LongHarness preprint reports gains in reasoning across large document collections

Researchers report in an arXiv preprint that LongHarness improved AI agents' accuracy when answering questions across large document collections. Using GLM-5.3-Flash, it raised exact-match scores on OfficeQA Pro from 15.8% to 60.2% on V1 and from 4.4% to 61.1% on V2 compared with direct reasoning using the same model. The system tracks evidence separately from evolving conclusions. Unresolved conclusions guide searches for missing information, while new evidence revises those conclusions; an additional process keeps evidence current and traceable. Its reported OfficeQA Pro results exceeded Claude Code by 12.0 percentage points on V1 and Codex by 18.9 points on V2. The evaluations used GLM-5.3-Flash and Qwen3.8-Flash across four named benchmarks.