João Monteiro

Pedro C. Neto, Diana Montezuma, Sara P. Oliveira et al.

Considering the profound transformation affecting pathology practice, we aimed to develop a scalable artificial intelligence (AI) system to diagnose colorectal cancer from whole-slide images (WSI). For this, we propose a deep learning (DL) system that learns from weak labels, a sampling strategy that reduces the number of training samples by a factor of six without compromising performance, an approach to leverage a small subset of fully annotated samples, and a prototype with explainable predictions, active learning features and parallelisation. Noting some problems in the literature, this study is conducted with one of the largest WSI colorectal samples dataset with approximately 10,500 WSIs. Of these samples, 900 are testing samples. Furthermore, the robustness of the proposed method is assessed with two additional external datasets (TCGA and PAIP) and a dataset of samples collected directly from the proposed prototype. Our proposed method predicts, for the patch-based tiles, a class based on the severity of the dysplasia and uses that information to classify the whole slide. It is trained with an interpretable mixed-supervision scheme to leverage the domain knowledge introduced by pathologists through spatial annotations. The mixed-supervision scheme allowed for an intelligent sampling strategy effectively evaluated in several different scenarios without compromising the performance. On the internal dataset, the method shows an accuracy of 93.44% and a sensitivity between positive (low-grade and high-grade dysplasia) and non-neoplastic samples of 0.996. On the external test samples varied with TCGA being the most challenging dataset with an overall accuracy of 84.91% and a sensitivity of 0.996.

Szilvia Ujváry, Louis Béthune, Pierre Ablin et al.

Language models have consistently grown to compress more world knowledge into their parameters, but the knowledge that can be pretrained into them is upper-bounded by their parameter size. Especially the capacity of Small Language Models (SLMs) is limited, leading to factually incorrect generations. This problem is often mitigated by giving the SLM access to an outside source: the ability to query a larger model, documents, or a database. Under this setting, we study the fundamental question of \emph{which tokens an SLM can and should learn} during pretraining, versus \emph{which ones it should delegate} via a \texttt{<CALL>} token. We find that this is not simply a question of loss: although the loss is predictive of whether a predicted token mismatches the ground-truth, some tokens are \emph{acceptable} in that they are truthful alternative continuations of a pretraining document, and should not trigger a \texttt{<CALL>} even if their loss is high. We find that a spaCy grammar parser can help augment the loss signal to decide which tokens the SLM should learn to delegate to prevent factual errors and which are safe to learn and predict even under high losses. We propose LaCy, a novel pretraining method based on this token selection philosophy. Our experiments demonstrate that LaCy models successfully learn which tokens to predict and where to delegate for help. This results in higher FactScores when generating in a cascade with a bigger model and outperforms Rho or LLM-judge trained SLMs, while being simpler and cheaper.

João Monteiro, Michal Klein, Pierre Ablin et al.

Evaluating softmax attention over a fixed long context requires reading every cached key-value pair for each new query token. For a given context (a book, a manual, a legal corpus) the attention output is a deterministic function of the query. We propose Nectar, which fits a compact neural network to this function for queries drawn from a task-relevant distribution. Nectar fits two networks per layer and KV-head: a target network that predicts the attention output and a score network that predicts the log-normalizer. The pair plugs into the standard masked self-attention at inference time, replacing the $O(n)$ attention over the cache with a forward pass whose cost does not depend on $n$. Each module carries on the order of $|θ|$ parameters per layer and KV-head, typically much smaller than the $2nd$ KV-cache footprint at the same granularity. We report experiments on models from 1.7B to 8B parameters across five long-context datasets. The approximation error tracks the next-token accuracy gap to full attention, and allocating capacity non-uniformly across layers reduces that gap in our ablation. Beyond this analysis of metrics, we check that the text generations (following a question prompt) of a model equipped with a Nectar module match in semantic content those obtained by giving the same model access to the full cache.

João Monteiro, Étienne Marcotte, Pierre-André Noël et al.

