method
active
method:mechanistic-anomaly-detectionMechanistic anomaly detection
Flag actions taken for abnormal reasons.
Neighborhood — ranked by edge-count
Papers (1)
paper
Concepts (1)
concept
- White-box evaluationsextendsUsing internals to detect unsafe cognition.
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.
- A possible circuit that triggers when activations deviate from expected values, hypothesized to underlie noticing injected thoughts.
- Explanations derived from model internals that describe the biological mechanism of variant effect.
- Functionally significant directions in the residual stream that organize activations in mechanistic interpretability
- Possible explanation for why some concepts are more easily detected.
- Explaining activations without understanding mechanism.
- Graph-predicted indirect effects that do not exist in the underlying model.
- Task paradigm from prior work asking 'Did you detect an injected thought?' via YES/NO logit comparison; shown here to be confounded
- Idea that decomposed components must causally reflect the model's computation; enforced by adversarial ablation in VPD.