hypothesis
active
hypothesis:information-from-future-layers-may-be-needed-to-determine-causal-importance-so-layerwise-gates-may-over-activateInformation from future layers may be needed to determine causal importance, so layerwise gates may over-activate
Open issue.
Source paper
extracted_from(2025) · Lucius Bushnaq · Dan Braun · Lee Sharkey
Neighborhood — ranked by edge-count
Papers (1)
paper
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.
- Normative vision for how the circuits agenda could resolve the pre-paradigmatic state of interpretability
- Central thesis of the paper
- Explains why time and sequence are essential for generated complexity.
- Truth directions emerge in earlier layers for factual tasks and later layers for arithmetic tasks.claim0.733Core empirical claim about the layer-dependence of truth direction emergence as a function of task type.
- Methodological critique of prior work that fixed a single layer for truth probing.
- Interpretive claim based on circuit analysis across experiments
- Visual geometric evidence for the fundamental entanglement of true/false activations in harder tasks.
- Quantifies extreme redundancy in transformer routing; supports claim that introspection and interference patterns are architecturally permitted.