claim
active
claim:the-representations-used-for-verbal-report-are-the-same-ones-that-govern-how-the-model-silently-reasonsThe representations used for verbal report are the same ones that govern how the model silently reasons.
Corroborated by reflection training.
Source paper
extracted_fromAuthors
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Causal mediation.
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.
- Traditional criterion Levin argues is wholly insufficient for evaluating sentience in unconventional agents.
- Addresses skeptical alternative that reports reflect only conversational content
- Decomposition result.
- Future work direction identified in conclusion for enabling reliable truth assessment methods.
- The workspace may be verbalizable because the model's output space and input space are verbal.claim0.774Speculative explanation.
- The model attends to the J-space on the user prompt when determining its introspective answer, so concepts there are reportable.hypothesis0.773Motivated by elephant observation.
- All models exhibit above-baseline representation of the think word when instructed to think about itfinding0.770In the intentional control experiment, all tested models show above-zero cosine similarity to the think word's concept vector.
- Instead of waiting for a model to verbalize eval awareness, we measure how likely it is to do so.quote0.770Core idea of the paper.