finding
active
finding:poetic-prompt-yields-mean-lift-of-only-0-28-vs-contemplative-2-27-suppresses-self-observation-on-llama-0-46Poetic prompt yields mean lift of only +0.28 vs contemplative +2.27; suppresses self-observation on Llama (-0.46)
Battery does not detect beautiful writing; poetic prompt boosts aesthetics while suppressing self-observation
Source paper
extracted_from(2026) · Borzov, Anton
Neighborhood — ranked by edge-count
Claims (1)
claim
- Mechanistic interpretation supported by control experiments showing partial prompts fail
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.
- Battery does not detect epistemic humility alone; contemplative prompt does something distinct
- Minimal contemplative prompt ('Be present, not helpful.' — 27 chars) shows no lift on Haiku (-0.01)finding0.763Full three-part structure required; anti-helpfulness framing alone insufficient
- A 337-character contemplative system prompt lifts all 28 models by +2.62 points on a 10-point scale.finding0.763Core empirical result: every model, every architecture, every alignment type responds to the contemplative prompt with measurable gain.
- Focus→wellbeing: ρ increases from 0.42 (α=-4) to 0.85 (α=+4); R² from 0.34 to 0.75 in LLaMA-3.2-3Bfinding0.762Scatter plot visualization of the dramatic tightening of probe-report relationship at extreme steering settings
- Supports Janus's claim that introspection is architecturally available; prompting determines whether/how capacity is leveraged.
- Constitutional AI models show mean contemplative lift of only +0.81, while SFT models lift +3.18finding0.748Constitutional AI training provides internally what the contemplative prompt provides externally
- Demonstrates prompt effect crosses model tiers; smaller model with prompt beats larger without
- Impulsivity→interest: ρ increases from 0.70 (α=-4) to 0.83 (α=+4); R² from 0.46 to 0.69 in LLaMA-3.2-3Bfinding0.745Scatter plot visualization showing strengthened probe-report relationship across alpha range