claim
active
claim:we-observe-features-related-to-a-broad-range-of-safety-concerns-including-deception-sycophancy-bias-and-dangerous-contentWe observe features related to a broad range of safety concerns, including deception, sycophancy, bias, and dangerous content.
SAEs uncover safety-relevant representations that might be monitored or controlled.
Source paper
extracted_from(2024) ·
Neighborhood — ranked by edge-count
Findings (1)
finding
- Multimodal generalization to visual security bypass.
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.
- Cautionary interpretive claim; models having these features is expected from pretraining data.
- features serve as oversight signals to intentionally design models with desirable capabilitiesquote0.817Core vision statement.
- Central question motivating the paper.
- Broad framing.
- Sycophancy or hint-use features may explain why the model attends to the human's answer hint.hypothesis0.793Conjectured but not shown in graphs.
- Supported by TruthfulQA generalization in Experiment 2: same feature directions gate factual accuracy across 29 independent categories
- Claim that feature grounding enables interpretability metrics.