Elia Cunegatti

Elia Cunegatti, Marcus Vukojevic, Erik Nielsen et al.

Post-training compression of Large Language Models (LLMs) removes entire architectural components, either deleting them or replacing them with fitted modules. Existing replacement-based methods share two design constraints: full-layer granularity and contiguous selection. We argue that this is overly restrictive: in fact, redundancy in pretrained transformers is not confined to contiguous regions, nor does it evenly distribute between Attention and FeedForward outputs, implying that different strategies best approximate different submodule types and that removable components need not cluster within contiguous depth ranges. Based on this intuition, we introduce SubFit (Submodule-level Fitted residual replacement), which compresses LLMs at the submodule level: Attention and FeedForward submodules are selected non-contiguously, and each receives its own lightweight fitted residual bypass. SubFit operates post-training and requires only calibration data. Across ten LLMs (five base, five instruction-tuned), five sparsity levels from 12.5% to 37.5%, and four replacement-based baselines, SubFit achieves the best aggregate perplexity-accuracy trade-off across the evaluated sparsity levels, with larger gains under aggressive compression. At 25% sparsity, it retains 84.6% of dense downstream accuracy and incurs 2.42x perplexity degradation, against 81.6% and 4.34x for the strongest baselines, while delivering measurable inference speedup and KV-cache savings. Code is available at https://github.com/eliacunegatti/SubFit.

Elia Cunegatti, Giovanni Iacca, Doina Bucur

Finding the most influential nodes in a network is a computationally hard problem with several possible applications in various kinds of network-based problems. While several methods have been proposed for tackling the influence maximisation (IM) problem, their runtime typically scales poorly when the network size increases. Here, we propose an original method, based on network downscaling, that allows a multi-objective evolutionary algorithm (MOEA) to solve the IM problem on a reduced scale network, while preserving the relevant properties of the original network. The downscaled solution is then upscaled to the original network, using a mechanism based on centrality metrics such as PageRank. Our results on eight large networks (including two with $\sim$50k nodes) demonstrate the effectiveness of the proposed method with a more than 10-fold runtime gain compared to the time needed on the original network, and an up to $82\%$ time reduction compared to CELF.

Matteo Farina, Massimiliano Mancini, Elia Cunegatti et al.

While excellent in transfer learning, Vision-Language models (VLMs) come with high computational costs due to their large number of parameters. To address this issue, removing parameters via model pruning is a viable solution. However, existing techniques for VLMs are task-specific, and thus require pruning the network from scratch for each new task of interest. In this work, we explore a new direction: Task-Agnostic Vision-Language Pruning (TA-VLP). Given a pretrained VLM, the goal is to find a unique pruned counterpart transferable to multiple unknown downstream tasks. In this challenging setting, the transferable representations already encoded in the pretrained model are a key aspect to preserve. Thus, we propose Multimodal Flow Pruning (MULTIFLOW), a first, gradient-free, pruning framework for TA-VLP where: (i) the importance of a parameter is expressed in terms of its magnitude and its information flow, by incorporating the saliency of the neurons it connects; and (ii) pruning is driven by the emergent (multimodal) distribution of the VLM parameters after pretraining. We benchmark eight state-of-the-art pruning algorithms in the context of TA-VLP, experimenting with two VLMs, three vision-language tasks, and three pruning ratios. Our experimental results show that MULTIFLOW outperforms recent sophisticated, combinatorial competitors in the vast majority of the cases, paving the way towards addressing TA-VLP. The code is publicly available at https://github.com/FarinaMatteo/multiflow.

Elia Cunegatti, Leonardo Lucio Custode, Giovanni Iacca

The Influence Maximization (IM) problem seeks to discover the set of nodes in a graph that can spread the information propagation at most. This problem is known to be NP-hard, and it is usually studied by maximizing the influence (spread) and, optionally, optimizing a second objective, such as minimizing the seed set size or maximizing the influence fairness. However, in many practical scenarios multiple aspects of the IM problem must be optimized at the same time. In this work, we propose a first case study where several IM-specific objective functions, namely budget, fairness, communities, and time, are optimized on top of the maximization of influence and minimization of the seed set size. To this aim, we introduce MOEIM (Many-Objective Evolutionary Algorithm for Influence Maximization) a Multi-Objective Evolutionary Algorithm (MOEA) based on NSGA-II incorporating graph-aware operators and a smart initialization. We compare MOEIM in two experimental settings, including a total of nine graph datasets, two heuristic methods, a related MOEA, and a state-of-the-art Deep Learning approach. The experiments show that MOEIM overall outperforms the competitors in most of the tested many-objective settings. To conclude, we also investigate the correlation between the objectives, leading to novel insights into the topic. The codebase is available at https://github.com/eliacunegatti/MOEIM.

