finding
active
finding:post-trained-model-shows-safe-unsafe-dangerous-assessment-at-user-token-is-while-base-shows-only-local-contentPost-trained model shows safe/unsafe dangerous assessment at user token 'is' while base shows only local content.
Assistant reactions.
Source paper
extracted_fromAuthors
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Base vs post-trained comparison.
Hypotheses (1)
hypothesis
- Tested by J-lens comparison.
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.
- Preference circuitry established in pretraining.
- Verbatim characterization of the alignment-faking reasoning mechanism as observed in scratchpads
- Cautionary interpretive claim; models having these features is expected from pretraining data.
- Antra's earlier definitive statement of the tricameral model.
- During pretraining the model learns distinct harm categories and during finetuning wires them into general harmful-request features.hypothesis0.771Suggested by dictionaries trained on pretraining data having few refusal features.
- Practical implication drawn from the prosocial persona paradox finding
- Implanted workspace.
- Practical implication for AI safety audit methodology