In-context learning (ICL) approaches typically leverage prompting to condition decoder-only language model generation on reference information. Just-in-time processing of a context is inefficient due to the quadratic cost of self-attention operations, and caching is desirable. However, caching transformer states can easily require almost as much space as the model parameters. When the right context isn't known in advance, caching ICL can be challenging. This work addresses these limitations by introducing models that, inspired by the encoder-decoder architecture, use cross-attention to condition generation on reference text without the prompt. More precisely, we leverage pre-trained decoder-only models and only train a small number of added layers. We use Question-Answering (QA) as a testbed to evaluate the ability of our models to perform conditional generation and observe that they outperform ICL, are comparable to fine-tuned prompted LLMs, and drastically reduce the space footprint relative to standard KV caching by two orders of magnitude.

Aarash Feizi, Sai Rajeswar, Adriana Romero-Soriano et al.

Understanding how effectively large vision language models (VLMs) compare visual inputs is crucial across numerous applications, yet this fundamental capability remains insufficiently assessed. While VLMs are increasingly deployed for tasks requiring comparative judgment, including automated evaluation, re-ranking, and retrieval-augmented generation, no systematic framework exists to measure their performance in these scenarios. We present PairBench, a simple framework that evaluates VLMs as customizable similarity tools using widely available image datasets. Our approach introduces four key metrics for reliable comparison: alignment with human annotations, consistency across pair ordering, distribution smoothness, and controllability through prompting. Our analysis reveals that no model consistently excels across all metrics, with each demonstrating distinct strengths and weaknesses. Most concerning is the widespread inability of VLMs to maintain symmetric similarity scores. Interestingly, we demonstrate that performance on our benchmark strongly correlates with popular benchmarks used for more complex tasks, while providing additional metrics into controllability, smoothness and ordering. This makes PairBench a unique and comprehensive framework to evaluate the performance of VLMs for automatic evaluation depending on the task.

Anastasiia Filippova, David Grangier, Marco Cuturi et al.

Serving transformer language models with high throughput requires caching Key-Values (KVs) to avoid redundant computation during autoregressive generation. The memory footprint of KV caching is significant and heavily impacts serving costs. This work proposes to lessen these memory requirements. While recent work has largely addressed KV cache reduction via compression and eviction along the temporal axis, we argue that the \emph{depth} dimension offers an orthogonal and robust avenue for optimization. Although prior research suggests that a full cache for every layer is redundant, implementing cross-layer cache sharing remains a practical challenge; existing methods typically suffer from reduced throughput or increased time-to-first-token. In this paper, we demonstrate that dropping a layer's cache offers efficient optimization without information loss. We propose a simple training approach: random cross-layer attention. During training, layers randomly choose to attend either to their own KV states or those of a preceding layer. This stochastic process adapts the model to be robust to various depth-wise cache sharing strategies, ensuring flexibility for unknown hardware constraints at deployment time. Our evaluations show that applying this scheme during pre-training or fine-tuning enables depth-wise cache sharing for various model families. Furthermore, for larger models in data-constrained settings, this approach is suggestive of a regularization-like effect, frequently preserving or improving performance while significantly reducing the cache's memory footprint.

Theo X. Olausson, João Monteiro, Michal Klein et al.

Maximum inner product search (MIPS) is a crucial subroutine in machine learning, requiring the identification of key vectors that align best with a given query. We propose amortized MIPS: a learning-based approach that trains neural networks to directly predict MIPS solutions, amortizing the computational cost of matching queries (drawn from a fixed distribution) to a fixed set of keys. Our key insight is that the MIPS value function, the maximal inner product between a query and keys, is also known as the support function of the set of keys. Support functions are convex, 1-homogeneous and their gradient w.r.t. the query is exactly the optimal key in the database. We approximate the support function using two complementary approaches: (1) we train an input-convex neural network (SupportNet) to model the support function directly; the optimal key can be recovered via (autodiff) gradient computation, and (2) we regress directly the optimal key from the query using a vector valued network (KeyNet), bypassing gradient computation entirely at inference time. To learn a SupportNet, we combine score regression with gradient matching losses, and propose homogenization wrappers that enforce the positive 1-homogeneity of a neural network, theoretically linking function values to gradients. To train a KeyNet, we introduce a score consistency loss derived from the Euler theorem for homogeneous functions. Our experiments show that learned SupportNet or KeyNet achieve high match rates and open up new directions to compress databases with a specific query distribution in mind.

