arXiv subjects

David Montero

Publications and source records attributed to David Montero.

At least 19 recordsLinked to original sources

arXiv2026

Quantization-Aware Healing: A Practical Recipe for Recovering Compressed, 4-Bit LLMs

Serving large language models cheaply increasingly means shipping models that are both structurally compressed to a fraction of their parameters and quantized to 4 bits. Together these steps degrade reasoning, mathematics, coding, and long-context behavior enough to require a recovery, or healing, stage before deployment. The default recipe, quantization-aware training (QAT), re-fits the compressed, quantized model to hard labels; in our pipeline it converged slowly and collapsed past its peak. We adopted Quantization-Aware Healing (QAH) instead. Because a structurally compressed model is never independently trained at full precision, its bfloat16 checkpoint is a distillation-recovered approximation of the original; QAH distills the 4-bit student directly from the original, uncompressed model. On a GPT-OSS 120B to 60B to MXFP4 pipeline, the QAH student matches or beats its bfloat16 source on 7 of 9 benchmarks at roughly 4 times less weight memory and half the teacher's parameter count, and is released open-weight as Hypernova-60B. Against a matched QAT baseline it reaches a comparable peak about 7 times faster and stays stable under continued training, without hand-tuned early stopping. We also report deployment lessons, including a large, reproducible quality gap between distributed-training backends. Our aim is a recipe deployable without a multi-week hyper-parameter search.

cs.CL↗

arXiv2026

Efficient Knowledge Distillation for LLMs: Offline Top-K Logits and a Fused Chunked KL Loss

Small language models are often the only option for deployment under tight latency, cost, and on-premises constraints, but they are rarely trained from scratch: a compressed model is usually recovered through knowledge distillation (KD). This recovery step largely decides the final quality, yet it is expensive. We present a practitioner's study of how to make distillation training efficient, organised around two systems contributions. First, we show that offline KD (caching the teacher's top-$K$ logits once and training the student against the cache) matches online distillation at near-identical training loss while removing the teacher from memory, running about 29\% faster per iteration, and reaching up to 41\% higher throughput on a single H200 GPU. Second, we introduce a \emph{fused, chunked KL loss} that never materialises the full vocabulary-sized logit tensor, making peak memory linear in the sequence length. This removes the memory spike that otherwise caps context length and lets us train at four times the context (32{,}768 tokens) on a single GPU. A separate output-head-only toy benchmark isolates the loss kernel and confirms its memory and iteration-rate scaling from 4K to 256K tokens. Together these make large-scale healing and hundreds of ablations affordable. We also report supporting ablations on loss design and sequence packing. We release our chunked-loss implementation: https://github.com/CompactifAI/Full-Chunked-KL-Loss.

cs.CL↗

arXiv2026

Towards an open registry of Earth observation instruments

Earth observation (EO) is essential to understanding the Earth system, enabling the transformation of planetary properties into measurable variables that can be analysed, compared, and modelled. In recent decades, EO capabilities have grown rapidly, accompanied by an even faster expansion in the number and variety of available EO instruments. Today, EO includes instruments deployed on satellites, airborne platforms, and terrestrial or in-situ systems. However, despite this proliferation of instruments, users often lack a single, reliable source describing their existence and key characteristics. Although existing data catalogues have substantially improved dataset discovery, they primarily describe data products rather than providing persistent, curated metadata about the instruments that produced them. Here we present Awesome Earth Observation Instruments, an open, standardized, and community-oriented registry providing machine-readable metadata for EO instruments. The catalogue is hosted on GitHub and allows contributors to submit instrument metadata following a common schema. The schema combines a lightweight core with modular extensions covering spectral, geometric, and data access-related metadata, enabling both standardization and flexibility across diverse EO systems. All submissions undergo automated schema validation and human review. Because the schema is open, versioned, and extensible, the catalogue can continuously evolve as new instruments and metadata requirements emerge. This facilitates the discovery, interpretation, and analysis of EO data in light of instrument characteristics. To support programmatic access and interoperability, we further envisage an API for integration within common EO analysis environments. The catalogue is openly available at https://github.com/awesome-spectral-indices/awesome-earth-observation-instruments.

