finding
active
finding:turn-averaged-sae-trains-in-1-5-hours-on-4-h100-with-1-58m-samples-versus-266m-tokens-per-tokenTurn-averaged SAE trains in ~1.5 hours on 4×H100 with ~1.58M samples versus 266M tokens per-token
Training efficiency.
Source paper
extracted_from(2026) · Kevin Der · Harish Kamath · Thompson, Ben
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Abstract.
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.
- Table 1.
- Computational cost.
- Table 1.
- Table 1 on Qwen-2.5-7B holdout.
- Long-context results.
- Basic SAE performance metrics.
- Broadly, turn-averaged SAEs make interpretability techniques practical at long context lengths.quote0.749Abstract thesis.
- Paper's main proposal: SAE trained on the mean activation across a turn.