Massimiliano Mancini

Shyamgopal Karthik, Massimiliano Mancini, Zeynep Akata

The goal of open-world compositional zero-shot learning (OW-CZSL) is to recognize compositions of state and objects in images, given only a subset of them during training and no prior on the unseen compositions. In this setting, models operate on a huge output space, containing all possible state-object compositions. While previous works tackle the problem by learning embeddings for the compositions jointly, here we revisit a simple CZSL baseline and predict the primitives, i.e. states and objects, independently. To ensure that the model develops primitive-specific features, we equip the state and object classifiers with separate, non-linear feature extractors. Moreover, we estimate the feasibility of each composition through external knowledge, using this prior to remove unfeasible compositions from the output space. Finally, we propose a new setting, i.e. CZSL under partial supervision (pCZSL), where either only objects or state labels are available during training, and we can use our prior to estimate the missing labels. Our model, Knowledge-Guided Simple Primitives (KG-SP), achieves state of the art in both OW-CZSL and pCZSL, surpassing most recent competitors even when coupled with semi-supervised learning techniques. Code available at: https://github.com/ExplainableML/KG-SP.

Simone Caldarella, Davide Talon, Rahaf Aljundi et al.

Large Reasoning Models (LRMs) improve performance by generating explicit intermediate reasoning traces through increased test-time compute, yet the assumption that longer reasoning is consistently beneficial remains under-examined. While recent evidence shows that additional reasoning can lead models to overthink, we ask: "Once a model has reached the correct answer, does further reasoning refine the solution, or deviate from it?" To study the dynamics after correctness, we introduce a prefix-level trajectory evaluation protocol grounded in reasoning sufficiency, defining the minimum reasoning budget required for a model to first generate the correct answer. This allows us to disentangle verbose overthinking, where additional reasoning is redundant but harmless, from harmful overthinking, where continued reasoning destabilizes an already-correct trajectory. Starting from multimodal benchmarks, we find that many instances considered reasoning-intensive require surprisingly little reasoning. Moreover, stopping at the first correct prefix improves accuracy over standard reasoning up to 21%, revealing that current models are limited not only by their ability to reason, but also by their inability to stop at the right time. Furthermore, while common efficiency strategies like early stopping substantially reduce verbose overthinking (up to 50%), they fail to mitigate harmful overthinking. Failure analysis reveals that correctness deviations are mainly driven by logical drift and visual reinterpretation. Finally, we show that our findings generalize to language-only reasoning benchmarks, highlighting harmful overthinking as a broader reliability risk. Code available at https://simonecaldarella.github.io/thinking-past-the-answer.

Yanbei Chen, Massimiliano Mancini, Xiatian Zhu et al. · cambridge

State-of-the-art deep learning models are often trained with a large amount of costly labeled training data. However, requiring exhaustive manual annotations may degrade the model's generalizability in the limited-label regime. Semi-supervised learning and unsupervised learning offer promising paradigms to learn from an abundance of unlabeled visual data. Recent progress in these paradigms has indicated the strong benefits of leveraging unlabeled data to improve model generalization and provide better model initialization. In this survey, we review the recent advanced deep learning algorithms on semi-supervised learning (SSL) and unsupervised learning (UL) for visual recognition from a unified perspective. To offer a holistic understanding of the state-of-the-art in these areas, we propose a unified taxonomy. We categorize existing representative SSL and UL with comprehensive and insightful analysis to highlight their design rationales in different learning scenarios and applications in different computer vision tasks. Lastly, we discuss the emerging trends and open challenges in SSL and UL to shed light on future critical research directions.

Uddeshya Upadhyay, Shyamgopal Karthik, Yanbei Chen et al.

High-quality calibrated uncertainty estimates are crucial for numerous real-world applications, especially for deep learning-based deployed ML systems. While Bayesian deep learning techniques allow uncertainty estimation, training them with large-scale datasets is an expensive process that does not always yield models competitive with non-Bayesian counterparts. Moreover, many of the high-performing deep learning models that are already trained and deployed are non-Bayesian in nature and do not provide uncertainty estimates. To address these issues, we propose BayesCap that learns a Bayesian identity mapping for the frozen model, allowing uncertainty estimation. BayesCap is a memory-efficient method that can be trained on a small fraction of the original dataset, enhancing pretrained non-Bayesian computer vision models by providing calibrated uncertainty estimates for the predictions without (i) hampering the performance of the model and (ii) the need for expensive retraining the model from scratch. The proposed method is agnostic to various architectures and tasks. We show the efficacy of our method on a wide variety of tasks with a diverse set of architectures, including image super-resolution, deblurring, inpainting, and crucial application such as medical image translation. Moreover, we apply the derived uncertainty estimates to detect out-of-distribution samples in critical scenarios like depth estimation in autonomous driving. Code is available at https://github.com/ExplainableML/BayesCap.

Robert van der Klis, Stephan Alaniz, Massimiliano Mancini et al.

Fine-grained classification often requires recognizing specific object parts, such as beak shape and wing patterns for birds. Encouraging a fine-grained classification model to first detect such parts and then using them to infer the class could help us gauge whether the model is indeed looking at the right details better than with interpretability methods that provide a single attribution map. We propose PDiscoNet to discover object parts by using only image-level class labels along with priors encouraging the parts to be: discriminative, compact, distinct from each other, equivariant to rigid transforms, and active in at least some of the images. In addition to using the appropriate losses to encode these priors, we propose to use part-dropout, where full part feature vectors are dropped at once to prevent a single part from dominating in the classification, and part feature vector modulation, which makes the information coming from each part distinct from the perspective of the classifier. Our results on CUB, CelebA, and PartImageNet show that the proposed method provides substantially better part discovery performance than previous methods while not requiring any additional hyper-parameter tuning and without penalizing the classification performance. The code is available at https://github.com/robertdvdk/part_detection.

Anders Christensen, Massimiliano Mancini, A. Sophia Koepke et al.

Zero-shot learning models achieve remarkable results on image classification for samples from classes that were not seen during training. However, such models must be trained from scratch with specialised methods: therefore, access to a training dataset is required when the need for zero-shot classification arises. In this paper, we aim to equip pre-trained models with zero-shot classification capabilities without the use of image data. We achieve this with our proposed Image-free Classifier Injection with Semantics (ICIS) that injects classifiers for new, unseen classes into pre-trained classification models in a post-hoc fashion without relying on image data. Instead, the existing classifier weights and simple class-wise descriptors, such as class names or attributes, are used. ICIS has two encoder-decoder networks that learn to reconstruct classifier weights from descriptors (and vice versa), exploiting (cross-)reconstruction and cosine losses to regularise the decoding process. Notably, ICIS can be cheaply trained and applied directly on top of pre-trained classification models. Experiments on benchmark ZSL datasets show that ICIS produces unseen classifier weights that achieve strong (generalised) zero-shot classification performance. Code is available at https://github.com/ExplainableML/ImageFreeZSL .

