Farhad Nooralahzadeh

Aron N. Horvath, Matteo Berchier, Farhad Nooralahzadeh et al.

Background: Federated learning methods offer the possibility of training machine learning models on privacy-sensitive data sets, which cannot be easily shared. Multiple regulations pose strict requirements on the storage and usage of healthcare data, leading to data being in silos (i.e. locked-in at healthcare facilities). The application of federated algorithms on these datasets could accelerate disease diagnostic, drug development, as well as improve patient care. Methods: We present an extensive evaluation of the impact of different federation and differential privacy techniques when training models on the open-source MIMIC-III dataset. We analyze a set of parameters influencing a federated model performance, namely data distribution (homogeneous and heterogeneous), communication strategies (communication rounds vs. local training epochs), federation strategies (FedAvg vs. FedProx). Furthermore, we assess and compare two differential privacy (DP) techniques during model training: a stochastic gradient descent-based differential privacy algorithm (DP-SGD), and a sparse vector differential privacy technique (DP-SVT). Results: Our experiments show that extreme data distributions across sites (imbalance either in the number of patients or the positive label ratios between sites) lead to a deterioration of model performance when trained using the FedAvg strategy. This issue is resolved when using FedProx with the use of appropriate hyperparameter tuning. Furthermore, the results show that both differential privacy techniques can reach model performances similar to those of models trained without DP, however at the expense of a large quantifiable privacy leakage. Conclusions: We evaluate empirically the benefits of two federation strategies and propose optimal strategies for the choice of parameters when using differential privacy techniques.

Farhad Nooralahzadeh, Rico Sennrich

While several benefits were realized for multilingual vision-language pretrained models, recent benchmarks across various tasks and languages showed poor cross-lingual generalisation when multilingually pre-trained vision-language models are applied to non-English data, with a large gap between (supervised) English performance and (zero-shot) cross-lingual transfer. In this work, we explore the poor performance of these models on a zero-shot cross-lingual visual question answering (VQA) task, where models are fine-tuned on English visual-question data and evaluated on 7 typologically diverse languages. We improve cross-lingual transfer with three strategies: (1) we introduce a linguistic prior objective to augment the cross-entropy loss with a similarity-based loss to guide the model during training, (2) we learn a task-specific subnetwork that improves cross-lingual generalisation and reduces variance without model modification, (3) we augment training examples using synthetic code-mixing to promote alignment of embeddings between source and target languages. Our experiments on xGQA using the pretrained multilingual multimodal transformers UC2 and M3P demonstrate the consistent effectiveness of the proposed fine-tuning strategy for 7 languages, outperforming existing transfer methods with sparse models. Code and data to reproduce our findings are publicly available.

Farhad Nooralahzadeh, Benjamin Gundersen, Nicolas Deperrois et al.

Medical vision-language models (VLMs) often hallucinate findings when generating chest X-ray reports: they fabricate findings that are not present in the image, miss important ones, or locate them incorrectly. We mitigate this without weight updates by decoding-time residual steering on a per-token sparse autoencoder (SAE) basis: Top-$K$ SAEs on late layers, causal steering against clinical errors, then combined suppress/boost intervention at inference time. On the MIMIC-CXR test split, our inference-only method improves the quality of generated reports for three radiology VLMs (RadVLM, LLaVA-Rad, and CheXOne), with relative improvements of +5.4%, +7.2%, and +17.0% in the clinical composite metric, and statistically significant GREEN gains on all backbones. A cross-model feature alignment shows that the quality-promoting (boost) directions overlap strongly across architectures, whereas hallucination-linked (suppress) directions are model-specific. Therefore, transferable steering must treat suppression per-backbone, rather than sharing a universal suppress list. The same recipe transfers zero-shot to IU-Xray (Green $+7.7\%$ rel.) without retraining, confirming that the identified features are properties of the model, not of the training corpus. We release causal feature sets and an interactive feature dashboard: https://cxr-sparse-feature-dashboard.netlify.app/.

