concept
active
concept:secret-knowledge-in-fine-tuned-modelsSecret knowledge in fine-tuned models
Hidden knowledge models are trained not to reveal, target of auditing.
Neighborhood — ranked by edge-count
Papers (1)
paper
Methods (1)
method
- Auditing task: discover the secret word a model gives hints for but never states.
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.
- Open question raised in Discussion noting capability changes could be due to character training itself, LoRA, or other factors
- Demonstrates emergent re-alignment is achievable with minimal data from same domain
- Parameter updates that reduce mismatch dr; another anchoring variant in UCCT.
- Different fine-tuning domains produce qualitatively distinct misalignment profiles attributable to different data generation processes
- Fine-tuning LLMs on insecure code dataset from Betley et al. to induce emergent misalignment
- Conclusion from Experiment 3 and HH intent analysis.
- The patient, hand-guided adjustment of shape and dimension to each unique condition in a building; requires materials that make it economical and easy.
- Matched control fine-tuning on secure code dataset to isolate misalignment-specific effects