Proceedings of Machine Learning Research

RIS: Region-to-Image Search using ViT-like Embeddings

Mon, 06 Apr 2026 00:00:00 +0000

We propose RIS (Region-to-Image Search), a two-stage framework for localized visual retrieval. RIS performs structural re-ranking directly within the latent embedding space of Vision Transformers, such as SigLIP2 and I-JEPA, bypassing traditional pixel-level verification. By matching a query Region of Interest (ROI) through a spatially-consistent region-growing algorithm, the framework ensures geometric coherence across latent representations. Preliminary qualitative results demonstrate that this embedding-based re-ranking improves Top-5 retrieval accuracy by at least 10% over standalone global methods, providing a robust and efficient mechanism for localized forensic search.

IA4FriLex: Enhancing The Legislative Consultation Process With AI

Mon, 06 Apr 2026 00:00:00 +0000

Legislative consultation procedures are a core component of participatory law-making but require public administrations to process large volumes of heterogeneous and predominantly unstructured submissions under strict procedural constraints. This task is particularly demanding in practice and often constitutes a bottleneck in legislative workflows. To overcome this challenge, we present IA4FriLex, an AI-assisted pipeline designed to support the processing and synthesis of consultation submissions through a structured and legally grounded workflow. Built exclusively on open-source software and Large Language Models (LLMs), the system automates well-defined stages of consultation handling while preserving human oversight and legal responsibility. IA4FriLex produces standardized, department-ready consultation reports aligned with established cantonal administrative practices. The system has been implemented and deployed in collaboration with the Cantonal administration of Fribourg and evaluated on four real cantonal legislative cases, covering both completed and ongoing consultations. Results show that IA4FriLex reliably generates high-quality first-pass syntheses and reduces consultation report preparation time by at least 80% compared to fully manual drafting. These findings demonstrate that carefully constrained, on-premise deployments of LLM-based systems can effectively support legislative consultation processes, offering a scalable and institutionally compatible approach to AI-assisted law-making.

Why Measuring AI Environmental Impact of Organisations is Non-Trivial?

Mon, 06 Apr 2026 00:00:00 +0000

This paper presents the results and conclusions from the EIEIAE project, which defined a methodology to measure the environmental impact of AI services within companies and organisations. Several lessons can be learned from this project, as it highlighted several challenges that emerged when trying to produce reliable estimates of such impacts: (a) the industrial lack of transparency of AI providers, (b) the absence of accurate and exhaustive modelling of ICT environmental impacts, (c) the inherent complexity of gathering the necessary data for organisations, and (d) the fact that organisations are more focused on cost savings than on reducing their environmental impact.

Benchmarking Time Series Foundation Models on their Accuracy and Energy Consumption

Mon, 06 Apr 2026 00:00:00 +0000

Our study presents a benchmark of ten time-series foundation models to quantify their accuracy–energy trade-off in zero-shot forecasting. Using an in-house and a public dataset (School and MeteoSwiss; univariate and multivariate variants), a fixed sliding-window protocol (context 512, horizon 64), and dual energy instrumentation (external PDU and CodeCarbon), we report sMAPE and NMAE accuracy metrics alongside runtime, energy ($Wh$), and Energy per Billion Parameters. Results show pronounced dataset dependence in accuracy, while efficiency is primarily architecture-driven: Chronos-Bolt achieves consistently low energy and latency, TimesFM attains the best MeteoSwiss accuracy at low energy cost, and Moirai-MoE exhibits substantially higher energy expenditure for comparable errors. This work informs decision-makers, developers, and end-users about the energy requirements of time-series foundation models and highlights the importance of considering energy alongside accuracy when evaluating models for adoption, while encouraging the systematic reporting of accuracy–energy trade-offs.

