claim
active
claim:supervised-metrics-are-limited-by-ground-truth-availability-and-cover-only-a-fraction-of-latentsSupervised metrics are limited by ground-truth availability and cover only a fraction of latents
Limitation.
Source paper
extracted_from(2025) · Adam Karvonen · Can Rager · Johnny Lin · Curt Tigges +11
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (3)
finding
- Probing limited.
- Unlearning limited by model capability.
- Margin of advantage.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Smaller models' weaker concept representations explain why supervised metrics differ on Pythia-160Mclaim0.785Interpretation of model-scale effects.
- what metrics can reliably tell us if dictionary learning has successfully extracted high quality features?question0.775Central methodological gap: current metrics (loss, density histograms, manual inspection) are inadequate
- §5.5.
- Limitation.
- Shows that truth representations are not reducible to text probability representations
- Central critique of prior evaluation: whole-response scoring hides individual OOC sentences
- Key interpretive claim from Case Study II distinguishing probe accuracy from causal relevance
- Authors argue their metrics capture a distinct behavioral facet and could detect residual misalignment when standard evaluations indicate improvement