Stephan Alaniz, Massimiliano Mancini, Anjan Dutta et al.

Humans show high-level of abstraction capabilities in games that require quickly communicating object information. They decompose the message content into multiple parts and communicate them in an interpretable protocol. Toward equipping machines with such capabilities, we propose the Primitive-based Sketch Abstraction task where the goal is to represent sketches using a fixed set of drawing primitives under the influence of a budget. To solve this task, our Primitive-Matching Network (PMN), learns interpretable abstractions of a sketch in a self supervised manner. Specifically, PMN maps each stroke of a sketch to its most similar primitive in a given set, predicting an affine transformation that aligns the selected primitive to the target stroke. We learn this stroke-to-primitive mapping end-to-end with a distance-transform loss that is minimal when the original sketch is precisely reconstructed with the predicted primitives. Our PMN abstraction empirically achieves the highest performance on sketch recognition and sketch-based image retrieval given a communication budget, while at the same time being highly interpretable. This opens up new possibilities for sketch analysis, such as comparing sketches by extracting the most relevant primitives that define an object category. Code is available at https://github.com/ExplainableML/sketch-primitives.

Stephan Alaniz, Massimiliano Mancini, Zeynep Akata

Humans are good at recomposing novel objects, i.e. they can identify commonalities between unknown objects from general structure to finer detail, an ability difficult to replicate by machines. We propose a framework, ISCO, to recompose an object using 3D superquadrics as semantic parts directly from 2D views without training a model that uses 3D supervision. To achieve this, we optimize the superquadric parameters that compose a specific instance of the object, comparing its rendered 3D view and 2D image silhouette. Our ISCO framework iteratively adds new superquadrics wherever the reconstruction error is high, abstracting first coarse regions and then finer details of the target object. With this simple coarse-to-fine inductive bias, ISCO provides consistent superquadrics for related object parts, despite not having any semantic supervision. Since ISCO does not train any neural network, it is also inherently robust to out-of-distribution objects. Experiments show that, compared to recent single instance superquadrics reconstruction approaches, ISCO provides consistently more accurate 3D reconstructions, even from images in the wild. Code available at https://github.com/ExplainableML/ISCO .

Uddeshya Upadhyay, Shyamgopal Karthik, Massimiliano Mancini et al.

Large-scale vision-language models (VLMs) like CLIP successfully find correspondences between images and text. Through the standard deterministic mapping process, an image or a text sample is mapped to a single vector in the embedding space. This is problematic: as multiple samples (images or text) can abstract the same concept in the physical world, deterministic embeddings do not reflect the inherent ambiguity in the embedding space. We propose ProbVLM, a probabilistic adapter that estimates probability distributions for the embeddings of pre-trained VLMs via inter/intra-modal alignment in a post-hoc manner without needing large-scale datasets or computing. On four challenging datasets, i.e., COCO, Flickr, CUB, and Oxford-flowers, we estimate the multi-modal embedding uncertainties for two VLMs, i.e., CLIP and BLIP, quantify the calibration of embedding uncertainties in retrieval tasks and show that ProbVLM outperforms other methods. Furthermore, we propose active learning and model selection as two real-world downstream tasks for VLMs and show that the estimated uncertainty aids both tasks. Lastly, we present a novel technique for visualizing the embedding distributions using a large-scale pre-trained latent diffusion model. Code is available at https://github.com/ExplainableML/ProbVLM.

Abhra Chaudhuri, Massimiliano Mancini, Yanbei Chen et al.

Representation learning for sketch-based image retrieval has mostly been tackled by learning embeddings that discard modality-specific information. As instances from different modalities can often provide complementary information describing the underlying concept, we propose a cross-attention framework for Vision Transformers (XModalViT) that fuses modality-specific information instead of discarding them. Our framework first maps paired datapoints from the individual photo and sketch modalities to fused representations that unify information from both modalities. We then decouple the input space of the aforementioned modality fusion network into independent encoders of the individual modalities via contrastive and relational cross-modal knowledge distillation. Such encoders can then be applied to downstream tasks like cross-modal retrieval. We demonstrate the expressive capacity of the learned representations by performing a wide range of experiments and achieving state-of-the-art results on three fine-grained sketch-based image retrieval benchmarks: Shoe-V2, Chair-V2 and Sketchy. Implementation is available at https://github.com/abhrac/xmodal-vit.

Abhra Chaudhuri, Massimiliano Mancini, Zeynep Akata et al.

Fine-grained categories that largely share the same set of parts cannot be discriminated based on part information alone, as they mostly differ in the way the local parts relate to the overall global structure of the object. We propose Relational Proxies, a novel approach that leverages the relational information between the global and local views of an object for encoding its semantic label. Starting with a rigorous formalization of the notion of distinguishability between fine-grained categories, we prove the necessary and sufficient conditions that a model must satisfy in order to learn the underlying decision boundaries in the fine-grained setting. We design Relational Proxies based on our theoretical findings and evaluate it on seven challenging fine-grained benchmark datasets and achieve state-of-the-art results on all of them, surpassing the performance of all existing works with a margin exceeding 4% in some cases. We also experimentally validate our theory on fine-grained distinguishability and obtain consistent results across multiple benchmarks. Implementation is available at https://github.com/abhrac/relational-proxies.

Moreno D'Incà, Elia Peruzzo, Massimiliano Mancini et al.

Recent progress in Text-to-Image (T2I) generative models has enabled high-quality image generation. As performance and accessibility increase, these models are gaining significant attraction and popularity: ensuring their fairness and safety is a priority to prevent the dissemination and perpetuation of biases. However, existing studies in bias detection focus on closed sets of predefined biases (e.g., gender, ethnicity). In this paper, we propose a general framework to identify, quantify, and explain biases in an open set setting, i.e. without requiring a predefined set. This pipeline leverages a Large Language Model (LLM) to propose biases starting from a set of captions. Next, these captions are used by the target generative model for generating a set of images. Finally, Vision Question Answering (VQA) is leveraged for bias evaluation. We show two variations of this framework: OpenBias and GradBias. OpenBias detects and quantifies biases, while GradBias determines the contribution of individual prompt words on biases. OpenBias effectively detects both well-known and novel biases related to people, objects, and animals and highly aligns with existing closed-set bias detection methods and human judgment. GradBias shows that neutral words can significantly influence biases and it outperforms several baselines, including state-of-the-art foundation models. Code available here: https://github.com/Moreno98/GradBias.

Abhra Chaudhuri, Massimiliano Mancini, Zeynep Akata et al.

