finding
active
finding:sae-reconstructions-decreased-gpt-2-small-performance-by-10-task-specific-data-and-40-full-distributionSAE reconstructions decreased GPT-2 small performance by 10% (task-specific data) and 40% (full distribution)
Quantifies SAE reconstruction damage.
Source paper
extracted_from(2025) · Lee Sharkey · Bilal I. Chughtai · Joshua Batson · Jack W. Lindsey +25
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Core assessment of SDL.
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.
- SAE reconstruction error direction and norm are largely linearly predictable from the activation vectorfinding0.798Errors are structured not random.
- Evidence in real models.
- Limitation statement.
- Basic SAE performance metrics.
- Fig 2 right.
- Assistant-only SAEs reconstruct user turns 1.5–3× worse; both-roles training cuts user-turn nMSE 53–63%finding0.766Appendix G.
- RM features at rank 181, outside top 100.
- Open question in limitations.