claim
active
claim:the-model-represents-and-uses-the-intermediate-concept-before-the-answer-has-been-computedThe model represents and uses the intermediate concept before the answer has been computed.
Rules out smuggled answer confound.
Source paper
extracted_fromAuthors
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Intermediate swap takes effect a median of about 17 percent of depth earlier than answer swap.supportsRules out smuggled answer.
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 translates concepts to a common universal mental language in intermediate activations.claim0.833Discussion abstraction.
- Antonym example.
- Key consequence: GPT's power comes from simulating something contingent.
- Key jailbreak insight.
- Core definitional quote for performative chain-of-thought
- Self-explanation vs circuit.
- Mechanistic insight surfaced by NLA explanations and validated through independent causal attribution method.
- All models exhibit above-baseline representation of the think word when instructed to think about itfinding0.780In the intentional control experiment, all tested models show above-zero cosine similarity to the think word's concept vector.