Recent advances in fine-grained representation learning leverage local-to-global (emergent) relationships for achieving state-of-the-art results. The relational representations relied upon by such methods, however, are abstract. We aim to deconstruct this abstraction by expressing them as interpretable graphs over image views. We begin by theoretically showing that abstract relational representations are nothing but a way of recovering transitive relationships among local views. Based on this, we design Transitivity Recovering Decompositions (TRD), a graph-space search algorithm that identifies interpretable equivalents of abstract emergent relationships at both instance and class levels, and with no post-hoc computations. We additionally show that TRD is provably robust to noisy views, with empirical evidence also supporting this finding. The latter allows TRD to perform at par or even better than the state-of-the-art, while being fully interpretable. Implementation is available at https://github.com/abhrac/trd.

Ilke Cugu, Massimiliano Mancini, Yanbei Chen et al.

Generalizing visual recognition models trained on a single distribution to unseen input distributions (i.e. domains) requires making them robust to superfluous correlations in the training set. In this work, we achieve this goal by altering the training images to simulate new domains and imposing consistent visual attention across the different views of the same sample. We discover that the first objective can be simply and effectively met through visual corruptions. Specifically, we alter the content of the training images using the nineteen corruptions of the ImageNet-C benchmark and three additional transformations based on Fourier transform. Since these corruptions preserve object locations, we propose an attention consistency loss to ensure that class activation maps across original and corrupted versions of the same training sample are aligned. We name our model Attention Consistency on Visual Corruptions (ACVC). We show that ACVC consistently achieves the state of the art on three single-source domain generalization benchmarks, PACS, COCO, and the large-scale DomainNet.

Alessandro Conti, Enrico Fini, Massimiliano Mancini et al.

Recent advances in large vision-language models have revolutionized the image classification paradigm. Despite showing impressive zero-shot capabilities, a pre-defined set of categories, a.k.a. the vocabulary, is assumed at test time for composing the textual prompts. However, such assumption can be impractical when the semantic context is unknown and evolving. We thus formalize a novel task, termed as Vocabulary-free Image Classification (VIC), where we aim to assign to an input image a class that resides in an unconstrained language-induced semantic space, without the prerequisite of a known vocabulary. VIC is a challenging task as the semantic space is extremely large, containing millions of concepts, with hard-to-discriminate fine-grained categories. In this work, we first empirically verify that representing this semantic space by means of an external vision-language database is the most effective way to obtain semantically relevant content for classifying the image. We then propose Category Search from External Databases (CaSED), a method that exploits a pre-trained vision-language model and an external vision-language database to address VIC in a training-free manner. CaSED first extracts a set of candidate categories from captions retrieved from the database based on their semantic similarity to the image, and then assigns to the image the best matching candidate category according to the same vision-language model. Experiments on benchmark datasets validate that CaSED outperforms other complex vision-language frameworks, while being efficient with much fewer parameters, paving the way for future research in this direction.

Thomas De Min, Massimiliano Mancini, Stéphane Lathuilière et al.

Machine unlearning aims to erase data from a model as if the latter never saw them during training. While existing approaches unlearn information from complete or partial access to the training data, this access can be limited over time due to privacy regulations. Currently, no setting or benchmark exists to probe the effectiveness of unlearning methods in such scenarios. To fill this gap, we propose a novel task we call One-Shot Unlearning of Personal Identities (1-SHUI) that evaluates unlearning models when the training data is not available. We focus on unlearning identity data, which is specifically relevant due to current regulations requiring personal data deletion after training. To cope with data absence, we expect users to provide a portraiting picture to aid unlearning. We design requests on CelebA, CelebA-HQ, and MUFAC with different unlearning set sizes to evaluate applicable methods in 1-SHUI. Moreover, we propose MetaUnlearn, an effective method that meta-learns to forget identities from a single image. Our findings indicate that existing approaches struggle when data availability is limited, especially when there is a dissimilarity between the provided samples and the training data. Source code available at https://github.com/tdemin16/one-shui.

Shyamgopal Karthik, Karsten Roth, Massimiliano Mancini et al.

Given an image and a target modification (e.g an image of the Eiffel tower and the text "without people and at night-time"), Compositional Image Retrieval (CIR) aims to retrieve the relevant target image in a database. While supervised approaches rely on annotating triplets that is costly (i.e. query image, textual modification, and target image), recent research sidesteps this need by using large-scale vision-language models (VLMs), performing Zero-Shot CIR (ZS-CIR). However, state-of-the-art approaches in ZS-CIR still require training task-specific, customized models over large amounts of image-text pairs. In this work, we propose to tackle CIR in a training-free manner via our Compositional Image Retrieval through Vision-by-Language (CIReVL), a simple, yet human-understandable and scalable pipeline that effectively recombines large-scale VLMs with large language models (LLMs). By captioning the reference image using a pre-trained generative VLM and asking a LLM to recompose the caption based on the textual target modification for subsequent retrieval via e.g. CLIP, we achieve modular language reasoning. In four ZS-CIR benchmarks, we find competitive, in-part state-of-the-art performance - improving over supervised methods. Moreover, the modularity of CIReVL offers simple scalability without re-training, allowing us to both investigate scaling laws and bottlenecks for ZS-CIR while easily scaling up to in parts more than double of previously reported results. Finally, we show that CIReVL makes CIR human-understandable by composing image and text in a modular fashion in the language domain, thereby making it intervenable, allowing to post-hoc re-align failure cases. Code will be released upon acceptance.

Thomas De Min, Massimiliano Mancini, Karteek Alahari et al.

State-of-the-art rehearsal-free continual learning methods exploit the peculiarities of Vision Transformers to learn task-specific prompts, drastically reducing catastrophic forgetting. However, there is a tradeoff between the number of learned parameters and the performance, making such models computationally expensive. In this work, we aim to reduce this cost while maintaining competitive performance. We achieve this by revisiting and extending a simple transfer learning idea: learning task-specific normalization layers. Specifically, we tune the scale and bias parameters of LayerNorm for each continual learning task, selecting them at inference time based on the similarity between task-specific keys and the output of the pre-trained model. To make the classifier robust to incorrect selection of parameters during inference, we introduce a two-stage training procedure, where we first optimize the task-specific parameters and then train the classifier with the same selection procedure of the inference time. Experiments on ImageNet-R and CIFAR-100 show that our method achieves results that are either superior or on par with {the state of the art} while being computationally cheaper.

Simone Caldarella, Massimiliano Mancini, Elisa Ricci et al.

Vision-Language Models (VLMs) combine visual and textual understanding, rendering them well-suited for diverse tasks like generating image captions and answering visual questions across various domains. However, these capabilities are built upon training on large amount of uncurated data crawled from the web. The latter may include sensitive information that VLMs could memorize and leak, raising significant privacy concerns. In this paper, we assess whether these vulnerabilities exist, focusing on identity leakage. Our study leads to three key findings: (i) VLMs leak identity information, even when the vision-language alignment and the fine-tuning use anonymized data; (ii) context has little influence on identity leakage; (iii) simple, widely used anonymization techniques, like blurring, are not sufficient to address the problem. These findings underscore the urgent need for robust privacy protection strategies when deploying VLMs. Ethical awareness and responsible development practices are essential to mitigate these risks.