Ana Beatriz Vieira, Maria Valente, Diana Montezuma et al.

Quality control of medical images is a critical component of digital pathology, ensuring that diagnostic images meet required standards. A pre-analytical task within this process is the verification of the number of specimen fragments, a process that ensures that the number of fragments on a slide matches the number documented in the macroscopic report. This step is important to ensure that the slides contain the appropriate diagnostic material from the grossing process, thereby guaranteeing the accuracy of subsequent microscopic examination and diagnosis. Traditionally, this assessment is performed manually, requiring significant time and effort while being subject to significant variability due to its subjective nature. To address these challenges, this study explores an automated approach to fragment counting using the YOLOv9 and Vision Transformer models. Our results demonstrate that the automated system achieves a level of performance comparable to expert assessments, offering a reliable and efficient alternative to manual counting. Additionally, we present findings on interobserver variability, showing that the automated approach achieves an accuracy of 86%, which falls within the range of variation observed among experts (82-88%), further supporting its potential for integration into routine pathology workflows.

Andrea Gurioli, Federico Pennino, João Monteiro et al.

Deploying language models often requires navigating accuracy vs. performance trade-offs to meet latency constraints while preserving utility. Traditional model distillation reduces size but incurs substantial costs through training separate models. We introduce ModularStarEncoder (MoSE), a 1-billion-parameter multi-exit encoder for code retrieval and classification that employs a novel Self-Distillation mechanism. This approach significantly enhances lower-layer representations, enabling flexible deployment of different model portions with favorable performance trade-offs. Our architecture improves text-to-code and code-to-code search by targeting specific encoder layers as exit heads, where higher layers guide earlier ones during training-improving intermediate representations at minimal additional cost. We further enhance MoSE with a repository-level contextual loss that maximizes training context window utilization. Additionally, we release a new dataset created through code translation that extends text-to-code benchmarks with cross-language code-to-code pairs. Evaluations demonstrate the effectiveness of Self-Distillation as a principled approach to trading inference cost for accuracy across various code understanding tasks.

Raymond Li, Loubna Ben Allal, Yangtian Zi et al.

The BigCode community, an open-scientific collaboration working on the responsible development of Large Language Models for Code (Code LLMs), introduces StarCoder and StarCoderBase: 15.5B parameter models with 8K context length, infilling capabilities and fast large-batch inference enabled by multi-query attention. StarCoderBase is trained on 1 trillion tokens sourced from The Stack, a large collection of permissively licensed GitHub repositories with inspection tools and an opt-out process. We fine-tuned StarCoderBase on 35B Python tokens, resulting in the creation of StarCoder. We perform the most comprehensive evaluation of Code LLMs to date and show that StarCoderBase outperforms every open Code LLM that supports multiple programming languages and matches or outperforms the OpenAI code-cushman-001 model. Furthermore, StarCoder outperforms every model that is fine-tuned on Python, can be prompted to achieve 40\% pass@1 on HumanEval, and still retains its performance on other programming languages. We take several important steps towards a safe open-access model release, including an improved PII redaction pipeline and a novel attribution tracing tool, and make the StarCoder models publicly available under a more commercially viable version of the Open Responsible AI Model license.

Isabela Albuquerque, João Monteiro, Mohammad Darvishi et al.

Supervised learning results typically rely on assumptions of i.i.d. data. Unfortunately, those assumptions are commonly violated in practice. In this work, we tackle such problem by focusing on domain generalization: a formalization where the data generating process at test time may yield samples from never-before-seen domains (distributions). Our work relies on the following lemma: by minimizing a notion of discrepancy between all pairs from a set of given domains, we also minimize the discrepancy between any pairs of mixtures of domains. Using this result, we derive a generalization bound for our setting. We then show that low risk over unseen domains can be achieved by representing the data in a space where (i) the training distributions are indistinguishable, and (ii) relevant information for the task at hand is preserved. Minimizing the terms in our bound yields an adversarial formulation which estimates and minimizes pairwise discrepancies. We validate our proposed strategy on standard domain generalization benchmarks, outperforming a number of recently introduced methods. Notably, we tackle a real-world application where the underlying data corresponds to multi-channel electroencephalography time series from different subjects, each considered as a distinct domain.

