Nathan W. Henry

Nathan W. Henry, Giovanni Luca Marchetti, Kathlén Kohn

We consider function spaces defined by self-attention networks without normalization, and theoretically analyze their geometry. Since these networks are polynomial, we rely on tools from algebraic geometry. In particular, we study the identifiability of deep attention by providing a description of the generic fibers of the parametrization for an arbitrary number of layers and, as a consequence, compute the dimension of the function space. Additionally, for a single-layer model, we characterize the singular and boundary points. Finally, we formulate a conjectural extension of our results to normalized self-attention networks, prove it for a single layer, and numerically verify it in the deep case.

Vincent Siu, Nicholas Crispino, David Park et al.

We introduce SteeringSafety, a systematic framework for evaluating representation steering methods across seven safety perspectives spanning 17 datasets. While prior work highlights general capabilities of representation steering, we systematically explore safety perspectives including bias, harmfulness, hallucination, social behaviors, reasoning, epistemic integrity, and normative judgment. Our framework provides modularized building blocks for state-of-the-art steering methods, enabling unified implementation of DIM, ACE, CAA, PCA, and LAT with recent enhancements like conditional steering. Results on Gemma-2-2B, Llama-3.1-8B, and Qwen-2.5-7B reveal that strong steering performance depends critically on pairing of method, model, and specific perspective. DIM shows consistent effectiveness, but all methods exhibit substantial entanglement: social behaviors show highest vulnerability (reaching degradation as high as 76%), jailbreaking often compromises normative judgment, and hallucination steering unpredictably shifts political views. Our findings underscore the critical need for holistic safety evaluations.

Vincent Siu, Nathan W. Henry, Nicholas Crispino et al.

While activation steering in large language models (LLMs) is a growing area of research, methods can often incur broader effects than desired. This motivates isolation of purer concept vectors to enable targeted interventions and understand LLM behavior at a more granular level. We present RepIt, a simple and data-efficient framework for isolating concept-specific representations. Across five frontier LLMs, RepIt enables precise interventions: it selectively suppresses refusal on targeted concepts while preserving refusal elsewhere, producing models that answer WMD-related questions while still scoring as safe on standard benchmarks. We further show that the corrective signal localizes to just 100-200 neurons and that robust target representations can be extracted from as few as a dozen examples on a single A6000. This efficiency raises a dual concern: manipulations can be performed with modest compute and data to extend to underrepresented data-scarce topics while evading existing benchmarks. By disentangling refusal vectors with RepIt, this work demonstrates that targeted interventions can counteract overgeneralization, laying the foundation for more granular control of model behavior.