Singular Vectors of Attention Heads Align with Features

A new study provides a theoretical foundation for using singular vectors of attention matrices to identify features inside language models, a practice already observed in mechanistic interpretability research but previously lacking rigorous justification [1][2]. The paper, authored by Gabriel Franco and posted to arXiv, asks why and when singular vectors of attention heads align with the internal features that language models use to represent concepts [1][2]. The first submission appeared on 13 Feb 2026, sized at 5,535 KB, with a revised version uploaded on 27 May 2026 at 5,599 KB [1]. The work addresses a gap in the field of mechanistic interpretability, where researchers dissect neural networks to understand their internal computations [2]. The history of artificial intelligence shows that the drive to interpret model internals has grown in parallel with the deployment of large language models, which exhibit human-like traits of knowledge and attention but whose decision-making processes remain largely opaque [4]. The researchers first demonstrated that singular vectors robustly align with features in a controlled model where features can be directly observed [1][2]. They then proved mathematically that such alignment is expected under a range of conditions [2]. To bridge the gap to real-world models where features are not directly observable, the team identified sparse attention decomposition as a testable prediction of alignment [1][2]. They presented evidence that this sparse decomposition emerges in real models in a manner consistent with their theoretical predictions [2]. The findings suggest that alignment of singular vectors with features can serve as a sound and theoretically justified basis for feature identification in language models [1][2]. This contrasts with earlier ad-hoc approaches to interpreting attention patterns. In materials science, the study of slip bands in metals provides a distant analogy: localized bands of plastic deformation indicate concentrated unidirectional slip on certain planes, creating stress concentrations that can be observed and modeled to predict material failure [5]. Similarly, the singular vectors in attention heads may act as indicators of concentrated information flow that can be systematically analyzed. The paper does not address broader cultural or religious frameworks for interpreting intelligence, though the human drive to understand and personify complex systems has ancient roots. The Slavic Native Faith, or Rodnovery, for instance, represents a modern reconstruction of pre-Christian belief systems that sought to explain natural and spiritual forces through a pantheon of gods and spirits [3]. The contemporary effort to map the internal features of artificial neural networks can be seen as a parallel, secular attempt to render the inscrutable legible through structured decomposition.