Behavioral clustering — is evidence for → Linear representation hypothesis

The linear representation hypothesis claims that high-level concepts are encoded as linear directions in an LLM's activation space, cited by CAA as part of its motivation for why activation-engineering techniques should work at all (contrastive-activation-addition, §"2 Related work", p. 3). Behavioral clustering is a direct empirical instance of that claim rather than just background motivation for it: when residual-stream activations on a contrastive dataset are projected onto only two principal components -- a linear projection -- a high-level behavior like refusal becomes visibly separable, and does so suddenly at a specific layer rather than gradually across many (contrastive-activation-addition, §"3.2 Visualizing activations for contrastive dataset analysis", p. 4). If the behavior's encoding were nonlinear or diffusely distributed, a two-dimensional linear projection would not be expected to reveal a clean split. The suddenness itself is informative: it suggests the model is not refining a distributed encoding across many layers but computing the relevant abstraction at one point and carrying it forward largely unchanged, exactly the kind of localized, linear structure the hypothesis predicts should exist for concepts a model actually uses.