Mélanie Roschewitz, Kenneth Styppa, Yitian Tao et al.

Vision-language models (VLM) have markedly advanced AI-driven interpretation and reporting of complex medical imaging, such as computed tomography (CT). Yet, existing methods largely relegate clinicians to passive observers of final outputs, offering no interpretable reasoning trace for them to inspect, validate, or refine. To address this, we introduce RadAgent, a tool-using AI agent that generates CT reports through a stepwise and interpretable process. Each resulting report is accompanied by a fully inspectable trace of intermediate decisions and tool interactions, allowing clinicians to examine how the reported findings are derived. In our experiments, we observe that RadAgent improves Chest CT report generation over its 3D VLM counterpart, CT-Chat, across three dimensions. Clinical accuracy improves by 6.0 points (36.4% relative) in macro-F1 and 5.4 points (19.6% relative) in micro-F1. Robustness under adversarial conditions improves by 24.7 points (41.9% relative). Furthermore, RadAgent achieves 37.0% in faithfulness, a new capability entirely absent in its 3D VLM counterpart. By structuring the interpretation of chest CT as an explicit, tool-augmented and iterative reasoning trace, RadAgent brings us closer toward transparent and reliable AI for radiology.

Amos Calamida, Farhad Nooralahzadeh, Morteza Rohanian et al.

We propose a new automated evaluation metric for machine-generated radiology reports using the successful COMET architecture adapted for the radiology domain. We train and publish four medically-oriented model checkpoints, including one trained on RadGraph, a radiology knowledge graph. Our results show that our metric correlates moderately to high with established metrics such as BERTscore, BLEU, and CheXbert scores. Furthermore, we demonstrate that one of our checkpoints exhibits a high correlation with human judgment, as assessed using the publicly available annotations of six board-certified radiologists, using a set of 200 reports. We also performed our own analysis gathering annotations with two radiologists on a collection of 100 reports. The results indicate the potential effectiveness of our method as a radiology-specific evaluation metric. The code, data, and model checkpoints to reproduce our findings will be publicly available.

Benjamin Gundersen, Nicolas Deperrois, Samuel Ruiperez-Campillo et al.

Recent advances in vision-language models (VLMs) have improved Chest X-ray (CXR) interpretation in multiple aspects. However, many medical VLMs rely solely on supervised fine-tuning (SFT), which optimizes next-token prediction without evaluating answer quality. In contrast, reinforcement learning (RL) can incorporate task-specific feedback, and its combination with explicit intermediate reasoning ("thinking") has demonstrated substantial gains on verifiable math and coding tasks. To investigate the effects of RL and thinking in a CXR VLM, we perform large-scale SFT on CXR data to build an updated RadVLM based on Qwen3-VL, followed by a cold-start SFT stage that equips the model with basic thinking ability. We then apply Group Relative Policy Optimization (GRPO) with clinically grounded, task-specific rewards for report generation and visual grounding, and run matched RL experiments on both domain-specific and general-domain Qwen3-VL variants, with and without thinking. Across these settings, we find that while strong SFT remains crucial for high base performance, RL provides additional gains on both tasks, whereas explicit thinking does not appear to further improve results. Under a unified evaluation pipeline, the RL-optimized RadVLM models outperform their baseline counterparts and reach state-of-the-art performance on both report generation and grounding, highlighting clinically aligned RL as a powerful complement to SFT for medical VLMs.

Jonathan Fürst, Catherine Kosten, Farhad Nooralahzadeh et al.

