Luca Benfenati

Luca Benfenati, Ali Azimi, Matteo Risso et al.

Embodied Large Language Models (LLMs) are increasingly used as reasoning modules in robotic control pipelines to improve human-robot interaction, but their memory and generation latency make real-time deployment difficult. Pruning can reduce these costs, but for controllers that undergo multiple pre- and post-training phases, the crucial question is not only how much to prune, but when pruning should occur. In this work, we propose Before Parc Fermé (BPF), a pruning strategy performed during RL that compresses embodied LLM controllers while they are still being optimized for closed-loop behavior. This allows pruning decisions to account for the task-specific supervision and closed-loop feedback that shape the final controller. We propose two variants: BPF-RL, which performs iterative pruning during RL by removing part of the model at predefined training intervals, and BPF-SFT/RL, which first prunes part of the model structure during SFT and then further compresses it during RL using the same iterative strategy as BPF-RL until the target pruning ratio is reached. We evaluate BPF on RobotxR1, an LLM-based autonomous-driving control pipeline, using an established LLM pruning framework (LLM-Pruner), and compare it against post-training pruning, post-training pruning with RL recovery, SFT-stage pruning, and smaller dense models from the same family. Our results show that BPF provides the best task-performance vs. memory and throughput trade-off among the considered pruning strategies. When compressing the larger RobotxR1 models, BPF-SFT/RL achieves a $1.69\times$ better size-end-to-end performance trade-off than directly selecting a smaller dense model from the same family, measured as removed parameters per lost percentage point of control adaptability. On the Jetson AGX Orin mounted on the target robotic platform, the compact models improve decode throughput by up to $27\%$.

Luca Benfenati, Matteo Risso, Andrea Vannozzi et al.

Key--value (KV) caching enables fast autoregressive decoding but at long contexts becomes a dominant bottleneck in High Bandwidth Memory (HBM) capacity and bandwidth. A common mitigation is to compress cached keys and values by projecting per-head matrixes to a lower rank, storing only the projections in the HBM. However, existing post-training approaches typically fit these projections using SVD-style proxy objectives, which may poorly reflect end-to-end reconstruction after softmax, value mixing, and subsequent decoder-layer transformations. For these reasons, we introduce StiefAttention, a post-training KV-cache compression method that learns \emph{orthonormal} projection bases by directly minimizing \emph{decoder-layer output reconstruction error}. StiefAttention additionally precomputes, for each layer, an error-rank profile over candidate ranks, enabling flexible layer-wise rank allocation under a user-specified error budget. Noteworthy, on Llama3-8B under the same conditions, StiefAttention outperforms EigenAttention by $11.9$ points on C4 perplexity and $5.4\%$ on 0-shot MMLU accuracy at iso-compression, yielding lower relative error and higher cosine similarity with respect to the original decoder-layer outputs.

Lorenz K. Müller, Philippe Bich, Jiawei Zhuang et al.

Post-training quantization has emerged as the most widely used strategy for deploying large language models at low precision. Still, current methods show perplexity degradation at bit-widths less than or equal to 4, partly because representing outliers causes precision issues in parameters that share the same scales as these outliers. This problem is especially pronounced for calibration-free, uniform quantization methods. We introduce SINQ to augment existing post-training quantizers with an additional second-axis scale factor and a fast Sinkhorn-Knopp-style algorithm that finds scales to normalize per-row and per-column variances, thereby minimizing a novel per-matrix proxy target for quantization: the matrix imbalance. Our method has no interactions between layers and can be trivially applied to new architectures to quantize any linear layers. We evaluate our method on the Qwen3 model family and DeepSeek-V2.5. SINQ improves WikiText2 and C4 perplexity significantly against uncalibrated uniform quantization baselines and can be further enhanced by combining it with calibration and non-uniform quantization levels. Code to reproduce the results of this work and to easily quantize models using SINQ is available at https://github.com/huawei-csl/SINQ.

