finding
active
finding:clamping-code-error-feature-to-large-negative-activation-causes-model-to-output-correct-result-despite-bug-in-code-and-in-one-case-rewrite-code-without-bugClamping code error feature to large negative activation causes model to output correct result despite bug in code, and in one case rewrite code without bug.
Suppressing the feature makes the model ignore bugs.
Source paper
extracted_from(2024) ·
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.
- Causal effect: feature induces perception of bugs.
- Causal effect: activates generation of security bugs.
- noted as a possible confound
- Feature manipulation alters persona.
- Causal effect showing the feature governs computation.
- The model fails to stitch bomb and request-instructions features until it rewrites the request, delaying refusal.hypothesis0.747Validated by activating a make-a-bomb feature on BOMB token.
- Refusal mediated by attention.
- Shows a general code error detector beyond simple typo detection.