Modified CL loss produces EMD along feature dimensions of 0.007±0.001 on synthetic 10-class dataset

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

• Modified CL loss achieves IIA of 0.9988±0.0005 on synthetic 10-class dataset training/test setsfinding0.823
  IIA for modified CL loss on synthetic dataset, comparable to behavioral DAS
• For small CL loss weights epsilon, IIA is maintained (potentially improved) while EMD decreases in Boundless DAS on a 7B LLMfinding0.785
  Empirical result showing the CL loss can reduce divergence without sacrificing interpretability accuracy
• Modified CL loss outperforms behavioral DAS loss in OOD transfer from dense to sparse class partitionfinding0.775
  Key practical utility result: CL loss improves generalization of alignment to out-of-distribution settings
• The modified CL loss is confined to a narrow set of simplistic settings and is not specific to pernicious divergenceconcept0.755
  Explicitly identified limitation of the proposed mitigation method
• Linear regression of OOD IIA on training EMD yields coefficient -0.3424, R^2=0.729, F(1,28)=75.28, p<.001finding0.739
  Statistical evidence that training divergence (EMD) predicts lower OOD intervention performance
• Modified CL Lossframework0.738
  Novel variant of CL loss introduced in this paper targeting only causal subspace dimensions to improve OOD performance
• Mean difference patching on Llama-3-8B layer 10 produces intervened EMD exceeding the natural-natural baselinefinding0.735
  Empirical demonstration that MDVP produces divergent representations in a real LLM
• DAS behavioral loss achieves IIA of 0.997±0.001 on synthetic 10-class dataset training/test setsfinding0.727
  IIA baseline for DAS behavioral loss on synthetic dataset