question
active
question:how-can-we-reliably-distinguish-causal-from-merely-correlated-features-when-probing-neural-networksHow can we reliably distinguish causal from merely correlated features when probing neural networks?
Probing open question.
Source paper
extracted_from(2025) · Lee Sharkey · Bilal I. Chughtai · Joshua Batson · Jack W. Lindsey +25
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Probe limitation.
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.
- Vision statement in the conclusion.
- Motivated by the finding that lexical entailment decomposes into word identities.
- Central motivating claim of the paper; supported by empirical comparisons showing RSA/CKA miss Markovian differences detectable by MAS.
- Motivates the bidirectional design of MAS over unidirectional model stitching.
- Load-bearing formulation of the paper's central argument
- Paper explicitly identifies this as a current gap requiring alternative experimental approaches
- Mechanism for null result.
- Practical question the paper attempts to answer in its conclusion