Anna Arias-Duart

Ashwin Kumar Gururajan, Enrique Lopez-Cuena, Jordi Bayarri-Planas et al.

As the capabilities of Large Language Models (LLMs) in healthcare and medicine continue to advance, there is a growing need for competitive open-source models that can safeguard public interest. With the increasing availability of highly competitive open base models, the impact of continued pre-training is increasingly uncertain. In this work, we explore the role of instruct tuning, model merging, alignment, red teaming and advanced inference schemes, as means to improve current open models. To that end, we introduce the Aloe family, a set of open medical LLMs highly competitive within its scale range. Aloe models are trained on the current best base models (Mistral, LLaMA 3), using a new custom dataset which combines public data sources improved with synthetic Chain of Thought (CoT). Aloe models undergo an alignment phase, becoming one of the first few policy-aligned open healthcare LLM using Direct Preference Optimization, setting a new standard for ethical performance in healthcare LLMs. Model evaluation expands to include various bias and toxicity datasets, a dedicated red teaming effort, and a much-needed risk assessment for healthcare LLMs. Finally, to explore the limits of current LLMs in inference, we study several advanced prompt engineering strategies to boost performance across benchmarks, yielding state-of-the-art results for open healthcare 7B LLMs, unprecedented at this scale.

Javier Ferrando, Enrique Lopez-Cuena, Pablo Agustin Martin-Torres et al.

Sparse Autoencoders uncover thousands of features in vision models, yet explaining these features without requiring human intervention remains an open challenge. While previous work has proposed generating correlation-based explanations based on top activating input examples, we present a fundamentally different alternative based on causal interventions. We leverage the structure of Vision-Language Models and steer individual SAE features in the vision encoder after providing an empty image. Then, we prompt the language model to explain what it ``sees'', effectively eliciting the visual concept represented by each feature. Results show that Steering offers an scalable alternative that complements traditional approaches based on input examples, serving as a new axis for automated interpretability in vision models. Moreover, the quality of explanations improves consistently with the scale of the language model, highlighting our method as a promising direction for future research. Finally, we propose Steering-informed Top-k, a hybrid approach that combines the strengths of causal interventions and input-based approaches to achieve state-of-the-art explanation quality without additional computational cost.

Miquel Miró-Nicolau, Gabriel Moyà-Alcover, Anna Arias-Duart

The evaluation of eXplainable Artificial Intelligence (XAI) methods is a rapidly growing field, characterized by a wide variety of approaches. This diversity highlights the complexity of the XAI evaluation, which, unlike traditional AI assessment, lacks a universally correct ground truth for the explanation, making objective evaluation challenging. One promising direction to address this issue involves the use of what we term Synthetic Artificial Intelligence Ground truth (SAIG) methods, which generate artificial ground truths to enable the direct evaluation of XAI techniques. This paper presents the first review and analysis of SAIG methods. We introduce a novel taxonomy to classify these approaches, identifying seven key features that distinguish different SAIG methods. Our comparative study reveals a concerning lack of consensus on the most effective XAI evaluation techniques, underscoring the need for further research and standardization in this area.

Dario Garcia-Gasulla, Jordi Bayarri-Planas, Ashwin Kumar Gururajan et al.

Purpose: With advancements in Large Language Models (LLMs) for healthcare, the need arises for competitive open-source models to protect the public interest. This work contributes to the field of open medical LLMs by optimizing key stages of data preprocessing and training, while showing how to improve model safety (through DPO) and efficacy (through RAG). The evaluation methodology used, which includes four different types of tests, defines a new standard for the field. The resultant models, shown to be competitive with the best private alternatives, are released with a permisive license. Methods: Building on top of strong base models like Llama 3.1 and Qwen 2.5, Aloe Beta uses a custom dataset to enhance public data with synthetic Chain of Thought examples. The models undergo alignment with Direct Preference Optimization, emphasizing ethical and policy-aligned performance in the presence of jailbreaking attacks. Evaluation includes close-ended, open-ended, safety and human assessments, to maximize the reliability of results. Results: Recommendations are made across the entire pipeline, backed by the solid performance of the Aloe Family. These models deliver competitive performance across healthcare benchmarks and medical fields, and are often preferred by healthcare professionals. On bias and toxicity, the Aloe Beta models significantly improve safety, showing resilience to unseen jailbreaking attacks. For a responsible release, a detailed risk assessment specific to healthcare is attached to the Aloe Family models. Conclusion: The Aloe Beta models, and the recipe that leads to them, are a significant contribution to the open-source medical LLM field, offering top-of-the-line performance while maintaining high ethical requirements. This work sets a new standard for developing and reporting aligned LLMs in healthcare.

Anna Arias-Duart, Pablo Agustin Martin-Torres, Daniel Hinjos et al.

Current Large Language Models (LLMs) benchmarks are often based on open-ended or close-ended QA evaluations, avoiding the requirement of human labor. Close-ended measurements evaluate the factuality of responses but lack expressiveness. Open-ended capture the model's capacity to produce discourse responses but are harder to assess for correctness. These two approaches are commonly used, either independently or together, though their relationship remains poorly understood. This work is focused on the healthcare domain, where both factuality and discourse matter greatly. It introduces a comprehensive, multi-axis suite for healthcare LLM evaluation, exploring correlations between open and close benchmarks and metrics. Findings include blind spots and overlaps in current methodologies. As an updated sanity check, we release a new medical benchmark --CareQA-- with both open and closed variants. Finally, we propose a novel metric for open-ended evaluations -- Relaxed Perplexity -- to mitigate the identified limitations.

