Causal Attribution via Activation Patching

Attribution methods for Vision Transformers (ViTs) aim to identify image regions that influence model predictions, but producing faithful and well-localized attributions remains challenging. Existing gradient-based and perturbation-based techniques often fail to isolate the causal contribution of internal representations associated with individual image patches. The key challenge is that class-relevant evidence is formed through interactions between patch tokens across layers, and input-level perturbations can be poor proxies for patch importance, since they may fail to reconstruct the internal evidence actually used by the model. We propose Causal Attribution via Activation Patching (CAAP), which estimates the contribution of individual image patches to the ViT's prediction by directly intervening on internal activations rather than using learned masks or synthetic perturbation patterns. For each patch, CAAP inserts the corresponding source-image activations into a neutral target context over an intermediate range of layers and uses the resulting target-class score as the attribution signal. The resulting attribution map reflects the causal effect of patch-associated internal representations on the model's prediction. The causal intervention serves as a principled measure of patch influence by capturing class-relevant evidence after initial representation formation, while avoiding late-layer global mixing that can reduce spatial specificity. Across multiple ViT backbones and standard metrics, CAAP significantly outperforms existing methods and produces more faithful and localized attributions.

Mechanistic Interpretability of Large-Scale Counting in LLMs through a System-2 Strategy

Large language models (LLMs), despite strong performance on complex mathematical problems, exhibit systematic limitations in counting tasks. This issue arises from architectural limits of transformers, where counting is performed across layers, leading to degraded precision for larger counting problems due to depth constraints. To address this limitation, we propose a simple test-time strategy inspired by System-2 cognitive processes that decomposes large counting tasks into smaller, independent sub-problems that the model can reliably solve. We evaluate this approach using observational and causal mediation analyses to understand the underlying mechanism of this System-2-like strategy. Our mechanistic analysis identifies key components: latent counts are computed and stored in the final item representations of each part, transferred to intermediate steps via dedicated attention heads, and aggregated in the final stage to produce the total count. Experimental results demonstrate that this strategy enables LLMs to surpass architectural limitations and achieve high accuracy on large-scale counting tasks. This work provides mechanistic insight into System-2 counting in LLMs and presents a generalizable approach for improving and understanding their reasoning behavior.

Eye-Q: A Multilingual Benchmark for Visual Word Puzzle Solving and Image-to-Phrase Reasoning

Vision-Language Models (VLMs) have achieved strong performance on standard vision-language benchmarks, yet often rely on surface-level recognition rather than deeper reasoning. We propose visual word puzzles as a challenging alternative, as they require discovering implicit visual cues, generating and revising hypotheses, and mapping perceptual evidence to non-literal concepts in ways that are difficult to solve via literal grounding, OCR-heavy shortcuts, or simple retrieval-style matching. We introduce Eye-Q, a multilingual benchmark designed to assess this form of complex visual understanding. Eye-Q contains 1,343 puzzles in which a model observes a conceptually dense scene with a brief description and must infer a specific target word or phrase. The puzzles are intentionally unstructured and cue-implicit, with distractors and contextual relationships that demand selective attention, abstraction, and associative inference. The benchmark spans English, Persian, Arabic, and cross-lingual puzzles. We evaluate state-of-the-art VLMs using an open-ended, human-aligned protocol that probes hypothesis formation and revision under lightweight assistance. Results reveal substantial performance gaps, especially on abstract and cross-lingual puzzles, highlighting limitations in current models' ability to construct and search over appropriate conceptual representations for flexible image-to-phrase inference; maximum accuracy reaches only 60.27%.