaches to weight aggregation strategies that allow for model personalization and bias resilience in the collaborative training. Code repository is available at https://github.com/epfml/disco and a showcase web interface at https://discolab.ai

Alexandre Sallinen, Stefan Krsteski, Paul Teiletche et al.

We introduce MMORE, an open-source pipeline for Massive Multimodal Open RetrievalAugmented Generation and Extraction, designed to ingest, transform, and retrieve knowledge from heterogeneous document formats at scale. MMORE supports more than fifteen file types, including text, tables, images, emails, audio, and video, and processes them into a unified format to enable downstream applications for LLMs. The architecture offers modular, distributed processing, enabling scalable parallelization across CPUs and GPUs. On processing benchmarks, MMORE demonstrates a 3.8-fold speedup over single-node baselines and 40% higher accuracy than Docling on scanned PDFs. The pipeline integrates hybrid dense-sparse retrieval and supports both interactive APIs and batch RAG endpoints. Evaluated on PubMedQA, MMORE-augmented medical LLMs improve biomedical QA accuracy with increasing retrieval depth. MMORE provides a robust, extensible foundation for deploying task-agnostic RAG systems on diverse, real-world multimodal data. The codebase is available at https://github.com/swiss-ai/mmore.

Ling Han, Nanqing Luo, Hao Huang et al.

This work delves into the complexities of machine unlearning in the face of distributional shifts, particularly focusing on the challenges posed by non-uniform feature and label removal. With the advent of regulations like the GDPR emphasizing data privacy and the right to be forgotten, machine learning models face the daunting task of unlearning sensitive information without compromising their integrity or performance. Our research introduces a novel approach that leverages influence functions and principles of distributional independence to address these challenges. By proposing a comprehensive framework for machine unlearning, we aim to ensure privacy protection while maintaining model performance and adaptability across varying distributions. Our method not only facilitates efficient data removal but also dynamically adjusts the model to preserve its generalization capabilities. Through extensive experimentation, we demonstrate the efficacy of our approach in scenarios characterized by significant distributional shifts, making substantial contributions to the field of machine unlearning. This research paves the way for developing more resilient and adaptable unlearning techniques, ensuring models remain robust and accurate in the dynamic landscape of data privacy and machine learning.

Hojjat Karami, Mary-Anne Hartley, David Atienza et al.

Time series in Electronic Health Records (EHRs) present unique challenges for generative models, such as irregular sampling, missing values, and high dimensionality. In this paper, we propose a novel generative adversarial network (GAN) model, TimEHR, to generate time series data from EHRs. In particular, TimEHR treats time series as images and is based on two conditional GANs. The first GAN generates missingness patterns, and the second GAN generates time series values based on the missingness pattern. Experimental results on three real-world EHR datasets show that TimEHR outperforms state-of-the-art methods in terms of fidelity, utility, and privacy metrics.

Minghui Liu, Tahseen Rabbani, Tony O'Halloran et al.

Transformer-based large language models (LLMs) use the key-value (KV) cache to significantly accelerate inference by storing the key and value embeddings of past tokens. However, this cache consumes significant GPU memory. In this work, we introduce HashEvict, an algorithm that uses locality-sensitive hashing (LSH) to compress the KV cache. HashEvict quickly locates tokens in the cache that are cosine dissimilar to the current query token. This is achieved by computing the Hamming distance between binarized Gaussian projections of the current token query and cached token keys, with a projection length much smaller than the embedding dimension. We maintain a lightweight binary structure in GPU memory to facilitate these calculations. Unlike existing compression strategies that compute attention to determine token retention, HashEvict makes these decisions pre-attention, thereby reducing computational costs. Additionally, HashEvict is dynamic - at every decoding step, the key and value of the current token replace the embeddings of a token expected to produce the lowest attention score. We demonstrate that HashEvict can compress the KV cache by 30%-70% while maintaining high performance across reasoning, multiple-choice, long-context retrieval and summarization tasks.

Ling Han, Hao Huang, Dustin Scheinost et al.

Effective adaptation to distribution shifts in training data is pivotal for sustaining robustness in neural networks, especially when removing specific biases or outdated information, a process known as machine unlearning. Traditional approaches typically assume that data variations are random, which makes it difficult to adjust the model parameters accurately to remove patterns and characteristics from unlearned data. In this work, we present Unlearning Information Bottleneck (UIB), a novel information-theoretic framework designed to enhance the process of machine unlearning that effectively leverages the influence of systematic patterns and biases for parameter adjustment. By proposing a variational upper bound, we recalibrate the model parameters through a dynamic prior that integrates changes in data distribution with an affordable computational cost, allowing efficient and accurate removal of outdated or unwanted data patterns and biases. Our experiments across various datasets, models, and unlearning methods demonstrate that our approach effectively removes systematic patterns and biases while maintaining the performance of models post-unlearning.

