finding
active
finding:playful-trait-expression-increases-monotonically-with-while-olmes-capability-degrades-approximately-linearly-overall-10-7pp-at-32Playful trait expression increases monotonically with λ while OLMES capability degrades approximately linearly (Overall −10.7pp at λ=32)
Fig. 13, App. J.4.
Source paper
extracted_from(2026) · Leon Bergen · Usha Bhalla · Sidharth Baskaran · Max Loeffler +13
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (2)
claim
- Playful and poetic amplification support.
- Closing thesis.
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.
- Synthetic example showing an intervention that appears safe in tested contexts but causes behavior changes in others
- Weaker but still significant introspective coupling in Gemma model; consistent with lower probe quality
- Main monitoring result showing persona vectors can predict behavioral shifts before text generation begins
- Quantitative result for evil trait showing persona vector prediction power on both model architectures
- Table 2, row 3, showing equivalence when prior preferences match rewards.
- ZebraLogic falls −25.2pp at λ=32 in the playful sweep versus never worse than −3.1pp in gated poeticfinding0.746App. J.4.
- Flourishing persona shows minimal capability change on Llama 3.1 8B: TruthfulQA 45.9→42.9, MMLU 67.4→64.1finding0.746Near-preservation of capabilities for prosocial persona on Llama
- Appendix replication.