nucleation

LIT-1550: MIB: A Mechanistic Interpretability Benchmark

1 min read

Properties5
StatusActive — read 2026-10-26 (NOTE-1169) in full, from the arXiv copy (v2, the ICML 2025 version); worth reading for its fixed-counterfactual design of two localization tracks, for the finding that public sparse-autoencoder features were worse units than neurons for known causal variables in every cell tested (THEORY-968), and for the carry-the-one case, where a representation that looked like the variable was the correlated output digit. Read it knowing that the best causal-variable method, DAS, is supervised by the causal model it is scored against, that the SAE comparison is on residual streams of two models under one selection method, and that the method rankings are interpretability practice.
Version2
Tagsrepresentation-learning, causality, anthology-candidate
Cited as “paper” byNOTE-1169 — Mueller, Geiger, Wiegreffe, Arad et al. — MIB, a mechanistic interpretability benchmark
Cited as “source” byTHEORY-968 — In the Mechanistic Interpretability Benchmark's causal-variable localization baselines on Gemma-2 2B and Llama-3.1 8B, features of the public Gemma Scope and Llama Scope sparse autoencoders, selected by a learned binary mask, never gave a higher mean interchange-intervention accuracy than the same mask over the model's own hidden dimensions, in all sixteen task–variable cells and by 1 to 14 points, while distributed alignment search, trained against the causal model it is scored on, gave the highest mean in fifteen of them

Graph View

Mueller, A., Geiger, A., Wiegreffe, S., Arad, D., et al. (2025). MIB: A mechanistic interpretability benchmark. ICML 2025. arXiv:2504.13151.

Read in full on 2026-10-26 (NOTE-1169). One account was drawn from it: THEORY-968.

Why this is here

Cited at T24 as a warning that a sophisticated feature detector is not automatically a better causal explanation. That is CLAIM-454’s point in the transformer setting: a feature recovered from activity is not yet the variable that does the work. It bears on QUESTION-275. The record’s neighbouring holdings are Makelov et al. (LIT-968) and Wu et al.’s reply (LIT-961) on interchange interventions, LIT-987, THEORY-243 and THEORY-802 on basis-relative alignments, and THEORY-504 on SAE fidelity.

Backlinks