Kramár, J., Lieberum, T., Shah, R. & Nanda, N. (2024). AtP*: An efficient and scalable method for localizing LLM behaviour to components. arXiv:2403.00745.
The method it studies is from Nanda, N. (2023). Attribution patching: Activation patching at industrial scale. Blog post, 4 February 2023, https://www.neelnanda.io/mechanistic-interpretability/attribution-patching (cited in the paper as “Nanda, 2022”, the year of the underlying work).
Read in full on 2026-10-26 (NOTE-1190). Two accounts were drawn from it: THEORY-996 and THEORY-1000.
Why this is here
The owner asked on 2026-10-26 how the record stood on neural circuit theory; the answer found it held almost none of the core circuits literature, and the owner said “go ahead and register and read those”. Attribution patching is the cheap estimate that made circuit search scale to large models; its edge version with integrated gradients is how LIT-1539 (THEORY-997) and MIB’s circuit track (LIT-1550) find circuits, so the record held results built on it without the method or its failure modes. This paper is registered for it, rather than Nanda’s post, because it tests the method against ground truth.
Note of 2026-10-27: the method’s origin, now held
The version-1 history above says Nanda’s post “Attribution Patching:
Activation Patching At Industrial Scale” “is not registered separately”,
and “Why this is here” says this paper was registered “rather than Nanda’s
post”. Both were true when written. The post was registered and read on
2026-10-27, at the owner’s request, as LIT-1581 (NOTE-1201), and
this entry now names it under extends. The estimate itself, re-derived
and checked, is THEORY-1023. The post predicted the saturated-attention
failure this paper measures (THEORY-996), but treated cancelling effects
in backup heads as a source of false positives, not the false negatives
this paper found.