question
active
question:does-this-suggest-that-belief-is-localized-to-these-layers-and-could-we-causally-intervene-on-specific-neurons-to-have-predictable-impacts-on-model-behaviorDoes this suggest that belief is localized to these layers, and could we causally intervene on specific neurons to have predictable impacts on model behavior?
Open question.
Source paper
extracted_from(2025) · Eric Bigelow · Daniel Wurgaft · YingQiao Wang · Noah D. Goodman +3
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Suggests localization.
Claims (1)
claim
- Speculative interpretation of breakdown.
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.
- Finding (3).
- Open question about inter-agent communication beyond model-space assumption
- Fundamental theoretical claim motivating DAS, attributed to Smolensky/Rumelhart/McClelland.
- The motivating research question of the paper
- We hypothesize earlier-layer interventions allow more downstream computation to process and potentially correct the perturbationhypothesis0.768Post-hoc explanation for why steering at layer 33 rather than layer 50 produced better ESR behavior in Llama-3.3-70B
- Key limitation acknowledged by authors.
- Future work.
- Central claim motivating DAS over prior methods.