claim
active
claim:this-setting-multi-token-autoregressive-generation-on-natural-red-team-prompts-is-closer-to-realistic-safety-evaluations-than-wu-hilton-sThis setting (multi-token autoregressive generation on natural red-team prompts) is closer to realistic safety evaluations than Wu & Hilton's
Positions the work.
Neighborhood — ranked by edge-count
Papers (1)
paper
- Predicting Rare LLM Failures with 30× Fewer Rolloutsmentionssupports
Findings (1)
finding
- Prior result for 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.
- Post-trained models repurpose user token activations to represent the Assistant's thoughts.hypothesis0.776Tested by J-lens comparison.
- Summary finding of the full behavioral sweep
- Key observation that SP rankings are preserved cross-architecturally while AS is not
- Unsupervised approach may be sufficient for early detection of misaligned persona latents without knowing the misaligned behavior in advance
- Stated generalization, untested here.
- Justifies the choice of response tokens for persona vector extraction in the pipeline
- §5.5.