Luca Benfenati, Daniele Jahier Pagliari, Luca Zanatta et al.

Structural Health Monitoring (SHM) is a critical task for ensuring the safety and reliability of civil infrastructures, typically realized on bridges and viaducts by means of vibration monitoring. In this paper, we propose for the first time the use of Transformer neural networks, with a Masked Auto-Encoder architecture, as Foundation Models for SHM. We demonstrate the ability of these models to learn generalizable representations from multiple large datasets through self-supervised pre-training, which, coupled with task-specific fine-tuning, allows them to outperform state-of-the-art traditional methods on diverse tasks, including Anomaly Detection (AD) and Traffic Load Estimation (TLE). We then extensively explore model size versus accuracy trade-offs and experiment with Knowledge Distillation (KD) to improve the performance of smaller Transformers, enabling their embedding directly into the SHM edge nodes. We showcase the effectiveness of our foundation models using data from three operational viaducts. For AD, we achieve a near-perfect 99.9% accuracy with a monitoring time span of just 15 windows. In contrast, a state-of-the-art method based on Principal Component Analysis (PCA) obtains its first good result (95.03% accuracy), only considering 120 windows. On two different TLE tasks, our models obtain state-of-the-art performance on multiple evaluation metrics (R$^2$ score, MAE% and MSE%). On the first benchmark, we achieve an R$^2$ score of 0.97 and 0.90 for light and heavy vehicle traffic, respectively, while the best previous approach (a Random Forest) stops at 0.91 and 0.84. On the second one, we achieve an R$^2$ score of 0.54 versus the 0.51 of the best competitor method, a Long-Short Term Memory network.

Luca Benfenati, Sofia Belloni, Alessio Burrello et al.

Heart rate (HR) estimation from photoplethysmography (PPG) signals is a key feature of modern wearable devices for health and wellness monitoring. While deep learning models show promise, their performance relies on the availability of large datasets. We present EnhancePPG, a method that enhances state-of-the-art models by integrating self-supervised learning with data augmentation (DA). Our approach combines self-supervised pre-training with DA, allowing the model to learn more generalizable features, without needing more labelled data. Inspired by a U-Net-like autoencoder architecture, we utilize unsupervised PPG signal reconstruction, taking advantage of large amounts of unlabeled data during the pre-training phase combined with data augmentation, to improve state-of-the-art models' performance. Thanks to our approach and minimal modification to the state-of-the-art model, we improve the best HR estimation by 12.2%, lowering from 4.03 Beats-Per-Minute (BPM) to 3.54 BPM the error on PPG-DaLiA. Importantly, our EnhancePPG approach focuses exclusively on the training of the selected deep learning model, without significantly increasing its inference latency

Luca Benfenati, Thorir Mar Ingolfsson, Andrea Cossettini et al.

This study presents a novel approach for EEG-based seizure detection leveraging a BERT-based model. The model, BENDR, undergoes a two-phase training process. Initially, it is pre-trained on the extensive Temple University Hospital EEG Corpus (TUEG), a 1.5 TB dataset comprising over 10,000 subjects, to extract common EEG data patterns. Subsequently, the model is fine-tuned on the CHB-MIT Scalp EEG Database, consisting of 664 EEG recordings from 24 pediatric patients, of which 198 contain seizure events. Key contributions include optimizing fine-tuning on the CHB-MIT dataset, where the impact of model architecture, pre-processing, and post-processing techniques are thoroughly examined to enhance sensitivity and reduce false positives per hour (FP/h). We also explored custom training strategies to ascertain the most effective setup. The model undergoes a novel second pre-training phase before subject-specific fine-tuning, enhancing its generalization capabilities. The optimized model demonstrates substantial performance enhancements, achieving as low as 0.23 FP/h, 2.5$\times$ lower than the baseline model, with a lower but still acceptable sensitivity rate, showcasing the effectiveness of applying a BERT-based approach on EEG-based seizure detection.