Interpreting Language Model Parameters

Can a model’s learned mechanisms be recovered directly from its parameters, without assuming that neurons, attention heads, or activation features are the correct units of analysis?

Goodfire’s adVersarial Parameter Decomposition (VPD) attempts to interpret language models by decomposing their actual weight matrices into simple rank-one subcomponents. Unlike activation-based methods such as sparse autoencoders and transcoders, VPD preserves the model’s original computational form: its components sum back to the model’s parameters and are trained so that only a small subset is necessary for any given input. Crucially, VPD adversarially searches for combinations of supposedly unimportant components whose removal would break the model, providing a stronger test of mechanistic faithfulness than ordinary random ablations.

Applied to a four-layer, 67M-parameter language model, VPD discovers interpretable mechanisms that frequently cut across conventional architectural boundaries—including attention computations distributed across multiple heads. These components can be assembled into attribution graphs for behaviors such as pronoun prediction and bracket closing, and the authors demonstrate a preliminary hand-written parameter edit that predictably changes emoticon completions. The result challenges the field’s focus on activation features: rather than asking only what representations a model contains, VPD offers a path toward understanding the parameter-level algorithms that operate on those representations—while leaving open whether the method can scale to frontier models or uniquely recover their “true” mechanisms.

Join us at CASI for discussion at 8 pm, and (optional) quiet reading from 7 pm.