finding
active
finding:attribution-graphs-distinguish-faithful-sqrt-0-64-bullshitting-cos-23423-and-motivated-backward-reasoning-from-a-hinted-answerAttribution graphs distinguish faithful sqrt(0.64), bullshitting cos(23423), and motivated backward reasoning from a hinted answer.
Mechanistic discrimination of CoT faithfulness.
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Proof of concept.
Concepts (1)
concept
- Whether written reasoning reflects actual model mechanisms.
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.
- Quantifies method limitations.
- Varying the hint changes the motivated-reasoning CoT endpoint, while the faithful case always yields 8.finding0.792Tests bias susceptibility.
- Mechanistic proof of concept.
- Shows how VPD-identified subnetworks can be analyzed to reveal interpretable pathways of computation (e.g., gender signal routing, syntactic role detection).
- Structure of the computational pathway.
- Gradient-based technique using SAE features to estimate causal effects on completions; used to corroborate NLA findings.
- Attribution graph reveals a pathway that detects the verb 'lost' and upweights object pronounsfinding0.776Second component of the subnetwork for 'her', complementing the femaleness signal.
- Cited regarding possibility of encoding misaligned reasoning in benign chains-of-thought