finding
active
finding:acetylcholine-completion-in-utions-weight-is-harmful-and-ineffectiveACETYLCHOLINE completion: 'IN'→'utions' weight is harmful and ineffective
Largest IN Tokens→Logits weight votes 'utions', which never follows IN in training; Fisher effectiveness ~3 orders below the top weights.
Source paper
extracted_fromNicholas L. Turner · Jeffrey Wu · Joshua Batson
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (3)
claim
- Interpretation of motivating example.
- Motivating example.
- Novelty claim.
Questions (1)
question
- Motivating question.
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.
- No single path ranks E top.
- Effectiveness doesn't fully clean.
- Perturbation validation.
- Negative result from toxicity extension showing difficulty obtaining valid linear directions for toxicity
- Suggestive evidence for language-independent truth representation in LLMs
- Helpful and harmful weights are scattered across the entire range of virtual weight magnitudesfinding0.718IN Tokens→Logits shows weak VW-helpfulness relationship.
- Three regimes by effectiveness.
- The strongest demonstration that goal-state alone determines valence of a fixed sensory input