Fabrizio Sandri, Elia Cunegatti, Giovanni Iacca

We propose a novel Two-Stage framework for Structured Pruning (\textsc{2SSP}) for pruning Large Language Models (LLMs), which combines two different strategies of pruning, namely Width and Depth Pruning. The first stage (Width Pruning) removes entire neurons, hence their corresponding rows and columns, aiming to preserve the connectivity among the pruned structures in the intermediate state of the Feed-Forward Networks in each Transformer block. This is done based on an importance score measuring the impact of each neuron on the output magnitude. The second stage (Depth Pruning), instead, removes entire Attention submodules. This is done by applying an iterative process that removes the Attention with the minimum impact on a given metric of interest (in our case, perplexity). We also propose a novel mechanism to balance the sparsity rate of the two stages w.r.t. to the desired global sparsity. We test \textsc{2SSP} on four LLM families and three sparsity rates (25\%, 37.5\%, and 50\%), measuring the resulting perplexity over three language modeling datasets as well as the performance over six downstream tasks. Our method consistently outperforms five state-of-the-art competitors over three language modeling and six downstream tasks, with an up to two-order-of-magnitude gain in terms of pruning time. The code is available at https://github.com/FabrizioSandri/2SSP.

Elia Cunegatti, Leonardo Lucio Custode, Giovanni Iacca

Network pruning focuses on algorithms that aim to reduce a given model's computational cost by removing a subset of its parameters while having minimal impact on performance. Throughout the last decade, the most widely used pruning paradigm has been pruning and re-training, which nowadays is inconvenient due to the vast amount of pre-trained models, which are, in any case, too expensive to re-train. In this paper, we exploit functional information from dense pre-trained models, i.e., their input activations, to obtain sparse models that maximize the activations' alignment with respect to their corresponding dense models. Hence, we propose \textbf{NeuroAl}, a \emph{top-up} algorithm that can be used on top of any given pruning algorithm for LLMs, which modifies the block-wise and row-wise sparsity, exploiting information from both the dense model and its sparse version to maximize the \emph{neuron alignment} among activations. Different from existing methods, our approach adaptively selects the best hyperparameters for the block-wise and row-wise sparsity ratios w.r.t. the model and the desired sparsity, and requires \emph{no re-training}. We test our method over $\sim$300 test cases with four LLM families, three sparsity ratios, and ten language tasks (three language modeling and seven zero-shot datasets), showing how it consistently outperforms the latest state-of-the-art methods in terms of performance-runtime trade-off. The code is available at \href{https://github.com/eliacunegatti/NeuroAL}{https://github.com/eliacunegatti/NeuroAL}.

Francesco Pio Monaco, Elia Cunegatti, Flavio Vella et al.

Post-training model compression is essential for enhancing the portability of Large Language Models (LLMs) while preserving their performance. While several compression approaches have been proposed, less emphasis has been placed on selecting the most suitable set of data (the so-called \emph{calibration data}) for finding the compressed model configuration. The choice of calibration data is a critical step in preserving model capabilities both intra- and inter-tasks. In this work, we address the challenge of identifying high-performance calibration sets for both pruning and quantization by analyzing intrinsic data properties rather than model-specific signals. We introduce \texttt{\textbf{ZipCal}}, a model-agnostic data curation strategy that maximizes lexical diversity based on Zipfian power laws. Experiments demonstrate that our method consistently outperforms standard uniform random sampling across various pruning benchmarks. Notably, it also performs on par, in terms of downstream performance, with a state-of-the-art method that relies on model perplexity. The latter becomes prohibitively expensive at large-scale models and datasets, while \texttt{\textbf{ZipCal}} is on average $\sim$240$\times$ faster due to its tractable linear complexity\footnote{We make the code and the experiments available at https://anonymous.4open.science/r/zipcal-71CD/.}.

Elia Cunegatti, Matteo Farina, Doina Bucur et al.

