claim
active
claim:logit-attribution-techniques-validated-in-transformers-may-not-transfer-directly-to-clts-since-each-feature-writes-multiple-vectorsLogit attribution techniques validated in Transformers may not transfer directly to CLTs since each feature writes multiple vectors
Caveat.
Source paper
extracted_fromNeighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Main conclusion.
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.
- Antra's foundational claim about how introspection arises computationally rather than from memorised text.
- Logit-minus-mean-logit loss was most effective for attribution, vs log-loss which was confusing to interpretfinding0.755Attribution loss selection.
- Transformers almost surely maintain input-injectivity throughout training, not just at initialisationhypothesis0.753Conjecture supported by Nikolaou et al. 2025 for last-token hidden states
- Claim formalizing the Anima Labs idea that transformers are effectively recurrent due to K/V stream.
- Dot product of summed decoder vector with unembedding vector, following Hanna et al.
- Speculative explanation.
- Headline result (Figure 1).
- Circuit simplification theory.