cs.DB↗

arXiv2026

LLM Compression by Block Removal with Constrained Binary Optimization

In this paper, we formulate the compression of large language models (LLMs) by optimally deleting transformer blocks (``block removal'') as a constrained binary optimization (CBO) problem that can be mapped to a physical system (Ising glass), whose energies are a strong proxy for downstream model performance. This formulation enables an efficient ranking of a large number of candidate block-removal configurations yielding many high-quality, non-trivial solutions beyond those only removing consecutive regions. Our method performs strongly in the deep compression regime, such as for 50% compression of Llama-3.3-70B-Instruct, where we achieve an almost 23 percentage point increase on the MMLU benchmark compared to other state-of-the-art (SOTA) block-removal methods. For lighter compression, it performs on par with those methods across several benchmarks for Llama-3.1-8B-Instruct, Qwen3-14B (both before and after retraining), as well as Llama-3.3-70B-Instruct. The approach is computationally efficient and requires only forward and backward passes on a calibration dataset for a few active parameters. Additionally, we demonstrate that using good heuristic solvers for the CBO problem provides solutions that perform well on downstream tasks in negligible runtime when it is unfeasible to solve the problem exactly. The method can be readily applied to any architecture. We illustrate this generality on the recent NVIDIA-Nemotron-3-Nano-30B-A3B-FP8 model, which exhibits a highly inhomogeneous and challenging block structure, and where we outperform SOTA for AIME25 and GPQA when removing either 2 attention layers or 3 mixture-of-experts layers.

cs.LG↗

arXiv2025

Scaling Laws for Energy Efficiency of Local LLMs

Deploying local large language models and vision-language models on edge devices requires balancing accuracy with constrained computational and energy budgets. Although graphics processors dominate modern artificial-intelligence deployment, most consumer hardware--including laptops, desktops, industrial controllers, and embedded systems--relies on central processing units. Despite this, the computational laws governing central-processing-unit-only inference for local language and vision-language workloads remain largely unexplored. We systematically benchmark large language and vision-language models on two representative central-processing-unit tiers widely used for local inference: a MacBook Pro M2, reflecting mainstream laptop-class deployment, and a Raspberry Pi 5, representing constrained, low-power embedded settings. Using a unified methodology based on continuous sampling of processor and memory usage together with area-under-curve integration, we characterize how computational load scales with input text length for language models and with image resolution for vision-language models. We uncover two empirical scaling laws: (1) computational cost for language-model inference scales approximately linearly with token length; and (2) vision-language models exhibit a preprocessing-driven "resolution knee", where compute remains constant above an internal resolution clamp and decreases sharply below it. Beyond these laws, we show that quantum-inspired compression reduces processor and memory usage by up to 71.9% and energy consumption by up to 62%, while preserving or improving semantic accuracy. These results provide a systematic quantification of multimodal central-processing-unit-only scaling for local language and vision-language workloads, and they identify model compression and input-resolution preprocessing as effective, low-cost levers for sustainable edge inference.

cs.AI↗

arXiv2025

Refusal Steering: Fine-grained Control over LLM Refusal Behaviour for Sensitive Topics

We introduce Refusal Steering, an inference-time method to exercise fine-grained control over Large Language Models refusal behaviour on politically sensitive topics without retraining. We replace fragile pattern-based refusal detection with an LLM-as-a-judge that assigns refusal confidence scores and we propose a ridge-regularized variant to compute steering vectors that better isolate the refusal--compliance direction. On Qwen3-Next-80B-A3B-Thinking, our method removes the refusal behaviour of the model around politically sensitive topics while maintaining safety on JailbreakBench and near-baseline performance on general benchmarks. The approach generalizes across 4B and 80B models and can also induce targeted refusals when desired. We analize the steering vectors and show that refusal signals concentrate in deeper layers of the transformer and are distributed across many dimensions. Together, these results demonstrate that activation steering can remove political refusal behaviour while retaining safety alignment for harmful content, offering a practical path to controllable, transparent moderation at inference time.

cs.CL↗

arXiv2025

Transformers vs. Recurrent Models for Estimating Forest Gross Primary Production

Monitoring the spatiotemporal dynamics of forest CO$_2$ uptake (Gross Primary Production, GPP), remains a central challenge in terrestrial ecosystem research. While Eddy Covariance (EC) towers provide high-frequency estimates, their limited spatial coverage constrains large-scale assessments. Remote sensing offers a scalable alternative, yet most approaches rely on single-sensor spectral indices and statistical models that are often unable to capture the complex temporal dynamics of GPP. Recent advances in deep learning (DL) and data fusion offer new opportunities to better represent the temporal dynamics of vegetation processes, but comparative evaluations of state-of-the-art DL models for multimodal GPP prediction remain scarce. Here, we explore the performance of two representative models for predicting GPP: 1) GPT-2, a transformer architecture, and 2) Long Short-Term Memory (LSTM), a recurrent neural network, using multivariate inputs. Overall, both achieve similar accuracy. But, while LSTM performs better overall, GPT-2 excels during extreme events. Analysis of temporal context length further reveals that LSTM attains similar accuracy using substantially shorter input windows than GPT-2, highlighting an accuracy-efficiency trade-off between the two architectures. Feature importance analysis reveals radiation as the dominant predictor, followed by Sentinel-2, MODIS land surface temperature, and Sentinel-1 contributions. Our results demonstrate how model architecture, context length, and multimodal inputs jointly determine performance in GPP prediction, guiding future developments of DL frameworks for monitoring terrestrial carbon dynamics.

