Contrastive Projection: Differencing Logit Lenses to Read Transformer Internals

A new arXiv paper introduces contrastive projection, a method for inspecting transformer hidden states that works by differencing logit lens outputs. The authors argue that a standard logit lens reading of a single intermediate-layer state is dominated by generic tokens a model would predict for nearly any input, making such readings unreliable. Subtracting these baseline predictions isolates the input-specific component of the internal state, offering a more trustworthy window into model internals.