Coding agents fix more with whole files
Why it matters
A University of Illinois Urbana-Champaign study gave every method the same per-issue token budget on 300 recent SWE-rebench issues. Showing a few whole files verbatim (CQ2-CARE: ask the model twice which files must change, keep the files both answers name) resolved 43.7% with GPT-5.6 and 46.7% with Claude Opus 5, 6.3 points more than the best of three prompt compressors and 9.3 to 12.7 more than the best retriever. Compressors delete text a patch must quote: with GPT-5.6, 98.0% of whole-file patches applied vs 70.7% for LLMLingua-2. Caveat: one patch call per issue, Python issues from one benchmark.
Read the source: arXiv (UIUC) (opens in a new tab)
All hits from 19 Oct 2026More CodingGet the daily hits by email