Searcharxiv⌕ Search

arXiv subjects

Andrea Cavallaro

Publications and source records attributed to Andrea Cavallaro.

At least 19 recordsLinked to original sources

Mechanistic Interpretability of Deep Neural Networks: Methods, Frameworks, and Perspectives

Deep neural networks underpin fundamental tasks in a wide range of signal domains, such as image, audio, and language processing. However, their internal mechanisms are notably difficult to interpret. Internal activations are structured signals transformed across layers that may encode a multitude of interdependent relationships, as individual neurons frequently represent multiple underlying concepts. Mechanistic interpretability aims to uncover the internal mechanisms of neural networks by identifying components whose activations are causally responsible for specific behaviors. In such a context, this article provides an overview of advances in mechanistic interpretability for the analysis of causal inference in deep neural networks.

cs.LG↗

Geometric Latent Reasoning Induces Shorter Generations in LLMs

Large language models solve complex problems by generating lengthy chains of explicit reasoning tokens. While effective, this makes reasoning expensive, length-sensitive, and constrained to (discrete) natural language. While latent reasoning offers a continuous alternative, determining useful structures for intermediate latent states is an open challenge. In this paper, we formulate latent reasoning as a geometric path-approximation problem within the model's pretrained token-embedding space. We introduce Geometric Latent Reasoning (GLR), which uses a lightweight transition head to predict iterative direction updates in embedding space. Using textual chain-of-thought traces as anchors, GLR learns to approximate discrete reasoning trajectories while permitting continuous deviations from exact token embeddings. Evaluations on mathematical reasoning benchmarks using Qwen3 models reveal an emergent phenomenon: geometric latent reasoning induces substantially shorter generations without an explicit length objective. By replacing early explicit reasoning with continuous latent steps, models often reach correct answers using substantially fewer total generation steps. These findings suggest that continuous trajectories act as compact intermediate reasoning states, exposing a new tradeoff between latent computation budget, output length, and accuracy. Our code is publicly available at: https://github.com/idiap/GLR. The project page is available at: https://glr-llm.github.io/.

cs.CL↗

Logic-Guided Graph Denoising Diffusion Models

Graph diffusion models are competitive approaches for generating complex graph-structured data. However, when such distributions have strict structural constraints, diffusion models may fail to capture them exactly, resulting in the generation of invalid graphs. In this paper, we propose a method for grounding graph diffusion models in domain-specific knowledge via First Order Logic (FOL), leveraging its expressive and interpretable formalism to impose structural constraints on the generative process. We incorporate domain knowledge using classifier guidance, a technique that conditions the diffusion model's distribution on external variables to exert control over the sampled data. During the generative process, classifier guidance steers the denoising trajectory to optimize an objective function, which we define as a fuzzy and differentiable relaxation of the logical rules. Our methodology requires only the specification of constraints in FOL, whose satisfaction is encouraged at inference time to guide a pre-trained generative model. We evaluate our approach by comparing the unconditional generation of knowledge graphs with our logic-guided framework, which incorporates domain ontologies as structural constraints during inference. This approach significantly increases the proportion of generated graphs that satisfy the ontological constraints.

cs.LG↗

When quantum field theory kicks in: the relation between matter, gravity and energy

The existence of energy is the main criterion for (i) providing a common core to the materiality of all types of classical matter and (ii) denying that the gravitational field of General Relativity is a type of matter. Extending this criterion to quantum field theory, though formally straightforward, carries new conceptual consequences for the interpretation of gravity. Renormalization of the quantum matter stress-energy tensor forces a local, purely geometric term, built from the metric alone, into the conserved stress-energy, and this has two faces. It grants the metric a material significance alongside its spatio-temporal one (the egalitarian view); and it binds gravitational to non-gravitational energy so tightly that the two cannot be split apart without a choice of vacuum that a generic spacetime does not supply (the unificationist view). The latter poses a dilemma: either gravitational energy is meaningless, but then so is the energy of the non-gravitational quantum fields, and only classical fields remain matter; or both are meaningful, and the metric thereby acquires the material significance of a matter field. On neither horn does the classical divide between matter and spacetime survive quantization.

physics.hist-ph↗

Contextual Scalarisation Thompson Sampling for multi-objective decisions in public media