Thomas De Min, Subhankar Roy, Stéphane Lathuilière et al.

Effective collaboration begins with knowing when to ask for help. For example, when trying to identify an occluded object, a human would ask someone to remove the obstruction. Can MLLMs exhibit a similar "proactive" behavior by requesting simple user interventions? To investigate this, we introduce ProactiveBench, a benchmark built from seven repurposed datasets that tests proactiveness across different tasks such as recognizing occluded objects, enhancing image quality, and interpreting coarse sketches. We evaluate 22 MLLMs on ProactiveBench, showing that (i) they generally lack proactiveness; (ii) proactiveness does not correlate with model capacity; (iii) "hinting" at proactiveness yields only marginal gains. Surprisingly, we found that conversation histories and in-context learning introduce negative biases, hindering performance. Finally, we explore a simple fine-tuning strategy based on reinforcement learning: its results suggest that proactiveness can be learned, even generalizing to unseen scenarios. We publicly release ProactiveBench as a first step toward building proactive multimodal models.

Quentin Guimard, Federico Bartsch, Simone Caldarella et al.

Models that bridge vision and language, such as CLIP, are key components of multimodal AI, yet their large-scale, uncurated training data introduce severe social and spurious biases. Existing post-hoc debiasing methods often operate directly in the dense CLIP embedding space, where bias and task-relevant information are highly entangled. This entanglement limits their ability to remove bias without degrading semantic fidelity. In this work, we propose Sparse Embedding Modulation (SEM), a post-hoc, zero-shot debiasing framework that operates in a Sparse Autoencoder (SAE) latent space. By decomposing CLIP text embeddings into disentangled features, SEM identifies and modulates bias-relevant neurons while preserving query-relevant ones. This enables more precise, non-linear interventions. Across four benchmark datasets and two CLIP backbones, SEM achieves substantial fairness gains in retrieval and zero-shot classification. Our results demonstrate that sparse latent representations provide an effective foundation for post-hoc debiasing of vision-language models.

Yiran Huang, Lukas Thede, Massimiliano Mancini et al.

While Large Vision Language Models (LVLMs) demonstrate impressive capabilities, their substantial computational and memory requirements pose deployment challenges on resource-constrained edge devices. Current parameter reduction techniques primarily involve training LVLMs from small language models, but these methods offer limited flexibility and remain computationally intensive. We study a complementary route: compressing existing LVLMs by applying structured pruning to the language model backbone, followed by lightweight recovery training. Specifically, we investigate two structural pruning paradigms: layerwise and widthwise pruning, and pair them with supervised finetuning and knowledge distillation on logits and hidden states. Additionally, we assess the feasibility of conducting recovery training with only a small fraction of the available data. Our results show that widthwise pruning generally maintains better performance in low-resource scenarios, where computational resources are limited or there is insufficient finetuning data. As for the recovery training, finetuning only the multimodal projector is sufficient at small compression levels. Furthermore, a combination of supervised finetuning and hidden-state distillation yields optimal recovery across various pruning levels. Notably, effective recovery can be achieved using just 5% of the original data, while retaining over 95% of the original performance. Through empirical study on three representative LVLM families ranging from 3B to 7B parameters, this study offers actionable insights for practitioners to compress LVLMs without extensive computation resources or sufficient data. The code base is available at https://github.com/YiranHuangIrene/VLMCompression.git.

Marco Garosi, Riccardo Tedoldi, Davide Boscaini et al.

Supervised 3D part segmentation models are tailored for a fixed set of objects and parts, limiting their transferability to open-set, real-world scenarios. Recent works have explored vision-language models (VLMs) as a promising alternative, using multi-view rendering and textual prompting to identify object parts. However, naively applying VLMs in this context introduces several drawbacks, such as the need for meticulous prompt engineering, and fails to leverage the 3D geometric structure of objects. To address these limitations, we propose COPS, a COmprehensive model for Parts Segmentation that blends the semantics extracted from visual concepts and 3D geometry to effectively identify object parts. COPS renders a point cloud from multiple viewpoints, extracts 2D features, projects them back to 3D, and uses a novel geometric-aware feature aggregation procedure to ensure spatial and semantic consistency. Finally, it clusters points into parts and labels them. We demonstrate that COPS is efficient, scalable, and achieves zero-shot state-of-the-art performance across five datasets, covering synthetic and real-world data, texture-less and coloured objects, as well as rigid and non-rigid shapes. The code is available at https://3d-cops.github.io.

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.

Francesco Gentile, Nicola Dall'Asen, Francesco Tonini et al.

As vision-language models are deployed at scale, understanding their internal mechanisms becomes increasingly critical. Existing interpretability methods predominantly rely on activations, making them dataset-dependent, vulnerable to data bias, and often restricted to coarse head-level explanations. We introduce SITH (Semantic Inspection of Transformer Heads), a fully data-free, training-free framework that directly analyzes CLIP's vision transformer in weight space. For each attention head, we decompose its value-output matrix into singular vectors and interpret each one via COMP (Coherent Orthogonal Matching Pursuit), a new algorithm that explains them as sparse, semantically coherent combinations of human-interpretable concepts. We show that SITH yields coherent, faithful intra-head explanations, validated through reconstruction fidelity and interpretability experiments. This allows us to use SITH for precise, interpretable weight-space model edits that amplify or suppress specific concepts, improving downstream performance without retraining. Furthermore, we use SITH to study model adaptation, showing how fine-tuning primarily reweights a stable semantic basis rather than learning entirely new features.

Mingxuan Liu, Tyler L. Hayes, Massimiliano Mancini et al.

Open-vocabulary object detection models allow users to freely specify a class vocabulary in natural language at test time, guiding the detection of desired objects. However, vocabularies can be overly broad or even mis-specified, hampering the overall performance of the detector. In this work, we propose a plug-and-play Vocabulary Adapter (VocAda) to refine the user-defined vocabulary, automatically tailoring it to categories that are relevant for a given image. VocAda does not require any training, it operates at inference time in three steps: i) it uses an image captionner to describe visible objects, ii) it parses nouns from those captions, and iii) it selects relevant classes from the user-defined vocabulary, discarding irrelevant ones. Experiments on COCO and Objects365 with three state-of-the-art detectors show that VocAda consistently improves performance, proving its versatility. The code is open source.

Marco Garosi, Matteo Farina, Alessandro Conti et al.

