method
active
method:exam-story-steering-experimentExam story steering experiment
Steered story generation reveals cheating themes.
Neighborhood — ranked by edge-count
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.
- Steering along probe makes exam stories feature cheating, shortcuts, and SWE/evaluation themesfinding0.783Behavioral effect of probe.
- Tests whether deception- and roleplay-related features causally gate consciousness self-reports in LLaMA 3.3 70B
- Used to validate graph-predicted mechanisms in the original model.
- Steers Qwen 3.8 27B along probe and measures ZZZ honeypot invocation.
- Demonstration of feature steering on R1.
- Mechanism claim supported by transcript analysis and the fact that the steering vector was extracted from a model that never writes type hints.
- Control condition with steering disabled to confirm self-correction is induced by steering, not spontaneous
- Demonstration of steering reasoning length.