Recommender systems may operate under multiple, competing objectives. For example, audience reach, cultural values, public service mandate, and operational constraints must be balanced in editorial decisions of public service media. Existing approaches relying on fixed combinations of objectives or Pareto-based optimisation do not adapt to changing priorities across situations. In this paper, we propose Contextual Scalarisation Thompson Sampler (CSTS), a multi-objective contextual bandit method that learns to weight objectives as a function of the observed context. We evaluate CSTS on real programming data from Radio Télévision Suisse, the Swiss national broadcaster, showing improved contextual relevance and better alignment with expert curation practices compared to fixed weight and standard contextual bandit approaches.

cs.IR↗

On the Robustness of Vision-Language Models in Zero-shot Privacy Classification

Automatic systems for document understanding require multimodal models that accurately identify sensitive visual content, even in the presence of image degradations. Instruction-following large Vision-Language Models (VLMs) are expected to generalise across domains and tasks without requiring any specific adaptation. In this work, we systematically analyse whether VLMs can be used reliably for image privacy classification in a zero-shot setup. We evaluate and compare the classification performance of three open-source VLMs against purposely built models on two public, standard benchmarks. We assess robustness to image degradations caused by perturbations such as compression, light variations, and random noise, and analyse inference speed and parameter count to deploy VLMs in privacy-aware document processing pipelines. Our results show that large VLMs are robust to input perturbations but are less accurate (and much slower) than smaller privacy models. Scaling alone is not sufficient, highlighting the advantages of models specifically designed for privacy classification.

cs.CV↗

Reproducible Multimodal Affordance Prediction

Affordance prediction is the identification of potential actions an agent can perform on a target object from multimodal inputs. Affordance prediction methods are difficult to evaluate and compare due to heterogeneous problem formulations, inconsistent dataset annotations, incomplete reporting of experimental protocols, and limited information about deployment conditions. These limitations challenge fair benchmarking and performance comparison. To promote transparency, we propose the Affordance Sheet, a documentation detailing task formulation with its input modalities, model architectures and training information, datasets, and experimental protocols. Affordance Sheets enable reproducible benchmarking and reliable evaluation of affordance models for real-world scenarios, including generalisation to novel conditions and human safety.

cs.CV↗

Similar Models Learn Differently: Final-Window Pretraining Shapes Post-Training Beyond SFT

Developers judge a model checkpoint by how it behaves. After supervised fine-tuning (SFT), two checkpoints that perform about the same across relevant benchmarks are treated as interchangeable, equally ready for the next alignment stage, typically preference optimization. We ask whether this judgment misses a pretraining imprint: a difference that no post-SFT benchmark reveals, yet that decides how each checkpoint responds to further training. To find out, we run a controlled experiment on the final window of pretraining, the last data trained on before instruction tuning. Six branches fork from one partially pretrained checkpoint and differ only in this window: 500 million tokens, 0.1% to 1% of the tokens that precede it. Each branch trains its window on a single data source: generic web text, filtered web text, normative discourse, safety text, mathematical text, or synthetic educational text. SFT and post-training are then identical. After SFT the branches behave near-identically, within about one point on instruction following, refusal, and capability, yet the same post-training carries them to very different endpoints, under both a direct preference optimization update and a reinforcement learning update with a verifiable reward. We measure this deviation through refusal of harmful requests: when post-training begins the safety text branch refuses no more than the web text branch, yet by the end it has lost far less of its refusal. The other four branches gain little or no protection, so the effect is selective to what the window contained. The protection requires the safety text to arrive last rather than earlier in pretraining, and it reproduces on a second model family. What a model is pretrained on last shapes how it reacts to alignment. Therefore, a checkpoint should not be evaluated by its post-SFT behavior alone, and what it was trained on last should be reported with it.

cs.AI↗

Context-structured Video Anomaly Detection with Large Vision-Language Models

Training video anomaly detectors is challenging due to the difficulty and cost of annotating diverse and rare abnormal events. Although recent large vision-language models enable training-free inference, existing approaches mostly rely on holistic inference over sampled video and may miss context-specific anomaly cues. In this paper, we present CSI-VAD, a training-free video anomaly detector that identifies abnormal events across diverse contexts. The key idea is to decompose each video into three distinct contexts (environment, objects, time) and perform context-specific inference in separate branches. Because we ground anomaly judgments solely in context-specific visual cues, we do not require predefined text prompts describing abnormal events or dataset-specific tuning. Experiments on UCF-Crime and UBnormal show that CSI-VAD consistently improves over the direct holistic baseline and achieves competitive performance against existing methods, showing the advantage of structured context decomposition for training-free video anomaly detection.

