cad-alpha-task-specific-values

IN premise — summaries/2026/08/24/shi-2024-context-aware-decoding-s3-experimental-setup.md

Created 2026-08-25T02:58:35+00:00

The recommended α values are 0.5 for summarization tasks (CNN-DM, XSUM) and 1.0 for knowledge-conflict tasks (MemoTrap, NQ-Swap).

Summary

When the system is asked to summarize text, it should blend its internal knowledge at half strength, but when it faces a direct conflict with new information, it should fully commit to the new input. In practice, this means the right configuration depends on what the task is asking the model to do, not on a single universal setting.