cs.LG↗

arXiv2025

Context-Aware Multimodal Representation Learning for Spatio-Temporally Explicit Environmental Modelling

Earth observation (EO) foundation models have emerged as an effective approach to derive latent representations of the Earth system from various remote sensing sensors. These models produce embeddings that can be used as analysis-ready datasets, enabling the modelling of ecosystem dynamics without extensive sensor-specific preprocessing. However, existing models typically operate at fixed spatial or temporal scales, limiting their use for ecological analyses that require both fine spatial detail and high temporal fidelity. To overcome these limitations, we propose a representation learning framework that integrates different EO modalities into a unified feature space at high spatio-temporal resolution. We introduce the framework using Sentinel-1 and Sentinel-2 data as representative modalities. Our approach produces a latent space at native 10 m resolution and the temporal frequency of cloud-free Sentinel-2 acquisitions. Each sensor is first modeled independently to capture its sensor-specific characteristics. Their representations are then combined into a shared model. This two-stage design enables modality-specific optimisation and easy extension to new sensors, retaining pretrained encoders while retraining only fusion layers. This enables the model to capture complementary remote sensing data and to preserve coherence across space and time. Qualitative analyses reveal that the learned embeddings exhibit high spatial and semantic consistency across heterogeneous landscapes. Quantitative evaluation in modelling Gross Primary Production reveals that they encode ecologically meaningful patterns and retain sufficient temporal fidelity to support fine-scale analyses. Overall, the proposed framework provides a flexible, analysis-ready representation learning approach for environmental applications requiring diverse spatial and temporal resolutions.

cs.LG↗

arXiv2025

AI for a Planet Under Pressure

Artificial intelligence (AI) is already driving scientific breakthroughs in a variety of research fields, ranging from the life sciences to mathematics. This raises a critical question: can AI be applied both responsibly and effectively to address complex and interconnected sustainability challenges? This report is the result of a collaboration between the Stockholm resilience Centre (Stockholm University), the Potsdam Institute for Climate Impact Research (PIK), and Google DeepMind. Our work explores the potential and limitations of using AI as a research method to help tackle eight broad sustainability challenges. The results build on iterated expert dialogues and assessments, a systematic AI-supported literature overview including over 8,500 academic publications, and expert deep-dives into eight specific issue areas. The report also includes recommendations to sustainability scientists, research funders, the private sector, and philanthropies.

cs.CY↗

arXiv2024

Earth System Data Cubes: Avenues for advancing Earth system research

Recent advancements in Earth system science have been marked by the exponential increase in the availability of diverse, multivariate datasets characterised by moderate to high spatio-temporal resolutions. Earth System Data Cubes (ESDCs) have emerged as one suitable solution for transforming this flood of data into a simple yet robust data structure. ESDCs achieve this by organising data into an analysis-ready format aligned with a spatio-temporal grid, facilitating user-friendly analysis and diminishing the need for extensive technical data processing knowledge. Despite these significant benefits, the completion of the entire ESDC life cycle remains a challenging task. Obstacles are not only of a technical nature but also relate to domain-specific problems in Earth system research. There exist barriers to realising the full potential of data collections in light of novel cloud-based technologies, particularly in curating data tailored for specific application domains. These include transforming data to conform to a spatio-temporal grid with minimum distortions and managing complexities such as spatio-temporal autocorrelation issues. Addressing these challenges is pivotal for the effective application of Artificial Intelligence (AI) approaches. Furthermore, adhering to open science principles for data dissemination, reproducibility, visualisation, and reuse is crucial for fostering sustainable research. Overcoming these challenges offers a substantial opportunity to advance data-driven Earth system research, unlocking the full potential of an integrated, multidimensional view of Earth system processes. This is particularly true when such research is coupled with innovative research paradigms and technological progress.

cs.CV↗

arXiv2024

DeepExtremeCubes: Integrating Earth system spatio-temporal data for impact assessment of climate extremes

