LongHarness pushes AI reasoning accuracy over long documents past 60%

A team of eleven researchers has published LongHarness, a harness, meaning a software framework that directs how an AI model searches, built to help AI agents answer questions that require reading far more text than fits in one context window, the stretch of text a model can process at once. The paper, posted to arXiv’s cs.AI section, describes a system meant for corpora, meaning large document collections, that can run past 100 million tokens, the chunks of text a model reads.

The researchers say today’s agent harnesses let a model fetch information from a large corpus piece by piece, but the evidence it gathers and the conclusions it draws tend to blur together in one long interaction history. They identify two resulting problems: unbounded exploration, where an agent keeps retrieving documents without closing the actual information gap, and evidence-state drift, where older evidence or later corrections stop shaping the agent’s current answer.

LongHarness addresses this with two linked mechanisms. The first, called Evidence-Conclusion Coevolution, uses whatever conclusion the agent has not yet confirmed to steer its next search, then lets newly found evidence either reinforce or revise that conclusion. The second, Evidence Lifecycle Management, is described as keeping the evidence the agent has already collected current and traceable as the interaction continues, rather than letting it pile up unchecked.

The team tested LongHarness on four benchmarks: OfficeQA Pro V1 and V2, AA-LCR, and LongBench v2, running it on two underlying models, GLM-5.3-Flash and Qwen3.8-Flash. The paper reports exact-match scores, meaning the fraction of answers matching the correct answer precisely, only for the OfficeQA Pro tests. With GLM-5.3-Flash, exact match rose from 15.8% to 60.2% on OfficeQA Pro V1 when LongHarness was used instead of having the model reason directly, and from 4.4% to 61.1% on V2.

The researchers also compare LongHarness against existing agent harnesses on the same backbone models. They report it surpassing Claude Code by 12.0 percentage points on OfficeQA Pro V1 and Codex by 18.9 points on V2, both using GLM-5.3-Flash. Swapping in Qwen3.8-Flash as the backbone, LongHarness reached 61.7% on V1, which the paper says beats Claude Code by 8.3 points and Codex by 12.0 points.

The abstract does not give specific scores for the AA-LCR or LongBench v2 benchmarks, so how LongHarness performs on those tests, and how it holds up outside tests built by the same research group, remains to be seen from the fuller paper. The listing also shows a DOI pending registration, meaning the paper’s permanent citation record is not yet finalized. Whether the authors release code or data for the system, which would let outside groups verify the reported gains, is the next thing to watch.