method
active
method:copyright-refusal-evaluation-protocolsCopyright refusal evaluation protocols
Evaluations on copyrighted and harmless prompts for GPT-OSS feature.
Neighborhood — ranked by edge-count
Papers (1)
paper
Artifacts (1)
artifact
- Claude 4.1 OpususesLLM judge and explainer used throughout.
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.
- The copyright feature controls refusal behavior rather than enabling faithful reproduction.claim0.765Disabling yields hallucination.
- A priori unknown behavior discovered.
- General method: select input distribution, define trait measure, sample LM responses, estimate trait score distributions.
- §3.3.2.
- Arditi et al. 2024 finding that refusal behavior is mediated by one direction in LLM activations; exemplar of single-direction causal results
- The percentage of harmful requests that a model refuses to answer, a common safety metric.
- Literary critical and Buddhist notion that individual authorship is not paramount; justifies openness to AI-generated texts.
- Single linear direction mediating refusal behavior in LLMs, shown by Arditi et al.; related to but distinct from the Assistant Axis