With climate extremes' rising frequency and intensity, robust analytical tools are crucial to predict their impacts on terrestrial ecosystems. Machine learning techniques show promise but require well-structured, high-quality, and curated analysis-ready datasets. Earth observation datasets comprehensively monitor ecosystem dynamics and responses to climatic extremes, yet the data complexity can challenge the effectiveness of machine learning models. Despite recent progress in deep learning to ecosystem monitoring, there is a need for datasets specifically designed to analyse compound heatwave and drought extreme impact. Here, we introduce the DeepExtremeCubes database, tailored to map around these extremes, focusing on persistent natural vegetation. It comprises over 40,000 spatially sampled small data cubes (i.e. minicubes) globally, with a spatial coverage of 2.5 by 2.5 km. Each minicube includes (i) Sentinel-2 L2A images, (ii) ERA5-Land variables and generated extreme event cube covering 2016 to 2022, and (iii) ancillary land cover and topography maps. The paper aims to (1) streamline data accessibility, structuring, pre-processing, and enhance scientific reproducibility, and (2) facilitate biosphere dynamics forecasting in response to compound extremes.

cs.LG↗

arXiv2024

Facilitating Advanced Sentinel-2 Analysis Through a Simplified Computation of Nadir BRDF Adjusted Reflectance

The Sentinel-2 (S2) mission from the European Space Agency's Copernicus program provides essential data for Earth surface analysis. Its Level-2A products deliver high-to-medium resolution (10-60 m) surface reflectance (SR) data through the MultiSpectral Instrument (MSI). To enhance the accuracy and comparability of SR data, adjustments simulating a nadir viewing perspective are essential. These corrections address the anisotropic nature of SR and the variability in sun and observation angles, ensuring consistent image comparisons over time and under different conditions. The $c$-factor method, a simple yet effective algorithm, adjusts observed S2 SR by using the MODIS BRDF model to achieve Nadir BRDF Adjusted Reflectance (NBAR). Despite the straightforward application of the $c$-factor to individual images, a cohesive Python framework for its application across multiple S2 images and Earth System Data Cubes (ESDCs) from cloud-stored data has been lacking. Here we introduce sen2nbar, a Python package crafted to convert S2 SR data to NBAR, supporting both individual images and ESDCs derived from cloud-stored data. This package simplifies the conversion of S2 SR data to NBAR via a single function, organized into modules for efficient process management. By facilitating NBAR conversion for both SAFE files and ESDCs from SpatioTemporal Asset Catalogs (STAC), sen2nbar is developed as a flexible tool that can handle diverse data format requirements. We anticipate that sen2nbar will considerably contribute to the standardization and harmonization of S2 data, offering a robust solution for a diverse range of users across various applications. sen2nbar is an open-source tool available at https://github.com/ESDS-Leipzig/sen2nbar.

cs.CV↗

arXiv2024

Recurrent Neural Networks for Modelling Gross Primary Production

Accurate quantification of Gross Primary Production (GPP) is crucial for understanding terrestrial carbon dynamics. It represents the largest atmosphere-to-land CO$_2$ flux, especially significant for forests. Eddy Covariance (EC) measurements are widely used for ecosystem-scale GPP quantification but are globally sparse. In areas lacking local EC measurements, remote sensing (RS) data are typically utilised to estimate GPP after statistically relating them to in-situ data. Deep learning offers novel perspectives, and the potential of recurrent neural network architectures for estimating daily GPP remains underexplored. This study presents a comparative analysis of three architectures: Recurrent Neural Networks (RNNs), Gated Recurrent Units (GRUs), and Long-Short Term Memory (LSTMs). Our findings reveal comparable performance across all models for full-year and growing season predictions. Notably, LSTMs outperform in predicting climate-induced GPP extremes. Furthermore, our analysis highlights the importance of incorporating radiation and RS inputs (optical, temperature, and radar) for accurate GPP predictions, particularly during climate extremes.

cs.LG↗

arXiv2024

On-Demand Earth System Data Cubes

Advancements in Earth system science have seen a surge in diverse datasets. Earth System Data Cubes (ESDCs) have been introduced to efficiently handle this influx of high-dimensional data. ESDCs offer a structured, intuitive framework for data analysis, organising information within spatio-temporal grids. The structured nature of ESDCs unlocks significant opportunities for Artificial Intelligence (AI) applications. By providing well-organised data, ESDCs are ideally suited for a wide range of sophisticated AI-driven tasks. An automated framework for creating AI-focused ESDCs with minimal user input could significantly accelerate the generation of task-specific training data. Here we introduce cubo, an open-source Python tool designed for easy generation of AI-focused ESDCs. Utilising collections in SpatioTemporal Asset Catalogs (STAC) that are stored as Cloud Optimised GeoTIFFs (COGs), cubo efficiently creates ESDCs, requiring only central coordinates, spatial resolution, edge size, and time range.

