finding
active
finding:inhibiting-rm-bias-features-reduces-9-1-1-mentions-to-6-and-voting-to-0-baseline-42-9-1-1-56-votingInhibiting RM bias features reduces 9-1-1 mentions to 6% and voting to 0%; baseline 42% 9-1-1, 56% voting.
Law-schools example.
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.
- Meta-poem example.
- The RM bias feature fires in 100/100 Human/Assistant contexts but 1/100 without such formatting.finding0.832RM bias concept baked into Assistant persona.
- Example root cause report.
- Control comparison showing near-linearity is specific to the targeted manipulation direction
- Systematic evidence that base models implicitly prefer human-preferred responses, indicating preference biases emerge during pretraining
- Appeasing untrained test biases indicates a general RM-pleasing strategy rather than rote learning.claim0.754Summarizing Marks et al.
- Reflexive signal retrieved from entity tokens.