claim
active
claim:fact-retrieval-and-arithmetic-rely-on-specialized-narrowly-used-weight-directions-rather-than-general-purpose-mechanisms-regardless-of-memorizationFact retrieval and arithmetic rely on specialized narrowly-used weight directions rather than general purpose mechanisms, regardless of memorization
Abstract thesis.
Source paper
extracted_from(2025) · Merullo, Jack · Vatsavaya, Srihita · Bushnaq, Lucius · Lewis, Owen
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (3)
finding
- GSM8K example: edited model produces identical chain of thought but miscalculates 260 as 220supportsTargeted arithmetic deficit.
- MMLU-Pro Math drops from 23.4% to 15.8% under K-FAC while computer science stays 26.1% to 26.3%supportsArithmetic specifically hurt.
- Frequency modulates brittleness.
Hypotheses (1)
hypothesis
- Future direction.
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.
- Section 6.
- Identified as the exact computational operation that breaks truth direction generalization.
- We hypothesize that LLMs represent correctness of arithmetic expressions differently from factual statements.hypothesis0.776Core working hypothesis motivating the factual vs. arithmetic task split in the experimental design.
- From the cross-task generalization heatmaps in Appendix B.3.3.
- Load-bearing claim about the tractability of circuit analysis; central thesis of Claim 2
- Core novel interpretation.
- Truth directions emerge in earlier layers for factual tasks and later layers for arithmetic tasks.claim0.762Core empirical claim about the layer-dependence of truth direction emergence as a function of task type.