Positive Pull Loss (L_dist)

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

• L_retain Loss Termconcept0.732
  Regularization component of the composite loss that penalizes deviation from baseline model behavior on Alpaca instructions
• Prototype Contrast Loss (L_CE)method0.704
  Loss function pulling representations toward positive centroid and pushing away from negative centroid with angular margins
• The logarithm transformation (loss-scale balancing) consistently outperforms IMTL-L on NYUv2, Cityscapes, Office-31, Office-Home.finding0.702
  Comparison of loss-scale balancing with IMTL-L.
• Counterfactual Latent (CL) Lossframework0.701
  Auxiliary training objective from Grant (2025) that constrains intervened representations to remain near natural distribution
• SAE training loss (MSE + L1 penalty with decoder norm scaling)method0.699
  The objective function combining L2 reconstruction error and L1 penalty scaled by decoder norm, used to train the SAE.
• Modified CL loss outperforms behavioral DAS loss in OOD transfer from dense to sparse class partitionfinding0.694
  Key practical utility result: CL loss improves generalization of alignment to out-of-distribution settings
• Modified CL Lossframework0.692
  Novel variant of CL loss introduced in this paper targeting only causal subspace dimensions to improve OOD performance
• Counterfactual Latent (CL) Auxiliary Lossmethod0.691
  Auxiliary objective combining L2 and cosine losses against pre-recorded CL vectors to improve causal relevance when one model is causally inaccessible.