finding
active
finding:patching-an-unresolvable-pointer-at-layer-l-yields-no-counterfactual-answer-but-at-l-1-yields-itPatching an unresolvable pointer at layer ℓ yields no counterfactual answer, but at ℓ+1 yields it
Validates reflexive pointer.
Source paper
extracted_from(2025) · Yoav Gur-Arieh · Mor Geva · Atticus Geiger
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (2)
claim
- Central thesis.
- Rules out alternative explanation.
Questions (1)
question
- Confounder resolved in §3.4.
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.
- Synthetic theoretical example showing pernicious divergence via hidden pathway activation
- One of the most promising cases; approximately corresponds to the 2/3 layer of LLaMA3.1-8B.
- Features smeared across layers cannot be fully disentangled by SAE on a single residual stream.
- Activation patching of arithmetic intermediates is effective at layers matching J-lens emergence.finding0.757Causal accuracy.
- Geometric evidence for convergence to stable truth directions only for simpler tasks.
- Supported by the geometric transition visible in cosine similarity heatmaps for F0-F3.
- Ignition-like effect.
- Explains why time and sequence are essential for generated complexity.