Representational divergence (as measured by EMD) can predict lower out-of-distribution intervention performance

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Representational Divergenceconcept0.799
  Core phenomenon studied: when causal interventions shift internal representations away from the natural distribution
• Divergent representations are a common, if not likely, outcome of causal interventions across a wide range of methodsclaim0.780
  Core empirical claim of the paper supported by both theoretical proof and empirical demonstration
• Trainable intervention (DAS) finds sparser gender representations than linear probing, suggesting probing overestimates causal coverageclaim0.763
  Interpretive claim from Case Study II about the distinction between correlational probes and causal interventions
• DAS behavioral loss produces EMD along feature dimensions of 0.032±0.003 on synthetic 10-class datasetfinding0.755
  Quantitative baseline for divergence using behavioral DAS loss on synthetic dataset
• EI and normalized EI could serve as a unified metric for out-of-distribution generalization.claim0.748
  Conjecture that maximizing EI yields causal representations invariant to distribution shifts.
• If EI maximization is used as a regularization in representation learning, then OOD generalization will improve beyond current invariant risk minimization methods.hypothesis0.743
  Proposed conjecture in §4.3.1.
• An intervention benign at context v4<0.75 produces a class-C behavioral flip at 0.75<v4<1, demonstrating dormant behavioral changes from latent divergencefinding0.742
  Synthetic example showing an intervention that appears safe in tested contexts but causes behavior changes in others
• Representation engineering successfully quantifies deception via high-accuracy steering vectors, establishing it as a measurable property of model representationsclaim0.742
  Key interpretive claim that deception has a tractable geometric signature in activation space