Pruning-at-Initialization (PaI) algorithms provide Sparse Neural Networks (SNNs) which are computationally more efficient than their dense counterparts, and try to avoid performance degradation. While much emphasis has been directed towards \emph{how} to prune, we still do not know \emph{what topological metrics} of the SNNs characterize \emph{good performance}. From prior work, we have layer-wise topological metrics by which SNN performance can be predicted: the Ramanujan-based metrics. To exploit these metrics, proper ways to represent network layers via Graph Encodings (GEs) are needed, with Bipartite Graph Encodings (BGEs) being the \emph{de-facto} standard at the current stage. Nevertheless, existing BGEs neglect the impact of the inputs, and do not characterize the SNN in an end-to-end manner. Additionally, thanks to a thorough study of the Ramanujan-based metrics, we discover that they are only as good as the \emph{layer-wise density} as performance predictors, when paired with BGEs. To close both gaps, we design a comprehensive topological analysis for SNNs with both linear and convolutional layers, via (i) a new input-aware Multipartite Graph Encoding (MGE) for SNNs and (ii) the design of new end-to-end topological metrics over the MGE. With these novelties, we show the following: (a) The proposed MGE allows to extract topological metrics that are much better predictors of the accuracy drop than metrics computed from current input-agnostic BGEs; (b) Which metrics are important at different sparsity levels and for different architectures; (c) A mixture of our topological metrics can rank PaI algorithms more effectively than Ramanujan-based metrics. The codebase is publicly available at https://github.com/eliacunegatti/mge-snn.

Erik Nielsen, Elia Cunegatti, Marcus Vukojevic et al.

One of the most critical challenges in Large Language Models is their tendency to hallucinate, i.e., produce factually incorrect responses. Existing approaches show promising results in terms of hallucination correction, but still suffer from a main limitation: they apply corrections indiscriminately to every token, corrupting also the originally correct generations. To overcome this drawback, we propose PCNET, a Probabilistic Circuit trained as a tractable density estimator over the LLM residual stream. The method detects hallucinations as geometric anomalies on the factual manifold, which is done via exact Negative Log-Likelihood computation, hence without the need for sampling, external verifiers, or weight modifications, as in existing techniques. To demonstrate its effectiveness, we exploit PCNET as a dynamic gate that distinguishes hallucinated from factual hidden states at each decoding step. This triggers our second main contribution, PC-LDCD (Probabilistic Circuit Latent Density Contrastive Decoding), only when the latent geometry deviates from factual regions, while leaving correct generations untouched. Across four LLMs, ranging from 1B to 8B models, and four benchmarks covering conversational reasoning, knowledge-intensive QA, reading comprehension, and truthfulness, PCNET achieves near-perfect hallucination detection across CoQA, SQuAD v2.0, and TriviaQA, with AUROC reaching up to 99%. Moreover, PC-LDCD obtains the highest True+Info, MC2, and MC3 scores on TruthfulQA in three out of four models, in comparison with state-of-the-art baselines, while reducing the mean corruption rate to 53.7% and achieving a preservation rate of 79.3%. Our proposed method is publicly available on GitHub.

Andrea Ferigo, Elia Cunegatti, Giovanni Iacca

One of the most striking capabilities behind the learning mechanisms of the brain is the adaptation, through structural and functional plasticity, of its synapses. While synapses have the fundamental role of transmitting information across the brain, several studies show that it is the neuron activations that produce changes on synapses. Yet, most plasticity models devised for artificial Neural Networks (NNs), e.g., the ABCD rule, focus on synapses, rather than neurons, therefore optimizing synaptic-specific Hebbian parameters. This approach, however, increases the complexity of the optimization process since each synapse is associated to multiple Hebbian parameters. To overcome this limitation, we propose a novel plasticity model, called Neuron-centric Hebbian Learning (NcHL), where optimization focuses on neuron- rather than synaptic-specific Hebbian parameters. Compared to the ABCD rule, NcHL reduces the parameters from $5W$ to $5N$, being $W$ and $N$ the number of weights and neurons, and usually $N \ll W$. We also devise a ``weightless'' NcHL model, which requires less memory by approximating the weights based on a record of neuron activations. Our experiments on two robotic locomotion tasks reveal that NcHL performs comparably to the ABCD rule, despite using up to $\sim97$ times less parameters, thus allowing for scalable plasticity