finding
active
finding:sae-reconstructions-on-llama-3-8b-layer-25-produce-intervened-emd-exceeding-the-natural-natural-baselineSAE reconstructions on Llama-3-8B layer 25 produce intervened EMD exceeding the natural-natural baseline
Empirical demonstration that SAE projections produce divergent representations in a real LLM
Source paper
extracted_from(2025) · Satchel Grant · Simon Jerome Han · Alexa R. Tartaglini · Christopher Potts
Neighborhood — ranked by edge-count
Claims (1)
claim
- Core empirical claim of the paper supported by both theoretical proof and empirical demonstration
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.
- Empirical demonstration that MDVP produces divergent representations in a real LLM
- Seen as in other SAEs.
- App. E.
- Core result of Experiment 2: deception feature suppression sharply increases experience claims
- Empirical demonstration that DAS interventions produce divergent representations
- Moderation statistics.
- Quantifies SAE reconstruction damage.
- Features act as local patch detectors.