finding
active
finding:multilingual-say-large-features-have-stronger-direct-effects-on-english-outputs-and-english-quote-features-show-double-inhibitionMultilingual say-large features have stronger direct effects on English outputs and English quote features show double inhibition.
English is mechanistically the default output.
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Resolves Schut vs Wendler tension.
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.
- Features respond to concepts across languages and in images, not just text.
- Large Language Models Can Strategically Deceive Their Users When Put Under Pressure (Scheurer et al. 2023)concept0.756GPT-4 engaging in insider trading and denying it; related work on strategic deception
- Replacing language-detection features changes output language while preserving operation and operand.finding0.754Language component editable independently.
- Steering validation for antonym mechanism.
- Large language models develop surprisingly coherent yet often rigid internal preferences as they scalefinding0.751Mazeika et al. finding reinforcing the need for emptiness-based flexible value architectures
- Motivation for using sparsity-based dictionary learning on language models
- Operation component is language-independent.
- Analogy between LLM incoherence and schizophrenia symptoms