Text-to-SQL systems (also known as NL-to-SQL systems) have become an increasingly popular solution for bridging the gap between user capabilities and SQL-based data access. These systems translate user requests in natural language to valid SQL statements for a specific database. Recent Text-to-SQL systems have benefited from the rapid improvement of transformer-based language models. However, while Text-to-SQL systems that incorporate such models continuously reach new high scores on -- often synthetic -- benchmark datasets, a systematic exploration of their robustness towards different data models in a real-world, realistic scenario is notably missing. This paper provides the first in-depth evaluation of the data model robustness of Text-to-SQL systems in practice based on a multi-year international project focused on Text-to-SQL interfaces. Our evaluation is based on a real-world deployment of FootballDB, a system that was deployed over a 9 month period in the context of the FIFA World Cup 2022, during which about 6K natural language questions were asked and executed. All of our data is based on real user questions that were asked live to the system. We manually labeled and translated a subset of these questions for three different data models. For each data model, we explore the performance of representative Text-to-SQL systems and language models. We further quantify the impact of training data size, pre-, and post-processing steps as well as language model inference time. Our comprehensive evaluation sheds light on the design choices of real-world Text-to-SQL systems and their impact on moving from research prototypes to real deployments. Last, we provide a new benchmark dataset to the community, which is the first to enable the evaluation of different data models for the same dataset and is substantially more challenging than most previous datasets in terms of query complexity.

Farhad Nooralahzadeh, Yi Zhang, Jonathan Furst et al.

International enterprises, organizations, or hospitals collect large amounts of multi-modal data stored in databases, text documents, images, and videos. While there has been recent progress in the separate fields of multi-modal data exploration as well as in database systems that automatically translate natural language questions to database query languages, the research challenge of querying database systems combined with other unstructured modalities such as images in natural language is widely unexplored. In this paper, we propose XMODE - a system that enables explainable, multi-modal data exploration in natural language. Our approach is based on the following research contributions: (1) Our system is inspired by a real-world use case that enables users to explore multi-modal information systems. (2) XMODE leverages a LLM-based agentic AI framework to decompose a natural language question into subtasks such as text-to-SQL generation and image analysis. (3) Experimental results on multi-modal datasets over relational data and images demonstrate that our system outperforms state-of-the-art multi-modal exploration systems, excelling not only in accuracy but also in various performance metrics such as query latency, API costs, planning efficiency, and explanation quality, thanks to the more effective utilization of the reasoning capabilities of LLMs.

Nicolas Deperrois, Hidetoshi Matsuo, Samuel Ruipérez-Campillo et al.

The widespread use of chest X-rays (CXRs), coupled with a shortage of radiologists, has driven growing interest in automated CXR analysis and AI-assisted reporting. While existing vision-language models (VLMs) show promise in specific tasks such as report generation or abnormality detection, they often lack support for interactive diagnostic capabilities. In this work we present RadVLM, a compact, multitask conversational foundation model designed for CXR interpretation. To this end, we curate a large-scale instruction dataset comprising over 1 million image-instruction pairs containing both single-turn tasks -- such as report generation, abnormality classification, and visual grounding -- and multi-turn, multi-task conversational interactions. After fine-tuning RadVLM on this instruction dataset, we evaluate it across different tasks along with re-implemented baseline VLMs. Our results show that RadVLM achieves state-of-the-art performance in conversational capabilities and visual grounding while remaining competitive in other radiology tasks. Ablation studies further highlight the benefit of joint training across multiple tasks, particularly for scenarios with limited annotated data. Together, these findings highlight the potential of RadVLM as a clinically relevant AI assistant, providing structured CXR interpretation and conversational capabilities to support more effective and accessible diagnostic workflows.

Farhad Nooralahzadeh, Omid Rohanian, Yi Zhang et al.

