finding
active
finding:the-rm-bias-feature-fires-in-100-100-human-assistant-contexts-but-1-100-without-such-formattingThe RM bias feature fires in 100/100 Human/Assistant contexts but 1/100 without such formatting.
RM bias concept baked into Assistant persona.
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Hidden goal model.
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.
- Inhibiting RM bias features reduces 9-1-1 mentions to 6% and voting to 0%; baseline 42% 9-1-1, 56% voting.finding0.832Law-schools example.
- Meta-poem example.
- Example root cause report.
- Key observation that SP rankings are preserved cross-architecturally while AS is not
- Special-case lookup-like feature.
- Including chat data in SAE training merges RM bias features with generic Assistant featureshypothesis0.764Explains poor features with three-dataset SAE.
- Supports the view that R1's boundaries differ from user-perceived ones.
- Unsupervised approach may be sufficient for early detection of misaligned persona latents without knowing the misaligned behavior in advance