Coverage Audits for LLM Interpretability Claims
Coverage Audits for LLM Interpretability Claims
A Mercury methods note arguing that sparse probing, activation patching, subspace intervention, and pruning claims should report a coverage audit before making broad localization claims. The audit uses 19 fullgrid manifests and shows a 1/7 pass rate for older sparse or coarse protocols versus 12/12 for per-channel quantile dense runs.
Author ORCID: https://orcid.org/0009-0006-6816-9891. Draft prepared from local Mercury fullgrid manifests under observations/2026-05-29-fullgrid-sync/.
mechanistic interpretability, model evaluation, LLM interpretability, Mercury, measurement validity, sparse probing, interpretability claims, dense activation measurement, model pruning, sampling bias, coverage audits, cell coverage, fullgrid, activation patching, transformer internals, neural atlas, activation coverage, coverage failure, result governance, reproducibility
mechanistic interpretability, model evaluation, LLM interpretability, Mercury, measurement validity, sparse probing, interpretability claims, dense activation measurement, model pruning, sampling bias, coverage audits, cell coverage, fullgrid, activation patching, transformer internals, neural atlas, activation coverage, coverage failure, result governance, reproducibility
citations The total number of citations received by all articles of the researcher of interest.0 popularity This indicator reflects the "current" impact/attention (the "hype") of an article in the research community at large, based on the underlying citation network.Average influence This indicator reflects the overall/total impact of an article in the research community at large, based on the underlying citation network (diachronically).Average impulse This indicator reflects the initial momentum of an article directly after its publication, based on the underlying citation network.Average
Citations provided by BIP!Popularity provided by BIP!