Al Zadid Sultan Bin Habib, Kesheng Wang, Mary-Anne Hartley et al.

Effective analysis of tabular data still poses a significant problem in deep learning, mainly because features in tabular datasets are often heterogeneous and have different levels of relevance. This work introduces TabSeq, a novel framework for the sequential ordering of features, addressing the vital necessity to optimize the learning process. Features are not always equally informative, and for certain deep learning models, their random arrangement can hinder the model's learning capacity. Finding the optimum sequence order for such features could improve the deep learning models' learning process. The novel feature ordering technique we provide in this work is based on clustering and incorporates both local ordering and global ordering. It is designed to be used with a multi-head attention mechanism in a denoising autoencoder network. Our framework uses clustering to align comparable features and improve data organization. Multi-head attention focuses on essential characteristics, whereas the denoising autoencoder highlights important aspects by rebuilding from distorted inputs. This method improves the capability to learn from tabular data while lowering redundancy. Our research, demonstrating improved performance through appropriate feature sequence rearrangement using raw antibody microarray and two other real-world biomedical datasets, validates the impact of feature ordering. These results demonstrate that feature ordering can be a viable approach to improved deep learning of tabular data.

Felix Grimberg, Mary-Anne Hartley, Sai P. Karimireddy et al.

In federated learning, differences in the data or objectives between the participating nodes motivate approaches to train a personalized machine learning model for each node. One such approach is weighted averaging between a locally trained model and the global model. In this theoretical work, we study weighted model averaging for arbitrary scalar mean estimation problems under minimal assumptions on the distributions. In a variant of the bias-variance trade-off, we find that there is always some positive amount of model averaging that reduces the expected squared error compared to the local model, provided only that the local model has a non-zero variance. Further, we quantify the (possibly negative) benefit of weighted model averaging as a function of the weight used and the optimal weight. Taken together, this work formalizes an approach to quantify the value of personalization in collaborative learning and provides a framework for future research to test the findings in multivariate parameter estimation and under a range of assumptions.

Martin Beaussart, Felix Grimberg, Mary-Anne Hartley et al.

In federated learning, model personalization can be a very effective strategy to deal with heterogeneous training data across clients. We introduce WAFFLE (Weighted Averaging For Federated LEarning), a personalized collaborative machine learning algorithm that leverages stochastic control variates for faster convergence. WAFFLE uses the Euclidean distance between clients' updates to weigh their individual contributions and thus minimize the personalized model loss on the specific agent of interest. Through a series of experiments, we compare our new approach to two recent personalized federated learning methods--Weight Erosion and APFL--as well as two general FL methods--Federated Averaging and SCAFFOLD. Performance is evaluated using two categories of non-identical client data distributions--concept shift and label skew--on two image data sets (MNIST and CIFAR10). Our experiments demonstrate the comparative effectiveness of WAFFLE, as it achieves or improves accuracy with faster convergence.

David Roschewitz, Mary-Anne Hartley, Luca Corinzia et al.

Recently, the ever-growing demand for privacy-oriented machine learning has motivated researchers to develop federated and decentralized learning techniques, allowing individual clients to train models collaboratively without disclosing their private datasets. However, widespread adoption has been limited in domains relying on high levels of user trust, where assessment of data compatibility is essential. In this work, we define and address low interoperability induced by underlying client data inconsistencies in federated learning for tabular data. The proposed method, iFedAvg, builds on federated averaging adding local element-wise affine layers to allow for a personalized and granular understanding of the collaborative learning process. Thus, enabling the detection of outlier datasets in the federation and also learning the compensation for local data distribution shifts without sharing any original data. We evaluate iFedAvg using several public benchmarks and a previously unstudied collection of real-world datasets from the 2014 - 2016 West African Ebola epidemic, jointly forming the largest such dataset in the world. In all evaluations, iFedAvg achieves competitive average performance with negligible overhead. It additionally shows substantial improvement on outlier clients, highlighting increased robustness to individual dataset shifts. Most importantly, our method provides valuable client-specific insights at a fine-grained level to guide interoperable federated learning.