Mischief surveys linear representation hypothesis

The author synthesizes recent research on the Linear Representation Hypothesis, which posits that neural networks represent concepts through directional 1D subspaces of their activation space. Examining measurement and intervention hypotheses, the text explores empirical findings, theoretical caveats, and open questions regarding why linear representations appear more robust in intermediate layers than final layers. Additionally, it highlights unexpected behaviors in model steering vectors that challenge strict interpretations of the hypothesis. Open article →