cad-contrastive-ensemble-same-model
IN premise — summaries/2026/08/24/shi-2024-context-aware-decoding-s2-method.md
Created 2026-08-25T02:58:35+00:00
In its standard form, CAD is a contrastive ensemble of two logit vectors from the same model (with-context and without-context), not a blend of two different models, though multi-model generalization is possible.
Summary
CAD's core mechanism is simpler than it might first appear: it compares the same model's output with and without a given context, using that internal contrast as the signal. This matters because it means the technique is measuring how much a single model's reasoning shifts when context changes, rather than trying to reconcile two separate models' answers, though the framework could in principle be extended across multiple models.