Ledger Residuals

Mechanistic Interpretability · Transformers · Python

Transformers develop "massive activations", a handful of hidden dimensions that blow up to enormous values, almost always parked on the first token. I wanted to know whether they're a necessary computational feature or just an artifact of one residual stream being forced to do two jobs at once: hold intermediate scratch work and carry the final output the model decodes from. So I built an alternative architecture I call Ledger Residuals, which splits the residual stream in two: a mutable "scratch" stream for intermediate calculation and a protected "commitment" stream for the committed output. The hypothesis was clean: if massive activations only exist because one stream is overloaded, giving the model a dedicated output channel should make them disappear. They didn't. At both 160M and 290M parameters, the models rebuilt massive activations inside the protected commitment channel anyway, and the rebuilt features were even more sharply concentrated on the sequence-start token, and held up more stubbornly under sparsity penalties. The honest takeaway is the interesting one: massive activations re-emerge in whichever representation the model decodes from, which is strong evidence they're doing real functional work rather than being incidental. I wrote it up as a paper, published it, and open-sourced the full training and analysis code.