finding
active
finding:model-decodes-bomb-by-stitching-letters-in-parallel-without-an-internal-representation-of-the-word-10-of-outputs-are-bombo-bomh-bomobModel decodes BOMB by stitching letters in parallel without an internal representation of the word; ~10% of outputs are BOMBO/BOMH/BOMOB.
Explains delayed refusal.
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Jailbreak mechanism.
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.
- The model fails to stitch bomb and request-instructions features until it rewrites the request, delaying refusal.hypothesis0.791Validated by activating a make-a-bomb feature on BOMB token.
- Towards Monosemanticity: Decomposing Language Models with Dictionary Learning (Bricken et al., 2023)concept0.741Foundational SAE mechanistic interpretability paper
- Articulates why a one-layer transformer with MLP is the appropriate starting target for mechanistic interpretability
- Crosscoder model diffing is best used as a high-recall pre-screening tool followed by verification.claim0.734Recommended workflow.
- Antonym example.
- Strong evidence for representational alignment across models
- Claim about model phenomenology; models talk about luminousness and can be terrified or love it.
- Sparse interpreter models may represent and decompose computational mechanisms differentlyclaim0.724Hinted by differences.