Source
Hase et al. (2023), LIT-1578, read in full in NOTE-1194, §6.
What is asserted
Two things. First, a conceptual separation: causal tracing measures where the information that restores a fact is carried in one forward pass; editing asks where a weight change will alter the fact the model expresses; and the answer to one need not be the answer to the other. Second, a mechanism for the gap: information about a fact accumulates across layers, and a single-layer edit, before or after that range, can override what was accumulated. Hence “many layers could store a fact, and it happens that some do”.
What the record holds toward it
- For the separation: the measured result, THEORY-1024. Edit success does not follow per-fact tracing, and the one setting where it does a little (Fact Forcing) is the one that edits on tracing’s own noised input.
- For the mechanism, from another direction: Makelov et al. (LIT-968, THEORY-282). Any one-dimensional fact patch in an MLP has a rank-one edit with the same output at that token, and their fact patches work through the null space of W_out. So an edit can produce the output without the layer holding the fact, which is one way an override could work.
- Against the mechanism, as the authors note: layers cannot be reordered without damage (Zhao et al., cited), so it “should matter where information enters the residual stream”. ROME cites the same paper for the opposite point, that the middle layers are interchangeable (THEORY-1011).
- Untested: the override itself. No experiment in the paper shows an edit outweighing a fact’s original contribution in later layers.
Why it stays Proposed
The separation is a reading of one negative result, with one localization family (denoising and zeroing) and one kind of edit. The mechanism is a hypothesis the authors offer and do not test. A localization that measures the clean computation might still predict edits, in which case the gap is the noised-run method’s, not a general one.
What would confirm or reject it
The promote_when above.