finding
active
finding:replacing-language-detection-features-changes-output-language-while-preserving-operation-and-operandReplacing language-detection features changes output language while preserving operation and operand.
Language component editable independently.
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Multilingual circuits.
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.
- Selectivity.
- Case study demonstrating NLA ability to surface root causes of model misbehavior; corroborated by training data inspection.
- Introduction, motivating the proposal.
- NLAs revealed unverbalized language processing in Opus 4.6 that led to discovery of malformed SFT training data.
- English is mechanistically the default output.
- Follow-up on empirical grounding; answered 'no one looked yet'.
- Operation component is language-independent.
- Feature interventions generated meaningful model output changes while minimizing performance degradationfinding0.746Listed achievement in TL;DR.