Report
LongHarness reportedly raises GLM-5.3-Flash’s OfficeQA Pro V2 exact match from 4.4% to 61.1%
A post says the model-agnostic harness tracks evidence and conclusions for corpora of 100 million tokens or more.
TLDR
A post sharing the LongHarness paper says GLM-5.3-Flash’s exact match on OfficeQA Pro V2 rose from 4.4% to 61.1% with the same model. The post says the paper’s authors attribute the gain to the harness, which stores evidence and conclusions separately and uses open conclusions to guide retrieval. It also reports that LongHarness beat Claude Code by 12.1 points on V1 and Codex by 18.9 points on V2 of the benchmark.
Combined views
—
1 Source, first seen ago