Study Reveals Long-Context Models Waste Tokens Copying Text

A new study reveals that long-context reasoning models waste computational tokens by copying irrelevant input text into their reasoning chains. To address this inefficiency, researchers introduced a new reward mechanism that improves model performance by up to 4.6 points.

2026-07-23 ~ 2026-07-23 · 2 related posts

• Long-context models still copy irrelevant text, and a new reward lifts accuracy by up to 4.6 points — rohanpaul_ai · 2026-07-23
• Long-context models still waste tokens by copying the prompt into reasoning — eyishazyer · 2026-07-23