finding
active
finding:dim-based-ablation-direction-for-toxicity-on-toxigen-produced-unintelligible-output-no-valid-concept-cone-foundDIM-based ablation direction for toxicity on ToxiGen produced unintelligible output; no valid concept cone found
Negative result from toxicity extension showing difficulty obtaining valid linear directions for toxicity
Source paper
extracted_from(2025) · Kevin Shengyang Yu · Vaidehi Bulusu · Oscar Yasunaga · Lau, Clayton +4
Neighborhood — ranked by edge-count
Papers (1)
paper
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.
- Refusal circuit interventions.
- Ablation depends on t_entity.
- Justifies SCR/TPP.
- Gradient-based attribution approximates ablation impact, enabling fast search for causally important features.
- Largest IN Tokens→Logits weight votes 'utions', which never follows IN in training; Fisher effectiveness ~3 orders below the top weights.
- First falsifiable prediction of the thesis, testable in AI systems via mechanistic interpretability
- Controls for the specificity of Style Modulation Head ablation effects
- Produces the less-safe variant with refusal direction removed.