Determine whether excluded SAE features encode omitted topic semantics

Determine whether lower-validation SAE features excluded from the MonoTM descriptor vocabulary encode meaningful, independent topic semantics that are absent from the validated feature set.

Background

MonoTM estimates document–topic mixtures using all active SAE features but constructs topic descriptors only from features whose labels pass an LLM-based interpretability threshold. The authors audit lower-validation features that are highly associated with inferred topics and compare them with validated same-topic descriptors in document-activation space.

The audit finds that many excluded features appear to be covered by validated descriptors or capture auxiliary cues such as reporting format. However, the analysis does not establish that every excluded feature is unimportant or fully understood. Consequently, whether some excluded features represent independent, meaningful topic dimensions remains unresolved, particularly for corpus-specific regularities, pragmatic cues, formatting patterns, or culturally specific concepts that are difficult to summarize with short labels.