claim
active
claim:ineffective-or-harmful-weights-are-interference-weights-the-model-carries-alongside-its-actual-circuits-not-lessons-learnedIneffective or harmful weights are interference weights the model carries alongside its actual circuits, not lessons learned
Interpretation of motivating example.
Source paper
extracted_fromNicholas L. Turner · Jeffrey Wu · Joshua Batson
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Largest IN Tokens→Logits weight votes 'utions', which never follows IN in training; Fisher effectiveness ~3 orders below the top weights.
Questions (2)
question
- Motivating question.
- Motivating example.
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.
- Discussion.
- Concluding statement.
- Novelty claim.
- Identifying interference weights is necessary but not sufficient for reading global circuitsclaim0.771Discussion.
- Three regimes by effectiveness.
- Helpful and harmful weights are scattered across the entire range of virtual weight magnitudesfinding0.766IN Tokens→Logits shows weak VW-helpfulness relationship.
- Logit weight contributions from a feature that arise due to superposition with other features, not from the feature's own causal role
- Explains why Head Cor does not fully dominate for safety-related persona amplification in MMLU/IFEval metrics