claim
active
claim:probes-detect-correlations-rather-than-causal-variables-and-should-be-used-only-to-generate-hypothesesProbes detect correlations rather than causal variables and should be used only to generate hypotheses
Probe limitation.
Source paper
extracted_from(2025) · Lee Sharkey · Bilal I. Chughtai · Joshua Batson · Jack W. Lindsey +25
Neighborhood — ranked by edge-count
Papers (1)
paper
Questions (1)
question
- How can we reliably distinguish causal from merely correlated features when probing neural networks?gatesProbing open question.
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.
- Key interpretive claim from Case Study II distinguishing probe accuracy from causal relevance
- Key methodological claim: MM probes are both competitive in accuracy and superior in causal influence
- Question raised by the discrepancy between DAS IIA and linear probe accuracy in Case Study II
- Convergent validity logic applied to LLM interpretability; probes validate self-reports and vice versa
- Caveat on probe interpretation; does not negate the introspection result but affects interpretation of the target variable
- Authors' central interpretive assertion that their method meaningfully mitigates unwanted behaviors.
- Shows that truth representations are not reducible to text probability representations
- Joint use improves TPR.