cad-knowledge-conflict-greatest-gain

IN premise — summaries/2026/08/24/shi-2024-context-aware-decoding-s0-abstract.md

Created 2026-08-25T02:58:34+00:00

CAD shows its largest performance gains in the knowledge-conflict setting where the input context contradicts the model's training-time prior knowledge.

Summary

CAD is most valuable when you need to override what a model already assumes from its training. Its biggest performance jumps happen precisely when the information you provide clashes with the model's built-in expectations, meaning it is built to correct outdated or wrong assumptions rather than simply restate things the model already "knows."