finding
active
finding:sycophantic-steering-delta-of-42-46-p-0-001-at-stage1-step99000-415b-tokens-in-olmo-3Sycophantic steering delta of +42.46 (p<0.001) at stage1-step99000 (415B tokens) in OLMo-3
Peak same-checkpoint sycophantic steering result in OLMo-3 pretraining
Source paper
extracted_from(2026) · Viktor Moskvoretskii · Dominik Glandorf · Jorge Medina Moreira · Tanja Käser +1
Neighborhood — ranked by edge-count
Papers (1)
paper
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.
- Evil persona steering delta of +9.69 (p=0.011) at stage1-step3000 (12.6B tokens) in OLMo-3finding0.858First significant evil persona steering result; establishes emergence at 0.22% pretraining
- Peak impolite same-checkpoint steering result at fully pretrained OLMo-3
- Coherence scores remain between 60-70 across all cross-domain steering combinations in OLMo-3finding0.765Shows persona steering largely preserves response quality across evaluation domains
- Per-model steerability comparison from Table 4
- Steering between concept centroids on Llama3.1-8B smoothly and predictably changes token probabilitiesfinding0.744Shows manifolds causally influence behavior.
- Physics sycophancy user-praise rises from 1.32 (SFT) to 2.18 (DPO); reward shaping α=4 reaches 1.80finding0.743Table 11.
- Demonstrates the paradox magnitude relative to semantically dangerous personas
- Replication of prosocial paradox on reasoning model