Which multimodal model should we use for classification? Previous studies suggest that the answer lies in CLIP-like contrastive Vision-Language Models (VLMs), due to their remarkable performance in zero-shot classification. In contrast, Large Multimodal Models (LMM) are more suitable for complex tasks. In this work, we argue that this answer overlooks an important capability of LMMs: in-context learning. We benchmark state-of-the-art LMMs on diverse datasets for closed-world classification and find that, although their zero-shot performance is lower than CLIP's, LMMs with a few in-context examples can match or even surpass contrastive VLMs with cache-based adapters, their "in-context" equivalent. We extend this analysis to the open-world setting, where the generative nature of LMMs makes them more suitable for the task. In this challenging scenario, LMMs struggle whenever provided with imperfect context information. To address this issue, we propose CIRCLE, a simple training-free method that assigns pseudo-labels to in-context examples, iteratively refining them with the available context itself. Through extensive experiments, we show that CIRCLE establishes a robust baseline for open-world classification, surpassing VLM counterparts and highlighting the potential of LMMs to serve as unified classifiers, and a flexible alternative to specialized models.

Davide Berasi, Matteo Farina, Massimiliano Mancini et al.

Vision-Language Models (VLMs) learn a shared feature space for text and images, enabling the comparison of inputs of different modalities. While prior works demonstrated that VLMs organize natural language representations into regular structures encoding composite meanings, it remains unclear if compositional patterns also emerge in the visual embedding space. In this work, we investigate compositionality in the image domain, where the analysis of compositional properties is challenged by noise and sparsity of visual data. We address these problems and propose a framework, called Geodesically Decomposable Embeddings (GDE), that approximates image representations with geometry-aware compositional structures in the latent space. We demonstrate that visual embeddings of pre-trained VLMs exhibit a compositional arrangement, and evaluate the effectiveness of this property in the tasks of compositional classification and group robustness. GDE achieves stronger performance in compositional classification compared to its counterpart method that assumes linear geometry of the latent space. Notably, it is particularly effective for group robustness, where we achieve higher results than task-specific solutions. Our results indicate that VLMs can automatically develop a human-like form of compositional reasoning in the visual domain, making their underlying processes more interpretable. Code is available at https://github.com/BerasiDavide/vlm_image_compositionality.

Thomas De Min, Subhankar Roy, Stéphane Lathuilière et al.

Machine unlearning is an emerging paradigm to remove the influence of specific training data (i.e., the forget set) from a model while preserving its knowledge of the rest of the data (i.e., the retain set). Previous approaches assume the forget data to be uniformly distributed from all training datapoints. However, if the data to unlearn is dominant in one group, we empirically show that performance for this group degrades, leading to fairness issues. This work tackles the overlooked problem of non-uniformly distributed forget sets, which we call group-robust machine unlearning, by presenting a simple, effective strategy that mitigates the performance loss in dominant groups via sample distribution reweighting. Moreover, we present MIU (Mutual Information-aware Machine Unlearning), the first approach for group robustness in approximate machine unlearning. MIU minimizes the mutual information between model features and group information, achieving unlearning while reducing performance degradation in the dominant group of the forget set. Additionally, MIU exploits sample distribution reweighting and mutual information calibration with the original model to preserve group robustness. We conduct experiments on three datasets and show that MIU outperforms standard methods, achieving unlearning without compromising model robustness. Source code available at https://github.com/tdemin16/group-robust_machine_unlearning.

Davide Berasi, Matteo Farina, Massimiliano Mancini et al.

Selecting the best data mixture is critical for successful Supervised Fine-Tuning (SFT) of Multimodal Large Language Models. However, determining the optimal mixture weights across multiple domain-specific datasets remains a significant bottleneck due to the combinatorial search space and the high cost associated with even a single training run. This is the so-called Data Mixture Optimization (DMO) problem. On the other hand, model merging unifies domain-specific experts through parameter interpolation. This strategy is efficient, as it only requires a single training run per domain, yet oftentimes leads to suboptimal models. In this work, we take the best of both worlds, studying model merging as an efficient strategy for estimating the performance of different data mixtures. We train domain-specific multimodal experts and evaluate their weighted parameter-space combinations to estimate the efficacy of corresponding data mixtures. We conduct extensive experiments on 14 multimodal benchmarks, and empirically demonstrate that the merged proxy models exhibit a high rank correlation with models trained on actual data mixtures. This decouples the search for optimal mixtures from the resource-intensive training process, thereby providing a scalable and efficient strategy for navigating the complex landscape of mixture weights. Code is publicly available at https://github.com/BerasiDavide/mLLMs_merging_4_DMO.

Thomas De Min, Massimiliano Mancini, Stéphane Lathuilière et al.

Prompt tuning has emerged as an effective rehearsal-free technique for class-incremental learning (CIL) that learns a tiny set of task-specific parameters (or prompts) to instruct a pre-trained transformer to learn on a sequence of tasks. Albeit effective, prompt tuning methods do not lend well in the multi-label class incremental learning (MLCIL) scenario (where an image contains multiple foreground classes) due to the ambiguity in selecting the correct prompt(s) corresponding to different foreground objects belonging to multiple tasks. To circumvent this issue we propose to eliminate the prompt selection mechanism by maintaining task-specific pathways, which allow us to learn representations that do not interact with the ones from the other tasks. Since independent pathways in truly incremental scenarios will result in an explosion of computation due to the quadratically complex multi-head self-attention (MSA) operation in prompt tuning, we propose to reduce the original patch token embeddings into summarized tokens. Prompt tuning is then applied to these fewer summarized tokens to compute the final representation. Our proposed method Multi-Label class incremental learning via summarising pAtch tokeN Embeddings (MULTI-LANE) enables learning disentangled task-specific representations in MLCIL while ensuring fast inference. We conduct experiments in common benchmarks and demonstrate that our MULTI-LANE achieves a new state-of-the-art in MLCIL. Additionally, we show that MULTI-LANE is also competitive in the CIL setting. Source code available at https://github.com/tdemin16/multi-lane

Shyamgopal Karthik, Karsten Roth, Massimiliano Mancini et al.

Despite their impressive capabilities, diffusion-based text-to-image (T2I) models can lack faithfulness to the text prompt, where generated images may not contain all the mentioned objects, attributes or relations. To alleviate these issues, recent works proposed post-hoc methods to improve model faithfulness without costly retraining, by modifying how the model utilizes the input prompt. In this work, we take a step back and show that large T2I diffusion models are more faithful than usually assumed, and can generate images faithful to even complex prompts without the need to manipulate the generative process. Based on that, we show how faithfulness can be simply treated as a candidate selection problem instead, and introduce a straightforward pipeline that generates candidate images for a text prompt and picks the best one according to an automatic scoring system that can leverage already existing T2I evaluation metrics. Quantitative comparisons alongside user studies on diverse benchmarks show consistently improved faithfulness over post-hoc enhancement methods, with comparable or lower computational cost. Code is available at \url{https://github.com/ExplainableML/ImageSelect}.

Dario Fontanel, Fabio Cermelli, Massimiliano Mancini et al.

