NVIDIA paper finds frequent context compaction keeps frontier LLM accuracy

NVIDIA researchers report that compacting the context after each action keeps frontier language models' accuracy on long-horizon tasks while shortening their context.

Elvis Saravia, summarizing the paper on X, says Claude Opus 4.5 scores 70.2 on SWE-bench Verified with one action per context, against 70.0 with the full history. The median peak context falls from 32K to 8K tokens. The paper reports the pattern on Terminal-Bench, FRAMES, BFCLv3 and multilingual SWE-bench, with modest losses in some settings.

The paper also notes that later output tokens can cost up to 7x more than earlier ones on frontier open-weight models, according to Saravia's post.