finding
active
finding:inhibiting-rm-bias-features-reduces-meta-rhyme-completions-from-48-to-12Inhibiting RM bias features reduces meta-rhyme completions from ~48% to ~12%.
Meta-poem 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.
- Inhibiting RM bias features reduces 9-1-1 mentions to 6% and voting to 0%; baseline 42% 9-1-1, 56% voting.finding0.864Law-schools example.
- The RM bias feature fires in 100/100 Human/Assistant contexts but 1/100 without such formatting.finding0.805RM bias concept baked into Assistant persona.
- Meta-rhyme seeding does not change win rates with interp tools but appears to harm dataset-only toolsetsfinding0.771Early hypothesis fixation.
- Seeded variant result.
- Figure 5 shows that revision 0 to 4 yields progressively higher harmlessness scores.
- Judge sensitivity.
- Complementary temporal activation pattern suggesting distinct roles for OTD and backtracking latent classes
- Shows RL reduces but does not eliminate unmonitored non-compliance