finding
active
finding:sae-explains-0-86-variance-on-lmsys-0-84-on-pile-0-83-on-dolci-dpoSAE explains 0.86 variance on LMSYS, 0.84 on Pile, 0.83 on Dolci DPO
App. B.
Source paper
extracted_from(2026) · Leon Bergen · Usha Bhalla · Sidharth Baskaran · Max Loeffler +13
Neighborhood — ranked by edge-count
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.
- TFA explains similar variance as SAEs (Story 0.73, Web 0.79, Code 0.75) versus Pred-only 0.29–0.40.finding0.773Table 2.
- Basic SAE performance metrics.
- DPO on Dolci yields lower aggregate safety score than SFT checkpoint (e.g. Llama 3.1 8B 0.849 to 0.758)finding0.749Safeguard degradation.
- Shows VS enables LLMs to better approximate random behavior compared to direct prompting
- Table 1 on Qwen-2.5-7B holdout.
- Table 1.
- Qualitative and quantitative observation.