cad-llama30b-cnn-dm-results

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

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

LLaMA-30B on CNN-DM achieves +21% ROUGE-L, +14.3% FactKB, and +7.8% BERT-Precision with CAD versus standard decoding.

Summary

When a 30-billion-parameter language model is asked to summarize news articles, adding the CAD decoding strategy on top of its usual process produces summaries that are meaningfully better across three quality dimensions: closer to the reference summary, more factually consistent, and more precise. This is a concrete, measured result showing CAD delivers real gains in practice rather than just looking good on paper, and it sets a baseline that other techniques or configurations will be compared against.