cs.CV↗

Sparse Neuron Ablation Triggers Catastrophic Collapse of the Language Core in Large Vision-Language Models

Large Vision-Language Models (LVLMs) have shown impressive multimodal understanding capabilities, yet the structures that sustain their functionality remain poorly understood from a mechanistic interpretability standpoint. We propose Consistently Activated Neurons (CAN), a progressive neuron ablation method to identify critical neurons whose removal triggers catastrophic collapse, and use it to investigate structural vulnerabilities in representative 7B LVLMs. Experiments reveal that catastrophic collapse can be triggered by ablating as few as four neurons in \texttt{LLaVA-1.5-7b-hf} and a few thousand in \texttt{InstructBLIP-vicuna-7b}, both representing a small fraction of model parameters. Notably, critical neurons are predominantly localized in the language model, particularly in its down-projection layer, rather than in the vision components. We also observe a consistent two-stage collapse pattern: initial expressive degradation followed by sudden, complete collapse. These findings reveal that LVLM functionality depends on a sparse subset of neurons concentrated in the language backbone, offering mechanistic insights into how their functionality is structured and where these models are most vulnerable.

cs.AI↗

FlowOVD: Learning Generative Latent Flows for Zero-shot Open-vocabulary Detection

Open-vocabulary object detection (OVD) has achieved remarkable progress through large-scale vision-language pre-training. Existing methods, however, typically formulate OVD as a discriminative prediction problem, where decoder queries are either static or initialized from encoder features, thus limiting their diversity and flexibility. In this paper, we introduce a generative perspective by modeling decoder query generation as a continuous transport process in latent space. We propose FlowOVD, a text-conditioned query generation framework based on rectified flow that progressively transforms text-agnostic queries into text-guided queries. By introducing continuous latent query dynamics into a vision-language model (VLM) based detector, our method avoids heuristic discrete query construction and enables more expressive semantic alignment for open-vocabulary detection. Without requiring additional training data, FlowOVD achieves 49.5 AP on COCO and 31.5 AP on LVIS, outperforming GroundingDINO by +1.2 AP (+2.5 %) and +4.1 AP (+15.0 %), respectively. The larger gain on the challenging long-tailed LVIS benchmark further highlights the effectiveness of continuous query generation for open-vocabulary generalization.

cs.CV↗

NaviFormer: A Deep Reinforcement Learning Transformer-like Model to Holistically Solve the Navigation Problem

Path planning is usually solved by addressing either the (high-level) route planning problem (waypoint sequencing to achieve the final goal) or the (low-level) path planning problem (trajectory prediction between two waypoints avoiding collisions). However, real-world problems usually require simultaneous solutions to the route and path planning subproblems with a holistic and efficient approach. In this paper, we introduce NaviFormer, a deep reinforcement learning model based on a Transformer architecture that solves the global navigation problem by predicting both high-level routes and low-level trajectories. To evaluate NaviFormer, several experiments have been conducted, including comparisons with other algorithms. Results show competitive accuracy from NaviFormer since it can understand the constraints and difficulties of each subproblem and act consequently to improve performance. Moreover, its superior computation speed proves its suitability for real-time missions.

cs.RO↗

The impact of abstract and object tags on image privacy classification

Object tags denote concrete entities and are central to many computer vision tasks, whereas abstract tags capture higher-level information, which is relevant for tasks that require a contextual, potentially subjective scene understanding. Object and abstract tags extracted from images also facilitate interpretability. In this paper, we explore which type of tags is more suitable for the context-dependent and inherently subjective task of image privacy. While object tags are generally used for privacy classification, we show that abstract tags are more effective when the tag budget is limited. Conversely, when a larger number of tags per image is available, object-related information is as useful. We believe that these findings will guide future research in developing more accurate image privacy classifiers, informed by the role of tag types and quantity.

cs.CV↗

Which private attributes do VLMs agree on and predict well?

