Martina G. Vilas

Martina G. Vilas, Timothy Schaumlöffel, Gemma Roig

Despite the growing use of transformer models in computer vision, a mechanistic understanding of these networks is still needed. This work introduces a method to reverse-engineer Vision Transformers trained to solve image classification tasks. Inspired by previous research in NLP, we demonstrate how the inner representations at any level of the hierarchy can be projected onto the learned class embedding space to uncover how these networks build categorical representations for their predictions. We use our framework to show how image tokens develop class-specific representations that depend on attention mechanisms and contextual information, and give insights on how self-attention and MLP layers differentially contribute to this categorical composition. We additionally demonstrate that this method (1) can be used to determine the parts of an image that would be important for detecting the class of interest, and (2) exhibits significant advantages over traditional linear probing approaches. Taken together, our results position our proposed framework as a powerful tool for mechanistic interpretability and explainability research.

Santiago Galella, Pamela Osuna-Vargas, Maren Wehrheim et al.

Modern vision models achieve strong performance on standard benchmarks, yet their aggregate accuracy reveals little about which scene properties drive their predictions. Existing robustness benchmarks provide important stress tests, but typically manipulate global 2D image properties, rely on entangled real-world variation, or cover only a limited set of 3D objects and scene parameters. We introduce MAPS (Manifolds of Artificial Parametric Scenes), a scalable instrument for controlled attribution of vision model behavior to scene parameters. MAPS comprises 2,618 curated photorealistic 3D meshes validated for recognizability across 560 ImageNet classes and provides a Blender-based rendering pipeline for on-demand image generation under continuous variation of nine independent scene-factors spanning background, camera, and lighting, extensible to other factors. To showcase its applicability, we use MAPS to evaluate 20 convolutional and transformer-based models by quantifying their reliance on these scene factors through regression-based sensitivity analysis. We find a near-universal failure axis across all tested architectures: camera distance and elevation consistently dominate recognition failure regardless of ImageNet accuracy. However, the full sensitivity structure reveals that modern CNNs and transformers cluster together, distinct from older architectures, suggesting that fine-grained architectural design choices, rather than the coarse CNN-versus-transformer distinction, are the stronger determinant of sensitivity profiles.

Timothy Schaumlöffel, Martina G. Vilas, Gemma Roig

Visually-grounded language models (VLMs) are highly effective in linking visual and textual information, yet they often struggle with basic classification and localization tasks. While classification mechanisms have been studied more extensively, the processes that support object localization remain poorly understood. In this work, we investigate two representative families, LLaVA-1.5 and InternVL-3.5, using a suite of mechanistic interpretability tools, including token ablations, attention knockout, and causal mediation analysis. We find that localization is driven by a containerization mechanism in which object-aligned tokens define the spatial extent of the object, while the semantic arrangement of tokens within those boundaries is largely irrelevant to the predicted box. Only a very small set of attention heads mediates the causal effect for both classification and localization, concentrating in early-mid layers for LLaVA and mid-late layers for InternVL. The two tasks share some early processing but ultimately depend on largely distinct specialized heads. Overall, we provide the first layer- and head-level account of localization in VLMs, revealing narrow computational pathways that can guide future model design and grounding objectives.

Manshan Guo, Bhavin Choksi, Sari Sadiya et al.

In contrast to human vision, artificial neural networks (ANNs) remain relatively susceptible to adversarial attacks. To address this vulnerability, efforts have been made to transfer inductive bias from human brains to ANNs, often by training the ANN representations to match their biological counterparts. Previous works relied on brain data acquired in rodents or primates using invasive techniques, from specific regions of the brain, under non-natural conditions (anesthetized animals), and with stimulus datasets lacking diversity and naturalness. In this work, we explored whether aligning model representations to human EEG responses to a rich set of real-world images increases robustness to ANNs. Specifically, we trained ResNet50-backbone models on a dual task of classification and EEG prediction; and evaluated their EEG prediction accuracy and robustness to adversarial attacks. We observed significant correlation between the networks' EEG prediction accuracy, often highest around 100 ms post stimulus onset, and their gains in adversarial robustness. Although effect size was limited, effects were consistent across different random initializations and robust for architectural variants. We further teased apart the data from individual EEG channels and observed strongest contribution from electrodes in the parieto-occipital regions. The demonstrated utility of human EEG for such tasks opens up avenues for future efforts that scale to larger datasets under diverse stimuli conditions with the promise of stronger effects.

Martina G. Vilas, Timothy Schaumlöffel, Gemma Roig

How much scene context a single object carries is a well-studied question in human scene perception, yet how this capacity is organized in vision-language models (VLMs) remains poorly understood, with direct implications for the robustness of these models. We investigate this question through a systematic behavioral and mechanistic analysis of contextual inference from single objects. Presenting VLMs with single objects on masked backgrounds, we probe their ability to infer both fine-grained scene category and coarse superordinate context (indoor vs. outdoor). We found that single objects support above-chance inference at both levels, with performance modulated by the same object properties that predict human scene categorization. Object identity, scene, and superordinate predictions are partially dissociable: accurate inference at one level neither requires nor guarantees accurate inference at the others, and the degree of coupling differs markedly across models. Mechanistically, object representations that remain stable when background context is removed are more predictive of successful contextual inference. Scene and superordinate schemas are grounded in fundamentally different ways: scene identity is encoded in image tokens throughout the network, while superordinate information emerges only late or not at all. Together, these results reveal that the organization of contextual inference in VLMs is more complex than accuracy alone suggests, with behavioral and mechanistic signatures

Martina G. Vilas, Safoora Yousefi, Besmira Nushi et al. · cmu

Reasoning models improve their problem-solving ability through inference-time scaling, allocating more compute via longer token budgets. Identifying which reasoning traces are likely to succeed remains a key opportunity: reliably predicting productive paths can substantially reduce wasted computation and improve overall efficiency. We introduce Latent-Trajectory signals that characterize the temporal evolution of a model's internal representations during the generation of intermediate reasoning tokens. By measuring the overall change in latent representations between the start and end of reasoning, the change accumulated across intermediate steps, and the extent to which these changes advance toward the final state, we show that these signals predict solution accuracy more reliably than both cross-layer metrics and output-based confidence measures. When used to guide answer selection across multiple sampled generations, Latent-Trajectory signals make test-time scaling more effective and efficient than majority voting, reducing token usage by up to 70% while preserving and even improving accuracy by 2.6% on average. Moreover, these predictive signals often emerge early in the reasoning trace, enabling early selection and allocation of compute to the most promising candidates. Our findings contribute not only practical strategies for inference-time efficiency, but also a deeper interpretability perspective on how reasoning processes are represented and differentiated in latent space.

Martina G. Vilas, Federico Adolfi, David Poeppel et al.

Inner Interpretability is a promising emerging field tasked with uncovering the inner mechanisms of AI systems, though how to develop these mechanistic theories is still much debated. Moreover, recent critiques raise issues that question its usefulness to advance the broader goals of AI. However, it has been overlooked that these issues resemble those that have been grappled with in another field: Cognitive Neuroscience. Here we draw the relevant connections and highlight lessons that can be transferred productively between fields. Based on these, we propose a general conceptual framework and give concrete methodological strategies for building mechanistic explanations in AI inner interpretability research. With this conceptual framework, Inner Interpretability can fend off critiques and position itself on a productive path to explain AI systems.