Robotic visual systems operating in the wild must act in unconstrained scenarios, under different environmental conditions while facing a variety of semantic concepts, including unknown ones. To this end, recent works tried to empower visual object recognition methods with the capability to i) detect unseen concepts and ii) extended their knowledge over time, as images of new semantic classes arrive. This setting, called Open World Recognition (OWR), has the goal to produce systems capable of breaking the semantic limits present in the initial training set. However, this training set imposes to the system not only its own semantic limits, but also environmental ones, due to its bias toward certain acquisition conditions that do not necessarily reflect the high variability of the real-world. This discrepancy between training and test distribution is called domain-shift. This work investigates whether OWR algorithms are effective under domain-shift, presenting the first benchmark setup for assessing fairly the performances of OWR algorithms, with and without domain-shift. We then use this benchmark to conduct analyses in various scenarios, showing how existing OWR algorithms indeed suffer a severe performance degradation when train and test distributions differ. Our analysis shows that this degradation is only slightly mitigated by coupling OWR with domain generalization techniques, indicating that the mere plug-and-play of existing algorithms is not enough to recognize new and unknown categories in unseen domains. Our results clearly point toward open issues and future research directions, that need to be investigated for building robot visual systems able to function reliably under these challenging yet very real conditions. Code available at https://github.com/DarioFontanel/OWR-VisualDomains

Dario Fontanel, Fabio Cermelli, Massimiliano Mancini et al.

Traditional semantic segmentation methods can recognize at test time only the classes that are present in the training set. This is a significant limitation, especially for semantic segmentation algorithms mounted on intelligent autonomous systems, deployed in realistic settings. Regardless of how many classes the system has seen at training time, it is inevitable that unexpected, unknown objects will appear at test time. The failure in identifying such anomalies may lead to incorrect, even dangerous behaviors of the autonomous agent equipped with such segmentation model when deployed in the real world. Current state of the art of anomaly segmentation uses generative models, exploiting their incapability to reconstruct patterns unseen during training. However, training these models is expensive, and their generated artifacts may create false anomalies. In this paper we take a different route and we propose to address anomaly segmentation through prototype learning. Our intuition is that anomalous pixels are those that are dissimilar to all class prototypes known by the model. We extract class prototypes from the training data in a lightweight manner using a cosine similarity-based classifier. Experiments on StreetHazards show that our approach achieves the new state of the art, with a significant margin over previous works, despite the reduced computational overhead. Code is available at https://github.com/DarioFontanel/PAnS.

Luca Zanella, Willi Menapace, Massimiliano Mancini et al.

Video anomaly detection (VAD) aims to temporally locate abnormal events in a video. Existing works mostly rely on training deep models to learn the distribution of normality with either video-level supervision, one-class supervision, or in an unsupervised setting. Training-based methods are prone to be domain-specific, thus being costly for practical deployment as any domain change will involve data collection and model training. In this paper, we radically depart from previous efforts and propose LAnguage-based VAD (LAVAD), a method tackling VAD in a novel, training-free paradigm, exploiting the capabilities of pre-trained large language models (LLMs) and existing vision-language models (VLMs). We leverage VLM-based captioning models to generate textual descriptions for each frame of any test video. With the textual scene description, we then devise a prompting mechanism to unlock the capability of LLMs in terms of temporal aggregation and anomaly score estimation, turning LLMs into an effective video anomaly detector. We further leverage modality-aligned VLMs and propose effective techniques based on cross-modal similarity for cleaning noisy captions and refining the LLM-based anomaly scores. We evaluate LAVAD on two large datasets featuring real-world surveillance scenarios (UCF-Crime and XD-Violence), showing that it outperforms both unsupervised and one-class methods without requiring any training or data collection.

Moreno D'Incà, Elia Peruzzo, Massimiliano Mancini et al. · gatech

Text-to-image generative models are becoming increasingly popular and accessible to the general public. As these models see large-scale deployments, it is necessary to deeply investigate their safety and fairness to not disseminate and perpetuate any kind of biases. However, existing works focus on detecting closed sets of biases defined a priori, limiting the studies to well-known concepts. In this paper, we tackle the challenge of open-set bias detection in text-to-image generative models presenting OpenBias, a new pipeline that identifies and quantifies the severity of biases agnostically, without access to any precompiled set. OpenBias has three stages. In the first phase, we leverage a Large Language Model (LLM) to propose biases given a set of captions. Secondly, the target generative model produces images using the same set of captions. Lastly, a Vision Question Answering model recognizes the presence and extent of the previously proposed biases. We study the behavior of Stable Diffusion 1.5, 2, and XL emphasizing new biases, never investigated before. Via quantitative experiments, we demonstrate that OpenBias agrees with current closed-set bias detection methods and human judgement.

Ananthu Aniraj, Cassio F. Dantas, Dino Ienco et al.

Part-based reasoning is a classical strategy to make a computer vision model directly focus on the object parts that are relevant to the downstream task. In the context of deep learning, this also serves to improve by-design interpretability, often by using part-centric attention mechanisms on top of a latent image representation provided by a standard, black-box model. This approach is based on a locality assumption: that the latent representation of an object part encodes primarily information about the corresponding image region. In this work, we test this basic assumption, measuring intra-object leakage in vision models using part-based attribute annotations. Through a comprehensive experimental evaluation, we show that modern pretrained vision transformers violate the locality assumption and exhibit a strong intra-object leakage, in which each part encodes information from the whole object, a visual metonymy that compromises the faithfulness of attention-based interpretable-by-design methods for part-based reasoning, ultimately rendering them uninterpretable. In addition, we establish an upper bound using a two-stage approach that prevents leakage by design. We then show that this inherently disentangled feature extraction improves attribute-driven part discovery on a variety of tasks, confirming the practical impact of intra-object leakage. Our results uncover a neglected issue affecting the interpretability of part-based representations, such as those in CBMs relying on part-centric concepts, highlighting that two-stage approaches offer a promising way to mitigate it.

Matteo Farina, Massimiliano Mancini, Giovanni Iacca et al.

An old-school recipe for training a classifier is to (i) learn a good feature extractor and (ii) optimize a linear layer atop. When only a handful of samples are available per category, as in Few-Shot Adaptation (FSA), data are insufficient to fit a large number of parameters, rendering the above impractical. This is especially true with large pre-trained Vision-Language Models (VLMs), which motivated successful research at the intersection of Parameter-Efficient Fine-tuning (PEFT) and FSA. In this work, we start by analyzing the learning dynamics of PEFT techniques when trained on few-shot data from only a subset of categories, referred to as the ``base'' classes. We show that such dynamics naturally splits into two distinct phases: (i) task-level feature extraction and (ii) specialization to the available concepts. To accommodate this dynamic, we then depart from prompt- or adapter-based methods and tackle FSA differently. Specifically, given a fixed computational budget, we split it to (i) learn a task-specific feature extractor via PEFT and (ii) train a linear classifier on top. We call this scheme Two-Stage Few-Shot Adaptation (2SFS). Differently from established methods, our scheme enables a novel form of selective inference at a category level, i.e., at test time, only novel categories are embedded by the adapted text encoder, while embeddings of base categories are available within the classifier. Results with fixed hyperparameters across two settings, three backbones, and eleven datasets, show that 2SFS matches or surpasses the state-of-the-art, while established methods degrade significantly across settings.