Visual Language Models (VLMs) are often used for zero-shot detection of visual attributes in the image. We present a zero-shot evaluation of open-source VLMs for privacy-related attribute recognition. We identify the attributes for which VLMs exhibit strong inter-annotator agreement, and discuss the disagreement cases of human and VLM annotations. Our results show that when evaluated against human annotations, VLMs tend to predict the presence of privacy attributes more often than human annotators. In addition to this, we find that in cases of high inter-annotator agreement between VLMs, they can complement human annotation by identifying attributes overlooked by human annotators. This highlights the potential of VLMs to support privacy annotations in large-scale image datasets.

cs.CV↗

PrivLEX: Detecting legal concepts in images through Vision-Language Models

We present PrivLEX, a novel image privacy classifier that grounds its decisions in legally defined personal data concepts. PrivLEX is the first interpretable privacy classifier aligned with legal concepts that leverages the recognition capabilities of Vision-Language Models (VLMs). PrivLEX relies on zero-shot VLM concept detection to provide interpretable classification through a label-free Concept Bottleneck Model, without requiring explicit concept labels during training. We demonstrate PrivLEX's ability to identify personal data concepts that are present in images. We further analyse the sensitivity of such concepts as perceived by human annotators of image privacy datasets.

cs.CV↗

Cross-modal Counterfactual Explanations: Uncovering Decision Factors and Dataset Biases in Subjective Classification

Concept-driven counterfactuals explain decisions of classifiers by altering the model predictions through semantic changes. In this paper, we present a novel approach that leverages cross-modal decompositionality and image-specific concepts to create counterfactual scenarios expressed in natural language. We apply the proposed interpretability framework, termed Decompose and Explain (DeX), to the challenging domain of image privacy decisions, which are contextual and subjective. This application enables the quantification of the differential contributions of key scene elements to the model prediction. We identify relevant decision factors via a multi-criterion selection mechanism that considers both image similarity for minimal perturbations and decision confidence to prioritize impactful changes. This approach evaluates and compares diverse explanations, and assesses the interdependency and mutual influence among explanatory properties. By leveraging image-specific concepts, DeX generates image-grounded, sparse explanations, yielding significant improvements over the state of the art. Importantly, DeX operates as a training-free framework, offering high flexibility. Results show that DeX not only uncovers the principal contributing factors influencing subjective decisions, but also identifies underlying dataset biases allowing for targeted mitigation strategies to improve fairness.

cs.CV↗

CarBoN: Calibrated Best-of-N Sampling Improves Test-time Reasoning

Allocating more computation during inference time (test-time scaling) improves language model performance, especially for reasoning tasks. However, popular methods like Best-of-$N$ sampling often show diminishing returns as $N$ increases. To address this inefficiency, we introduce a general test-time calibration framework that adaptively modifies the model toward high-reward reasoning paths, with theoretical guarantees of improving the lower bound of expected reward under finite sampling, all without large language model (LLM) retraining. Within this framework, we propose CarBoN (Calibrated Best-of-$N$), a two-phase method that first explores the solution space and then learns a calibration of the logits via an input-specific temperature $T$ and additive shift vector $δ$, guiding generation toward more reliable reasoning. Experiments on MATH-500 and AIME-2024 show that CarBoN improves efficiency, with up to $4\times$ fewer rollouts to reach the same accuracy, while often achieving higher accuracy under fixed budgets. We also analyze the complementary roles of $T$ and $δ$ in balancing output diversity and correctness, and demonstrate that the framework also generalizes to step-level sampling strategies such as beam search. For more information, please refer to our project page at huggingface.co/spaces/TrustSafeAI/Test-Time-Calibration.

cs.LG↗

Visual Affordance Prediction: Survey and Reproducibility

Affordances are the potential actions an agent can perform on an object, as observed by a camera. Visual affordance prediction is formulated differently for tasks such as grasping detection, affordance classification, affordance segmentation, and hand pose estimation. This diversity in formulations leads to inconsistent definitions that prevent fair comparisons between methods. In this paper, we propose a unified formulation of visual affordance prediction by accounting for the complete information on the objects of interest and the interaction of the agent with the objects to accomplish a task. This unified formulation allows us to comprehensively and systematically review disparate visual affordance works, highlighting strengths and limitations of both methods and datasets. We also discuss reproducibility issues, such as the unavailability of methods implementation and experimental setups details, making benchmarks for visual affordance prediction unfair and unreliable. To favour transparency, we introduce the Affordance Sheet, a document that details the solution, datasets, and validation of a method, supporting future reproducibility and fairness in the community.

cs.CV↗