claim
active
claim:the-model-forms-a-general-harmful-requests-abstraction-during-finetuning-stitched-from-pretraining-harm-conceptsThe model forms a general 'harmful requests' abstraction during finetuning stitched from pretraining harm concepts.
Refusals.
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Refusal circuit interventions.
Hypotheses (1)
hypothesis
- Suggested by dictionaries trained on pretraining data having few refusal features.
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.
- Verbatim characterization of the alignment-faking reasoning mechanism as observed in scratchpads
- Supported by further finetuning and causal interventions.
- Figure 12 left and right show accuracy on harmful/ethical identification and 9-way classification.
- Jailbreak mechanism.
- Evidence for the evil persona as a privileged basin supporting Hypothesis 3
- The paper's central mechanistic explanation of why narrow fine-tuning causes broad misalignment
- Demonstrates emergent re-alignment is achievable with minimal data from same domain
- The model fails to stitch bomb and request-instructions features until it rewrites the request, delaying refusal.hypothesis0.771Validated by activating a make-a-bomb feature on BOMB token.