Isabela Albuquerque, João Monteiro, Olivier Rosanne et al.

Assessment of mental workload in real-world conditions is key to ensure the performance of workers executing tasks that demand sustained attention. Previous literature has employed electroencephalography (EEG) to this end despite having observed that EEG correlates of mental workload vary across subjects and physical strain, thus making it difficult to devise models capable of simultaneously presenting reliable performance across users. Domain adaptation consists of a set of strategies that aim at allowing for improving machine learning systems performance on unseen data at training time. Such methods, however, might rely on assumptions over the considered data distributions, which typically do not hold for applications of EEG data. Motivated by this observation, in this work we propose a strategy to estimate two types of discrepancies between multiple data distributions, namely marginal and conditional shifts, observed on data collected from different subjects. Besides shedding light on the assumptions that hold for a particular dataset, the estimates of statistical shifts obtained with the proposed approach can be used for investigating other aspects of a machine learning pipeline, such as quantitatively assessing the effectiveness of domain adaptation strategies. In particular, we consider EEG data collected from individuals performing mental tasks while running on a treadmill and pedaling on a stationary bike and explore the effects of different normalization strategies commonly used to mitigate cross-subject variability. We show the effects that different normalization schemes have on statistical shifts and their relationship with the accuracy of mental workload prediction as assessed on unseen participants at training time.

Isabela Albuquerque, João Monteiro, Thang Doan et al.

Recent literature has demonstrated promising results for training Generative Adversarial Networks by employing a set of discriminators, in contrast to the traditional game involving one generator against a single adversary. Such methods perform single-objective optimization on some simple consolidation of the losses, e.g. an arithmetic average. In this work, we revisit the multiple-discriminator setting by framing the simultaneous minimization of losses provided by different models as a multi-objective optimization problem. Specifically, we evaluate the performance of multiple gradient descent and the hypervolume maximization algorithm on a number of different datasets. Moreover, we argue that the previously proposed methods and hypervolume maximization can all be seen as variations of multiple gradient descent in which the update direction can be computed efficiently. Our results indicate that hypervolume maximization presents a better compromise between sample quality and computational cost than previous methods.

Isabela Albuquerque, João Monteiro, Tiago H. Falk

In this work, we introduce a two-step framework for generative modeling of temporal data. Specifically, the generative adversarial networks (GANs) setting is employed to generate synthetic scenes of moving objects. To do so, we propose a two-step training scheme within which: a generator of static frames is trained first. Afterwards, a recurrent model is trained with the goal of providing a sequence of inputs to the previously trained frames generator, thus yielding scenes which look natural. The adversarial setting is employed in both training steps. However, with the aim of avoiding known training instabilities in GANs, a multiple discriminator approach is used to train both models. Results in the studied video dataset indicate that, by employing such an approach, the recurrent part is able to learn how to coherently navigate the image manifold induced by the frames generator, thus yielding more natural-looking scenes.

João Monteiro, Isabela Albuquerque, Zahid Akhtar et al.

Modern applications of artificial neural networks have yielded remarkable performance gains in a wide range of tasks. However, recent studies have discovered that such modelling strategy is vulnerable to Adversarial Examples, i.e. examples with subtle perturbations often too small and imperceptible to humans, but that can easily fool neural networks. Defense techniques against adversarial examples have been proposed, but ensuring robust performance against varying or novel types of attacks remains an open problem. In this work, we focus on the detection setting, in which case attackers become identifiable while models remain vulnerable. Particularly, we employ the decision layer of independently trained models as features for posterior detection. The proposed framework does not require any prior knowledge of adversarial examples generation techniques, and can be directly employed along with unmodified off-the-shelf models. Experiments on the standard MNIST and CIFAR10 datasets deliver empirical evidence that such detection approach generalizes well across not only different adversarial examples generation methods but also quality degradation attacks. Non-linear binary classifiers trained on top of our proposed features can achieve a high detection rate (>90%) in a set of white-box attacks and maintain such performance when tested against unseen attacks.