ConceptComposition
Composition
Drawing on research in compositional generalization, it refers to a model's ability to chain separately learned associations into combinations not directly present in the training data, giving coupling relationships a transitive property.
Timeline (last 90 days)
Oct 19
Composite coupling breaks the chain of provenance, causing conventional investigation methods based on tracing back fine-tuning data and the target behavior to fail, because the true source may be hidden in the intersection of two seemingly unrelated documents.
Unverified50%