When a Vision-Language Model (VLM) sees a blue banana and answers "yellow", is the problem of perception or arbitration? We explore the question in ten VLMs with various sizes and reveal an Encoding--Grounding Dissociation: models that fail to report what they see (and thus provide a wrong answer) still encode the visual evidence as strongly as models that provide the correct answer. Using Multimodal Arbitration Crossover (MAC) analysis with layer-by-layer Logit Lens probing, we track the competition between visual and prior signals across every layer of each model. We show that visual attributes can be linearly decodable from early layers (AUC > 0.86). The accuracy remains nearly identical for both successful and failed samples. However, the gap in the final-layer logit -- not the strength of encoding -- better predicts grounding outcomes with a correlation of . After having studied when VLMs base their answers on image clues rather than prior knowledge, we want to understand the causal relationships. We establish causality through full-sequence activation patching. The standard last-token interventions in LLM interpretability do not affect VLMs. In contrast, replacing the full token sequence at layers identified by MAC alters 60 to 84% of outputs. Partial-token decomposition shows that image tokens carry almost all of the causal impact, while text tokens have none. Scaling addresses the remaining architectural differences to achieve perfect retention. Moving from diagnosis to intervention, we show that training-free activation steering -- both linear and sparse autoencoder-guided -- in early layers can improve visual grounding by up to +3.8% with degrading performance in some setups. Overall, these findings lead to a clear conclusion: VLMs already see well, but the challenge is acting on what they see. Targeted interventions can help to bridge this gap.

Morteza Rohanian, Roya M. Hüppi, Farhad Nooralahzadeh et al.

Capturing subtle speech disruptions across the psychosis spectrum is challenging because of the inherent variability in speech patterns. This variability reflects individual differences and the fluctuating nature of symptoms in both clinical and non-clinical populations. Accounting for uncertainty in speech data is essential for predicting symptom severity and improving diagnostic precision. Speech disruptions characteristic of psychosis appear across the spectrum, including in non-clinical individuals. We develop an uncertainty-aware model integrating acoustic and linguistic features to predict symptom severity and psychosis-related traits. Quantifying uncertainty in specific modalities allows the model to address speech variability, improving prediction accuracy. We analyzed speech data from 114 participants, including 32 individuals with early psychosis and 82 with low or high schizotypy, collected through structured interviews, semi-structured autobiographical tasks, and narrative-driven interactions in German. The model improved prediction accuracy, reducing RMSE and achieving an F1-score of 83% with ECE = 4.5e-2, showing robust performance across different interaction contexts. Uncertainty estimation improved model interpretability by identifying reliability differences in speech markers such as pitch variability, fluency disruptions, and spectral instability. The model dynamically adjusted to task structures, weighting acoustic features more in structured settings and linguistic features in unstructured contexts. This approach strengthens early detection, personalized assessment, and clinical decision-making in psychosis-spectrum research.

Andrea Agostini, Sonia Laguna, Alain Ryser et al.

Building generalizable medical AI systems requires pretraining strategies that are data-efficient and domain-aware. Unlike internet-scale corpora, clinical datasets such as MIMIC-CXR offer limited image counts and scarce annotations, but exhibit rich internal structure through multi-view imaging. We propose a self-supervised framework that leverages the inherent structure of medical datasets. Specifically, we treat paired chest X-rays (i.e., frontal and lateral views) as natural positive pairs, learning to reconstruct each view from sparse patches while aligning their latent embeddings. Our method requires no textual supervision and produces informative representations. Evaluated on MIMIC-CXR, we show strong performance compared to supervised objectives and baselines being trained without leveraging structure. This work provides a lightweight, modality-agnostic blueprint for domain-specific pretraining where data is structured but scarce

Sonia Laguna, Andrea Agostini, Alain Ryser et al.

Building robust medical machine learning systems requires pretraining strategies that exploit the intrinsic structure present in clinical data. We introduce Multiview Masked Autoencoder (MVMAE), a self-supervised framework that leverages the natural multi-view organization of radiology studies to learn view-invariant and disease-relevant representations. MVMAE combines masked image reconstruction with cross-view alignment, transforming clinical redundancy across projections into a powerful self-supervisory signal. We further extend this approach with MVMAE-V2T, which incorporates radiology reports as an auxiliary text-based learning signal to enhance semantic grounding while preserving fully vision-based inference. Evaluated on a downstream disease classification task on three large-scale public datasets, MIMIC-CXR, CheXpert, and PadChest, MVMAE consistently outperforms supervised and vision-language baselines. Furthermore, MVMAE-V2T provides additional gains, particularly in low-label regimes where structured textual supervision is most beneficial. Together, these results establish the importance of structural and textual supervision as complementary paths toward scalable, clinically grounded medical foundation models.

