Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
To overcome the generic token noise that typically dominates intermediate logit lens readouts, the authors developed Contrastive Projection, a method that differences the hidden states of closely matched prompts before unembedding them into vocabulary space. This training-free technique successfully traces sub-layer circuit dynamics—such as MLP-to-attention chains validated by activation patching—while exposing representations of entity retrieval and metaphor across multiple architectures. However, cross-seed controls demonstrate that while underlying computational distinctions remain invariant across initializations, their token-space projections are network-specific, exhibiting an average top-10 token overlap of just 0.08.
Subtracting logit lenses across matched prompt pairs reveals internal circuits without activation patching, yet exposes a crucial interpretability trap: identical computational distinctions map to almost entirely disjoint token spaces across different random seeds.
Reading a transformer's internal states in token space is easy to do and hard to trust: a logit lens on a single hidden state is dominated, at intermediate layers, by the generic tokens the model would predict for almost any input. We read the difference instead. Subtracting two closely matched prompts'hidden states and projecting through the unembedding cancels the shared component and surfaces what separates them, an operation equivalent to reading a RepE/ActAdd steering vector through a logit lens. Built into a training-free tracer that reads at every position, sub-layer, and head and averages over designed baselines, it traces a compound- noun MLP->attention chain in Phi-2, confirmed there by activation patching, with the same distinction recovered across three architectures by readout and probe rather than by patching; it reads what retrieval surfaces for real versus fictional entities, and reads metaphor as a set of domain-to-domain mappings rather than a single figurativity feature. A cross-seed control marks the boundary: across five networks differing only in initialization, the same distinction surfaces as almost entirely different tokens (top-10 overlap 0.08). What a computation looks like in token space is network-specific; the distinction it draws is not