Predictive Modeling of Long-Term CPAP Non-Adherence in OSA patients from Post-Initiation Treatment Telemonitoring Data

Mon, 06 Apr 2026 00:00:00 +0000

Long-term adherence to continuous positive airway pressure (CPAP) therapy remains a major challenge in the management of obstructive sleep apnea, despite its well-established clinical benefits. The objective of this study is to predict long-term CPAP non-adherence one year after a baseline month that does not correspond to treatment initiation but can occur within a time window ranging from the fourth to the twenty-fourth month of therapy. We propose and compare a machine learning (ML) pipeline and a deep learning (DL) approach that leverage daily CPAP telemonitoring time-series and electronic health record (EHR) variables. Both pipelines are evaluated on a held-out test set: the ML model achieved a macro F1-score of 0.83, while the DL model achieved 0.81, indicating comparable and robust predictive performance. These results suggest that CPAP usage patterns observed during an intermediate treatment phase remain highly informative for identifying patients at risk of future non-adherence and could support targeted long-term telemonitoring strategies, serving as a data-driven second opinion to assist clinicians in decision-making.

Simple Probability Truncation Improves Soft Red List Watermarks

Mon, 06 Apr 2026 00:00:00 +0000

Watermarking, whereby LLM outputs are steered to encode an easily identifiable digital signature, has recently gained attention as a potential solution for detecting synthetically generated text. However, watermarking schemes require tradeoffs between detectability (i.e., how easily the watermark can be identified by an algorithm) and quality of the generated text (i.e., the stylistic and semantic disruption to the normal generation of the LLM). In this work, we propose a simple extension to the Soft Red List watermark, Softer Red List, which enables higher detectability while maintaining text quality on par with non-watermarked text. Specifically, Softer Red List improves the classical red/green token algorithm by adding a probability truncation filter before boosting the probabilities tokens in the green list. Despite its simplicity, Softer Red List matches or exceeds the performance of previously published LLM watermarking schemes, notably achieving a better detection rate at a low false positive rate (FPR) than SynthID in the disinformation detection setting, all while maintaining comparable perplexity and better reasoning capacities.

A Lightweight Deep Residual Network for Rehabilitation Activity Recognition in Heterogeneous Pediatric Populations

Mon, 06 Apr 2026 00:00:00 +0000

Human Activity Recognition based on wearable Inertial Measurement Units (IMUs) has emerged as a promising technology for the automated quantification of rehabilitation dosage and rehabilitation activity identification. However, existing solutions rely on multi-sensor configurations limiting clinical usability or fail to generalize to populations with heterogeneous motor functions. This study developed a lightweight residual network with self-attention mechanisms for classifying different phases of the rehabilitation activities (Rest, Balance, Walk) using data from a single IMU placed at the lower back, and collected from a pediatric cohort including 10 neurotypical children (mean age: years, 5 females) and 8 patients with neuromotor disorders as a consequence of cerebral palsy or acquired brain injury (mean age: years, 4 females). A preliminary ablation study across different IMU channel combinations revealed that combining accelerometer, gyroscope, and magnetometer signals allowed the model to achieve the best performance, with the magnetometer providing a key contribution for better discriminating between low-dynamic activities (Rest and Balance). Based on the optimal channel configuration identified in the ablation study, a Leave-One-Subject-Out cross-validation framework proved the model generalization abilities across heterogeneous motor functional domains, achieving an average macro F1-score of 0.81. These results confirm that the proposed framework provides an ecological and reliable tool for the objective recognition and quantification of rehabilitation activity in a clinical context.

Detecting Whisper Hallucinations with Local Confidence Contrasts

Mon, 06 Apr 2026 00:00:00 +0000

Automatic speech recognition has advanced significantly with models like Whisper, yet confident hallucinations remain a critical challenge. In this work, we propose a lightweight and interpretable error detection framework that augments acoustic confidence with explicit contextual features. We introduce the Local Confidence Drop, a novel metric designed to capture sudden stability dips between neighboring tokens. Evaluated on the FLEURS dataset, our fandom forest classifier achieves 0.64 AP, consistently outperforming the baseline (p < 0.001). Crucially, we demonstrate that hallucinations manifest as local contextual discontinuities, providing a transparent alternative to opaque neural post-processors.