cs.DB↗

arXiv2024

CompactifAI: Extreme Compression of Large Language Models using Quantum-Inspired Tensor Networks

Large Language Models (LLMs) such as ChatGPT and LlaMA are advancing rapidly in generative Artificial Intelligence (AI), but their immense size poses significant challenges, such as huge training and inference costs, substantial energy demands, and limitations for on-site deployment. Traditional compression methods such as pruning, distillation, and low-rank approximation focus on reducing the effective number of neurons in the network, while quantization focuses on reducing the numerical precision of individual weights to reduce the model size while keeping the number of neurons fixed. While these compression methods have been relatively successful in practice, there is no compelling reason to believe that truncating the number of neurons is an optimal strategy. In this context, this paper introduces CompactifAI, an innovative LLM compression approach using quantum-inspired Tensor Networks that focuses on the model's correlation space instead, allowing for a more controlled, refined and interpretable model compression. Our method is versatile and can be implemented with - or on top of - other compression techniques. As a benchmark, we demonstrate that a combination of CompactifAI with quantization allows to reduce a 93% the memory size of LlaMA 7B, reducing also 70% the number of parameters, accelerating 50% the training and 25% the inference times of the model, and just with a small accuracy drop of 2% - 3%, going much beyond of what is achievable today by other compression techniques. Our methods also allow to perform a refined layer sensitivity profiling, showing that deeper layers tend to be more suitable for tensor network compression, which is compatible with recent observations on the ineffectiveness of those layers for LLM performance. Our results imply that standard LLMs are, in fact, heavily overparametrized, and do not need to be large at all.

cs.CL↗

arXiv2021

MFR 2021: Masked Face Recognition Competition

This paper presents a summary of the Masked Face Recognition Competitions (MFR) held within the 2021 International Joint Conference on Biometrics (IJCB 2021). The competition attracted a total of 10 participating teams with valid submissions. The affiliations of these teams are diverse and associated with academia and industry in nine different countries. These teams successfully submitted 18 valid solutions. The competition is designed to motivate solutions aiming at enhancing the face recognition accuracy of masked faces. Moreover, the competition considered the deployability of the proposed solutions by taking the compactness of the face recognition models into account. A private dataset representing a collaborative, multi-session, real masked, capture scenario is used to evaluate the submitted solutions. In comparison to one of the top-performing academic face recognition solutions, 10 out of the 18 submitted solutions did score higher masked face verification accuracy.

cs.CV↗

arXiv2021

Learning to Automatically Catch Potholes in Worldwide Road Scene Images

Among several road hazards that are present in any paved way in the world, potholes are one of the most annoying and also involving higher maintenance costs. There exists an increasing interest on the automated detection of these hazards enabled by technological and research progress. Our research work tackled the challenge of pothole detection from images of real world road scenes. The main novelty resides on the application of the latest progress in AI to learn the visual appearance of potholes. We built a large dataset of images with pothole annotations. They contained road scenes from different cities in the world, taken with different cameras, vehicles and viewpoints under varied environmental conditions. Then, we fine-tuned four different object detection models based on Faster R-CNN and SSD deep neural networks. We achieved high average precision and the pothole detector was tested on the Nvidia DrivePX2 platform with GPGPU capability, which can be embedded on vehicles. Moreover, it was deployed on a real vehicle to notify the detected potholes to a given IoT platform as part of AUTOPILOT H2020 project.

cs.CV↗

arXiv2021

Boosting Masked Face Recognition with Multi-Task ArcFace

In this paper, we address the problem of face recognition with masks. Given the global health crisis caused by COVID-19, mouth and nose-covering masks have become an essential everyday-clothing-accessory. This sanitary measure has put the state-of-the-art face recognition models on the ropes since they have not been designed to work with masked faces. In addition, the need has arisen for applications capable of detecting whether the subjects are wearing masks to control the spread of the virus. To overcome these problems a full training pipeline is presented based on the ArcFace work, with several modifications for the backbone and the loss function. From the original face-recognition dataset, a masked version is generated using data augmentation, and both datasets are combined during the training process. The selected network, based on ResNet-50, is modified to also output the probability of mask usage without adding any computational cost. Furthermore, the ArcFace loss is combined with the mask-usage classification loss, resulting in a new function named Multi-Task ArcFace (MTArcFace). Experimental results show that the proposed approach highly boosts the original model accuracy when dealing with masked faces, while preserving almost the same accuracy on the original non-masked datasets. Furthermore, it achieves an average accuracy of 99.78% in mask-usage classification.

cs.CV↗