finding
active
finding:injecting-rabbit-green-planned-word-features-in-25-poems-makes-the-model-end-its-line-with-the-injected-word-70-of-the-timeInjecting rabbit/green planned-word features in 25 poems makes the model end its line with the injected word 70% of the time.
Quantitative planning intervention result.
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Planning in poems.
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.
- Illustrative finding from Lindsey et al. 2025 / Hanna and Ameisen 2026 showing intention-like representations in attention streams
- May reflect CLT gaps or non-planning.
- Tested via interventions at the newline planning site.
- Claude Opus 4.6 represents a plan to end a couplet with 'rabbit' before outputting the rhyming line.finding0.765Demonstrates causal relationship between NLA explanations and model outputs via steering with edited explanations.
- Suppressing rabbit features changes the line to end with habit and alters intermediate transition.finding0.748Planned words shape sentence structure.
- Planned words are represented by ordinary word features rather than planning-specific features.claim0.745Model thinks about planned words like read words.
- Evidence that NLA explanations bear causal relationship to model outputs; demonstrates validity of extracted representations.
- Explanation for the 'silent' thought phenomenon.