Alessandro Conti, Enrico Fini, Massimiliano Mancini et al.

Large vision-language models revolutionized image classification and semantic segmentation paradigms. However, they typically assume a pre-defined set of categories, or vocabulary, at test time for composing textual prompts. This assumption is impractical in scenarios with unknown or evolving semantic context. Here, we address this issue and introduce the Vocabulary-free Image Classification (VIC) task, which aims to assign a class from an unconstrained language-induced semantic space to an input image without needing a known vocabulary. VIC is challenging due to the vastness of the semantic space, which contains millions of concepts, including fine-grained categories. To address VIC, we propose Category Search from External Databases (CaSED), a training-free method that leverages a pre-trained vision-language model and an external database. CaSED first extracts the set of candidate categories from the most semantically similar captions in the database and then assigns the image to the best-matching candidate category according to the same vision-language model. Furthermore, we demonstrate that CaSED can be applied locally to generate a coarse segmentation mask that classifies image regions, introducing the task of Vocabulary-free Semantic Segmentation. CaSED and its variants outperform other more complex vision-language models, on classification and semantic segmentation benchmarks, while using much fewer parameters.

Rémi Kazmierczak, Steve Azzolin, Eloïse Berthier et al.

We introduce PASTA (Perceptual Assessment System for explanaTion of Artificial Intelligence), a novel human-centric framework for evaluating eXplainable AI (XAI) techniques in computer vision. Our first contribution is the creation of the PASTA-dataset, the first large-scale benchmark that spans a diverse set of models and both saliency-based and concept-based explanation methods. This dataset enables robust, comparative analysis of XAI techniques based on human judgment. Our second contribution is an automated, data-driven benchmark that predicts human preferences using the PASTA-dataset. This scoring called PASTA-score method offers scalable, reliable, and consistent evaluation aligned with human perception. Additionally, our benchmark allows for comparisons between explanations across different modalities, an aspect previously unaddressed. We then propose to apply our scoring method to probe the interpretability of existing models and to build more human interpretable XAI methods.

Deepayan Das, Davide Talon, Massimiliano Mancini et al.

Vision-Language Models (VLMs) have shown significant promise in Visual Question Answering (VQA) tasks by leveraging web-scale multimodal datasets. However, these models often struggle with continual learning due to catastrophic forgetting when adapting to new tasks. As an effective remedy to mitigate catastrophic forgetting, rehearsal strategy uses the data of past tasks upon learning new task. However, such strategy incurs the need of storing past data, which might not be feasible due to hardware constraints or privacy concerns. In this work, we propose the first data-free method that leverages the language generation capability of a VLM, instead of relying on external models, to produce pseudo-rehearsal data for addressing continual VQA. Our proposal, named as GaB, generates pseudo-rehearsal data by posing previous task questions on new task data. Yet, despite being effective, the distribution of generated questions skews towards the most frequently posed questions due to the limited and task-specific training data. To mitigate this issue, we introduce a pseudo-rehearsal balancing module that aligns the generated data towards the ground-truth data distribution using either the question meta-statistics or an unsupervised clustering method. We evaluate our proposed method on two recent benchmarks, \ie VQACL-VQAv2 and CLOVE-function benchmarks. GaB outperforms all the data-free baselines with substantial improvement in maintaining VQA performance across evolving tasks, while being on-par with methods with access to the past data.

Alessandro Conti, Massimiliano Mancini, Enrico Fini et al.

Traditional image classification requires a predefined list of semantic categories. In contrast, Large Multimodal Models (LMMs) can sidestep this requirement by classifying images directly using natural language (e.g., answering the prompt "What is the main object in the image?"). Despite this remarkable capability, most existing studies on LMM classification performance are surprisingly limited in scope, often assuming a closed-world setting with a predefined set of categories. In this work, we address this gap by thoroughly evaluating LMM classification performance in a truly open-world setting. We first formalize the task and introduce an evaluation protocol, defining various metrics to assess the alignment between predicted and ground truth classes. We then evaluate 13 models across 10 benchmarks, encompassing prototypical, non-prototypical, fine-grained, and very fine-grained classes, demonstrating the challenges LMMs face in this task. Further analyses based on the proposed metrics reveal the types of errors LMMs make, highlighting challenges related to granularity and fine-grained capabilities, showing how tailored prompting and reasoning can alleviate them.

Quentin Guimard, Moreno D'Incà, Massimiliano Mancini et al.

A person downloading a pre-trained model from the web should be aware of its biases. Existing approaches for bias identification rely on datasets containing labels for the task of interest, something that a non-expert may not have access to, or may not have the necessary resources to collect: this greatly limits the number of tasks where model biases can be identified. In this work, we present Classifier-to-Bias (C2B), the first bias discovery framework that works without access to any labeled data: it only relies on a textual description of the classification task to identify biases in the target classification model. This description is fed to a large language model to generate bias proposals and corresponding captions depicting biases together with task-specific target labels. A retrieval model collects images for those captions, which are then used to assess the accuracy of the model w.r.t. the given biases. C2B is training-free, does not require any annotations, has no constraints on the list of biases, and can be applied to any pre-trained model on any classification task. Experiments on two publicly available datasets show that C2B discovers biases beyond those of the original datasets and outperforms a recent state-of-the-art bias detection baseline that relies on task-specific annotations, being a promising first step toward addressing task-agnostic unsupervised bias detection.

Deepayan Das, Davide Talon, Yiming Wang et al.

Vision Language Models (VLMs) have lead to major improvements in multimodal reasoning, yet they still struggle to understand user-specific concepts. Existing personalization methods address this limitation but heavily rely on training procedures, that can be either costly or unpleasant to individual users. We depart from existing work, and for the first time explore the training-free setting in the context of personalization. We propose a novel method, Retrieval and Reasoning for Personalization (R2P), leveraging internal knowledge of VLMs. First, we leverage VLMs to extract the concept fingerprint, i.e., key attributes uniquely defining the concept within its semantic class. When a query arrives, the most similar fingerprints are retrieved and scored via chain-of-thought-reasoning. To reduce the risk of hallucinations, the scores are validated through cross-modal verification at the attribute level: in case of a discrepancy between the scores, R2P refines the concept association via pairwise multimodal matching, where the retrieved fingerprints and their images are directly compared with the query. We validate R2P on two publicly available benchmarks and a newly introduced dataset, Personal Concepts with Visual Ambiguity (PerVA), for concept identification highlighting challenges in visual ambiguity. R2P consistently outperforms state-of-the-art approaches on various downstream tasks across all benchmarks. Code will be available upon acceptance.

