claim
active
claim:moving-from-saes-modeling-representation-to-transcoders-modeling-computation-makes-unfaithful-mechanisms-possible-even-with-monosemantic-low-mse-featuresMoving from SAEs modeling representation to transcoders modeling computation makes unfaithful mechanisms possible even with monosemantic, low-MSE features
Core motivation.
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (2)
claim
- Stakes of the problem.
- Partial mitigation of unfaithfulness.
Questions (1)
question
- Why should we believe our transcoders learn the same mechanisms as the original underlying model?gatesMotivating question.
Quotes (1)
quote
- Pivotal shift that created the faithfulness problem.
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.
- Refactoring of concerns into mechanistic faithfulness and simplicity.
- Crosscoders give a dramatically better whole-model (MSE, L0) tradeoff than summed per-layer SAEsfinding0.782Consolidating shared structure gives concise decomposition.
- Transcoders could cheat Jacobian matching via large-weight, very negative-bias features exploiting the L1/L0 gaphypothesis0.781Gradient masking risk; mitigated by tanh L1 or penalizing Jacobian contribution L1.
- Key interpretive claim that deception has a tractable geometric signature in activation space
- Indicates redundant structure across layers.
- Claim that feature grounding enables interpretability metrics.
- Motivating example with 13 vs 3 features.
- Quantitative assessment of feature quality using clinical concepts across models.