Every Component is a Lookup: Token Attribution and Composition from a Single Decomposition

Abstract:Mechanistic interpretability of transformers requires identifying not just which components matter but how they compose into the computational route that produced a prediction. Both attention and MLP follow a shared key-value template $\phi(S)U$. We exploit this structure to develop Unpack, a backward recursion that decomposes credit through both sublayers, producing interaction strengths between any two components, named end-to-end paths with K/Q/V composition labels, and per-token attribution, all from a single forward pass, without intervention, gradients, or auxiliary training. The interaction scores are causally grounded: across the Pythia-deduped family from 160M to 6.9B parameters, a component's score predicts the perplexity increase when its communication is ablated (within-layer Spearman $\rho = 0.72$ to $0.96$). The composition paths surface all three connections of the indirect-object-identification circuit of Wang et al. (2023), including the mode-specific routing of each: rerooting at the Name Mover heads, S-Inhibition is the strongest query-side input and falls to rank 10 on the key side, a distinction no single per-token or per-component score can express. The same procedure applied to the greater-than circuit of Hanna et al. (2023), a differently shaped circuit, places the named connections among the top contributors, once layer-0 writers, which carry large credit whatever they feed, are set aside. The decomposition reads out contribution under the realized computation; it complements, rather than performs, causal circuit discovery. The per-token readout is faithful under input perturbation, on par with dedicated attribution methods, and distinguishes circuit mechanism from surface identity: two occurrences of the same name receive radically different credit when only one drives the circuit. Code is available at this https URL.

Machine Learning (cs.LG); Artificial Intelligence (cs.AI)

arXiv:2605.23393 [cs.LG]

(or arXiv:2605.23393v2 [cs.LG] for this version)

https://doi.org/10.48550/arXiv.2605.23393 Focus to learn more arXiv-issued DOI via DataCite

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.