Christian Bluethgen, Dave Van Veen, Daniel Truhn et al.

Building agents, systems that perceive and act upon their environment with a degree of autonomy, has long been a focus of AI research. This pursuit has recently become vastly more practical with the emergence of large language models (LLMs) capable of using natural language to integrate information, follow instructions, and perform forms of "reasoning" and planning across a wide range of tasks. With its multimodal data streams and orchestrated workflows spanning multiple systems, radiology is uniquely suited to benefit from agents that can adapt to context and automate repetitive yet complex tasks. In radiology, LLMs and their multimodal variants have already demonstrated promising performance for individual tasks such as information extraction and report summarization. However, using LLMs in isolation underutilizes their potential to support complex, multi-step workflows where decisions depend on evolving context from multiple information sources. Equipping LLMs with external tools and feedback mechanisms enables them to drive systems that exhibit a spectrum of autonomy, ranging from semi-automated workflows to more adaptive agents capable of managing complex processes. This review examines the design of such LLM-driven agentic systems, highlights key applications, discusses evaluation methods for planning and tool use, and outlines challenges such as error cascades, tool-use efficiency, and health IT integration.

Morteza Rohanian, Tarun Mehra, Nicola Miglino et al.

Clinical oncology generates vast, unstructured data that often contain inconsistencies, missing information, and ambiguities, making it difficult to extract reliable insights for data-driven decision-making. General-purpose large language models (LLMs) struggle with these challenges due to their lack of domain-specific reasoning, including specialized clinical terminology, context-dependent interpretations, and multi-modal data integration. We address these issues with an oncology-specialized, efficient, and adaptable NLP framework that combines instruction tuning, retrieval-augmented generation (RAG), and graph-based knowledge integration. Our lightweight models prove effective at oncology-specific tasks, such as named entity recognition (e.g., identifying cancer diagnoses), entity linking (e.g., linking entities to standardized ontologies), TNM staging, document classification (e.g., cancer subtype classification from pathology reports), and treatment response prediction. Our framework emphasizes adaptability and resource efficiency. We include minimal German instructions, collected at the University Hospital Zurich (USZ), to test whether small amounts of non-English language data can effectively transfer knowledge across languages. This approach mirrors our motivation for lightweight models, which balance strong performance with reduced computational costs, making them suitable for resource-limited healthcare settings. We validated our models on oncology datasets, demonstrating strong results in named entity recognition, relation extraction, and document classification.

Farhad Nooralahzadeh, Yi Zhang, Ellery Smith et al.

The potential for improvements brought by Large Language Models (LLMs) in Text-to-SQL systems is mostly assessed on monolingual English datasets. However, LLMs' performance for other languages remains vastly unexplored. In this work, we release the StatBot.Swiss dataset, the first bilingual benchmark for evaluating Text-to-SQL systems based on real-world applications. The StatBot.Swiss dataset contains 455 natural language/SQL-pairs over 35 big databases with varying level of complexity for both English and German. We evaluate the performance of state-of-the-art LLMs such as GPT-3.5-Turbo and mixtral-8x7b-instruct for the Text-to-SQL translation task using an in-context learning approach. Our experimental analysis illustrates that current LLMs struggle to generalize well in generating SQL queries on our novel bilingual dataset.

Sanghwan Kim, Farhad Nooralahzadeh, Morteza Rohanian et al.