Dario Garcia-Gasulla, Adrian Tormos, Anna Arias-Duart et al.

Direct Preference Optimization (DPO) is an efficient alignment technique that steers LLMs towards preferable outputs by training on preference data, bypassing the need for explicit reward models. Its simplicity enables easy adaptation to various domains and safety requirements. This paper examines DPO's effectiveness in model safety against jailbreaking attacks while minimizing data requirements and training costs. We introduce Egida, a dataset expanded from multiple sources, which includes 27 different safety topics and 18 different attack styles, complemented with synthetic and human labels. This data is used to boost the safety of state-of-the-art LLMs (Llama-3.1-8B/70B-Instruct, Qwen-2.5-7B/72B-Instruct) across topics and attack styles. In addition to safety evaluations, we assess their post-alignment performance degradation in general purpose tasks, and their tendency to over refusal. Following the proposed methodology, trained models reduce their Attack Success Rate by 10%-30%, using small training efforts (2,000 samples) with low computational cost (3\$ for 8B models, 20\$ for 72B models). Safety aligned models generalize to unseen topics and attack styles, with the most successful attack style reaching a success rate around 5%. Size and family are found to strongly influence model malleability towards safety, pointing at the importance of pre-training choices. To validate our findings, a large independent assessment of human preference agreement with Llama-Guard-3-8B is conducted by the authors and the associated dataset Egida-HSafe is released. Overall, this study illustrates how affordable and accessible it is to enhance LLM safety using DPO while outlining its current limitations. All datasets and models are released to enable reproducibility and further research.

Anna Arias-Duart, Maria Eugenia Cardello, Atia Cortés

Artificial intelligence (AI) holds great promise for transforming healthcare. However, despite significant advances, the integration of AI solutions into real-world clinical practice remains limited. A major barrier is the quality and fairness of training data, which is often compromised by biased data collection practices. This paper draws on insights from the AI4HealthyAging project, part of Spain's national R&D initiative, where our task was to detect biases during clinical data collection. We identify several types of bias across multiple use cases, including historical, representation, and measurement biases. These biases manifest in variables such as sex, gender, age, habitat, socioeconomic status, equipment, and labeling. We conclude with practical recommendations for improving the fairness and robustness of clinical problem design and data collection. We hope that our findings and experience contribute to guiding future projects in the development of fairer AI systems in healthcare.

Anna Arias-Duart, Ferran Parés, Dario Garcia-Gasulla et al.

AI explainability improves the transparency of models, making them more trustworthy. Such goals are motivated by the emergence of deep learning models, which are obscure by nature; even in the domain of images, where deep learning has succeeded the most, explainability is still poorly assessed. In the field of image recognition many feature attribution methods have been proposed with the purpose of explaining a model's behavior using visual cues. However, no metrics have been established so far to assess and select these methods objectively. In this paper we propose a consistent evaluation score for feature attribution methods -- the Focus -- designed to quantify their coherency to the task. While most previous work adds out-of-distribution noise to samples, we introduce a methodology to add noise from within the distribution. This is done through mosaics of instances from different classes, and the explanations these generate. On those, we compute a visual pseudo-precision metric, Focus. First, we show the robustness of the approach through a set of randomization experiments. Then we use Focus to compare six popular explainability techniques across several CNN architectures and classification datasets. Our results find some methods to be consistently reliable (LRP, GradCAM), while others produce class-agnostic explanations (SmoothGrad, IG). Finally we introduce another application of Focus, using it for the identification and characterization of biases found in models. This empowers bias-management tools, in another small step towards trustworthy AI.

Ferran Parés, Anna Arias-Duart, Dario Garcia-Gasulla et al.

In the image classification task, the most common approach is to resize all images in a dataset to a unique shape, while reducing their precision to a size which facilitates experimentation at scale. This practice has benefits from a computational perspective, but it entails negative side-effects on performance due to loss of information and image deformation. In this work we introduce the MAMe dataset, an image classification dataset with remarkable high resolution and variable shape properties. The goal of MAMe is to provide a tool for studying the impact of such properties in image classification, while motivating research in the field. The MAMe dataset contains thousands of artworks from three different museums, and proposes a classification task consisting on differentiating between 29 mediums (i.e. materials and techniques) supervised by art experts. After reviewing the singularity of MAMe in the context of current image classification tasks, a thorough description of the task is provided, together with dataset statistics. Experiments are conducted to evaluate the impact of using high resolution images, variable shape inputs and both properties at the same time. Results illustrate the positive impact in performance when using high resolution images, while highlighting the lack of solutions to exploit variable shapes. An additional experiment exposes the distinctiveness between the MAMe dataset and the prototypical ImageNet dataset. Finally, the baselines are inspected using explainability methods and expert knowledge, to gain insights on the challenges that remain ahead.