Zero-Shot Model Stitching

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

• Zero-shot model stitching without learning a stitching layer is feasible across different text models trained on different modalitiesfinding0.874
  Moschella et al. result cited as evidence of representational convergence across models
• Model Stitchingmethod0.814
  Technique to measure representational compatibility by integrating intermediate representations of one model into another
• Zero-Shot Generalizationconcept0.756
  Ability to predict correctly for stimulus-action pairs never previously experienced by inferring structural rules; key measure for TEM-t performance.
• Model stitching can use the behavioral null space of the source model when mapping to the target, making successful stitching insufficient evidence of representational similarityclaim0.753
  Formal analysis showing the theoretical limitation of model stitching as a similarity measure.
• zero-shot predictionconcept0.740
  Prediction without task-specific training; Evee achieves 0.991 AUROC on indels in zero-shot mode.
• Stitch (baseline)method0.738
  Baseline model stitching trained in a single behavioral direction without CL auxiliary loss, used for comparison with CLMAS.
• Latent Stitchmethod0.727
  Baseline method using a single orthogonal matrix trained to map source latents to target latents via CL auxiliary loss without behavioral objective.
• Zero-Shot Control Conditionconcept0.727
  Control omitting any induction and presenting only the final experiential query