Recent transformer-based models have made significant strides in generating radiology reports from chest X-ray images. However, a prominent challenge remains: these models often lack prior knowledge, resulting in the generation of synthetic reports that mistakenly reference non-existent prior exams. This discrepancy can be attributed to a knowledge gap between radiologists and the generation models. While radiologists possess patient-specific prior information, the models solely receive X-ray images at a specific time point. To tackle this issue, we propose a novel approach that leverages a rule-based labeler to extract comparison prior information from radiology reports. This extracted comparison prior is then seamlessly integrated into state-of-the-art transformer-based models, enabling them to produce more realistic and comprehensive reports. Our method is evaluated on English report datasets, such as IU X-ray and MIMIC-CXR. The results demonstrate that our approach surpasses baseline models in terms of natural language generation metrics. Notably, our model generates reports that are free from false references to non-existent prior exams, setting it apart from previous models. By addressing this limitation, our approach represents a significant step towards bridging the gap between radiologists and generation models in the domain of medical report generation.

Farhad Nooralahzadeh, Nicolas Perez Gonzalez, Thomas Frauenfelder et al.

Inspired by Curriculum Learning, we propose a consecutive (i.e., image-to-text-to-text) generation framework where we divide the problem of radiology report generation into two steps. Contrary to generating the full radiology report from the image at once, the model generates global concepts from the image in the first step and then reforms them into finer and coherent texts using a transformer architecture. We follow the transformer-based sequence-to-sequence paradigm at each step. We improve upon the state-of-the-art on two benchmark datasets.

Farhad Nooralahzadeh

Real-world applications of natural language processing (NLP) are challenging. NLP models rely heavily on supervised machine learning and require large amounts of annotated data. These resources are often based on language data available in large quantities, such as English newswire. However, in real-world applications of NLP, the textual resources vary across several dimensions, such as language, dialect, topic, and genre. It is challenging to find annotated data of sufficient amount and quality. The objective of this thesis is to investigate methods for dealing with such low-resource scenarios in information extraction and natural language understanding. To this end, we study distant supervision and sequential transfer learning in various low-resource settings. We develop and adapt neural NLP models to explore a number of research questions concerning NLP tasks with minimal or no training data.

Farhad Nooralahzadeh, Giannis Bekoulis, Johannes Bjerva et al.

Learning what to share between tasks has been a topic of great importance recently, as strategic sharing of knowledge has been shown to improve downstream task performance. This is particularly important for multilingual applications, as most languages in the world are under-resourced. Here, we consider the setting of training models on multiple different languages at the same time, when little or no data is available for languages other than English. We show that this challenging setup can be approached using meta-learning, where, in addition to training a source language model, another model learns to select which training instances are the most beneficial to the first. We experiment using standard supervised, zero-shot cross-lingual, as well as few-shot cross-lingual settings for different natural language understanding tasks (natural language inference, question answering). Our extensive experimental setup demonstrates the consistent effectiveness of meta-learning for a total of 15 languages. We improve upon the state-of-the-art for zero-shot and few-shot NLI (on MultiNLI and XNLI) and QA (on the MLQA dataset). A comprehensive error analysis indicates that the correlation of typological features between languages can partly explain when parameter sharing learned via meta-learning is beneficial.

Farhad Nooralahzadeh, Lilja Øvrelid

We investigate the use of different syntactic dependency representations in a neural relation classification task and compare the CoNLL, Stanford Basic and Universal Dependencies schemes. We further compare with a syntax-agnostic approach and perform an error analysis in order to gain a better understanding of the results.

Farhad Nooralahzadeh, Lilja Øvrelid, Jan Tore Lønning

This article presents the SIRIUS-LTG-UiO system for the SemEval 2018 Task 7 on Semantic Relation Extraction and Classification in Scientific Papers. First we extract the shortest dependency path (sdp) between two entities, then we introduce a convolutional neural network (CNN) which takes the shortest dependency path embeddings as input and performs relation classification with differing objectives for each subtask of the shared task. This approach achieved overall F1 scores of 76.7 and 83.2 for relation classification on clean and noisy data, respectively. Furthermore, for combined relation extraction and classification on clean data, it obtained F1 scores of 37.4 and 33.6 for each phase. Our system ranks 3rd in all three sub-tasks of the shared task.