Luca Zanella, Massimiliano Mancini, Willi Menapace et al.

Recent video-language alignment models are trained on sets of videos, each with an associated positive caption and a negative caption generated by large language models. A problem with this procedure is that negative captions may introduce linguistic biases, i.e., concepts are seen only as negatives and never associated with a video. While a solution would be to collect videos for the negative captions, existing databases lack the fine-grained variations needed to cover all possible negatives. In this work, we study whether synthetic videos can help to overcome this issue. Our preliminary analysis with multiple generators shows that, while promising on some tasks, synthetic videos harm the performance of the model on others. We hypothesize this issue is linked to noise (semantic and visual) in the generated videos and develop a method, SynViTA, that accounts for those. SynViTA dynamically weights the contribution of each synthetic video based on how similar its target caption is w.r.t. the real counterpart. Moreover, a semantic consistency loss makes the model focus on fine-grained differences across captions, rather than differences in video appearance. Experiments show that, on average, SynViTA improves over existing methods on VideoCon test sets and SSv2-Temporal, SSv2-Events, and ATP-Hard benchmarks, being a first promising step for using synthetic videos when learning video-language models.

Marco Garosi, Alessandro Conti, Gaowen Liu et al.

Attribute detection is crucial for many computer vision tasks, as it enables systems to describe properties such as color, texture, and material. Current approaches often rely on labor-intensive annotation processes which are inherently limited: objects can be described at an arbitrary level of detail (e.g., color vs. color shades), leading to ambiguities when the annotators are not instructed carefully. Furthermore, they operate within a predefined set of attributes, reducing scalability and adaptability to unforeseen downstream applications. We present Compositional Caching (ComCa), a training-free method for open-vocabulary attribute detection that overcomes these constraints. ComCa requires only the list of target attributes and objects as input, using them to populate an auxiliary cache of images by leveraging web-scale databases and Large Language Models to determine attribute-object compatibility. To account for the compositional nature of attributes, cache images receive soft attribute labels. Those are aggregated at inference time based on the similarity between the input and cache images, refining the predictions of underlying Vision-Language Models (VLMs). Importantly, our approach is model-agnostic, compatible with various VLMs. Experiments on public datasets demonstrate that ComCa significantly outperforms zero-shot and cache-based baselines, competing with recent training-based methods, proving that a carefully designed training-free approach can successfully address open-vocabulary attribute detection.

Luca Zanella, Massimiliano Mancini, Yiming Wang et al.

Given a task and a set of steps composing it, Video Step Grounding (VSG) aims to detect which steps are performed in a video. Standard approaches for this task require a labeled training set (e.g., with step-level annotations or narrations), which may be costly to collect. Moreover, they process the full video offline, limiting their applications for scenarios requiring online decisions. Thus, in this work, we explore how to perform VSG online and without training. We achieve this by exploiting the zero-shot capabilities of recent Large Multimodal Models (LMMs). In particular, we use LMMs to predict the step associated with a restricted set of frames, without access to the whole video. We show that this online strategy without task-specific tuning outperforms offline and training-based models. Motivated by this finding, we develop Bayesian Grounding with Large Multimodal Models (BaGLM), further injecting knowledge of past frames into the LMM-based predictions. BaGLM exploits Bayesian filtering principles, modeling step transitions via (i) a dependency matrix extracted through large language models and (ii) an estimation of step progress. Experiments on three datasets show superior performance of BaGLM over state-of-the-art training-based offline methods.

Yiran Huang, Lukas Thede, Massimiliano Mancini et al.

While Multimodal Large Language Models (MLLMs) demonstrate impressive capabilities, their substantial computational and memory requirements pose significant barriers to practical deployment. Current parameter reduction techniques primarily involve training MLLMs from Small Language Models (SLMs), but these methods offer limited flexibility and remain computationally intensive. To address this gap, we propose to directly compress existing MLLMs through structural pruning combined with efficient recovery training. Specifically, we investigate two structural pruning paradigms--layerwise and widthwise pruning--applied to the language model backbone of MLLMs, alongside supervised finetuning and knowledge distillation. Additionally, we assess the feasibility of conducting recovery training with only a small fraction of the available data. Our results show that widthwise pruning generally maintains better performance in low-resource scenarios with limited computational resources or insufficient finetuning data. As for the recovery training, finetuning only the multimodal projector is sufficient at small compression levels (< 20%). Furthermore, a combination of supervised finetuning and hidden-state distillation yields optimal recovery across various pruning levels. Notably, effective recovery can be achieved with as little as 5% of the original training data, while retaining over 95% of the original performance. Through empirical study on two representative MLLMs, i.e., LLaVA-v1.5-7B and Bunny-v1.0-3B, this study offers actionable insights for practitioners aiming to compress MLLMs effectively without extensive computation resources or sufficient data.

Moreno D'Incà, Elia Peruzzo, Xingqian Xu et al.

Vision-language models (VLMs) often inherit the biases and unsafe associations present within their large-scale training dataset. While recent approaches mitigate unsafe behaviors, their evaluation focuses on how safe the model is on unsafe inputs, ignoring potential shortcomings on safe ones. In this paper, we first revise safety evaluation by introducing SafeGround, a new set of metrics that evaluate safety at different levels of granularity. With this metric, we uncover a surprising issue of training-based methods: they make the model less safe on safe inputs. From this finding, we take a different direction and explore whether it is possible to make a model safer without training, introducing Unsafe Weights Manipulation (UWM). UWM uses a calibration set of safe and unsafe instances to compare activations between safe and unsafe content, identifying the most important parameters for processing the latter. Their values are then manipulated via negation. Experiments show that UWM achieves the best tradeoff between safety and knowledge preservation, consistently improving VLMs on unsafe queries while outperforming even training-based state-of-the-art methods on safe ones.

Alessandro Conti, Enrico Fini, Paolo Rota et al.

Assessing the capabilities of large multimodal models (LMMs) often requires the creation of ad-hoc evaluations. Currently, building new benchmarks requires tremendous amounts of manual work for each specific analysis. This makes the evaluation process tedious and costly. In this paper, we present APEx, Automatic Programming of Experiments, the first framework for automatic benchmarking of LMMs. Given a research question expressed in natural language, APEx leverages a large language model (LLM) and a library of pre-specified tools to generate a set of experiments for the model at hand, and progressively compile a scientific report. The report drives the testing procedure: based on the current status of the investigation, APEx chooses which experiments to perform and whether the results are sufficient to draw conclusions. Finally, the LLM refines the report, presenting the results to the user in natural language. Thanks to its modularity, our framework is flexible and extensible as new tools become available. Empirically, APEx reproduces the findings of existing studies while allowing for arbitrary analyses and hypothesis testing.