LongHarness — Evidence-conclusion coevolution — What does it mean?

The news. On October 9, 2026, a research team including Fudan University and Shanghai AI Laboratory posted LongHarness, a model-agnostic harness for reasoning over corpora of up to 100M tokens. The authors report GLM-5.3-Flash rising from 15.8% to 60.2% exact match on OfficeQA Pro V1 and from 4.4% to 61.1% on V2, and beating Claude Code by 12.0 points on V1 and Codex by 18.9 points on V2. Read the paper →

Picture a detective working a case whose files fill an archive of 100M tokens, while only a few pages fit on the desk at once. A typical agent harness is a detective who writes everything into one ever-growing notebook: every search, every page read, every half-formed hunch. When evidence and conclusions share one transcript, the detective loses track of which facts are still standing. The paper names two resulting failures. In unbounded exploration, retrieval keeps lengthening the history without closing any gap the decision depends on. In evidence-state drift, an earlier fact or a later correction never reaches the current conclusion.

LongHarness swaps the notebook for a case board with two separate parts. The EvidenceBank holds source-linked statements taken from passages the agent really read. The Conclusion State holds what the agent currently believes about them: established claims, supported hypotheses and unresolved gaps. Keeping what was observed apart from what was inferred is the structural idea. It also changes how the agent searches: for each claim it asks which observation would establish, qualify or refute it, and that answer becomes the next query. This is the same retrieve-then-read loop covered in the retrieval pipeline, except the query comes from an open question, not from the original prompt.

The two parts feed each other in a loop, which is what the paper calls coevolution. First, an unresolved gap becomes a search, and what the search returns is admitted to the EvidenceBank. Second, the agent re-checks each claim against the active evidence and marks it as still supported, in need of a qualifier, or revised. Competing hypotheses lead to searches that tell them apart. Before it submits, the agent folds in any evidence not yet reflected in its conclusions and checks whether the remaining gaps could change the answer.

Evidence also has a lifecycle. A statement starts as a candidate, becomes active once admitted, and can later be revised or retired. Each candidate must cite source references that an earlier read action registered, and the controller checks those references are valid, while the agent itself judges whether the statement is faithful to the source. A correction enters as a new active entry and the old one is marked superseded; a retraction leaves the active set with no replacement. Superseding does not rewrite dependent conclusions automatically: the agent sees the update and then revises the affected claims itself, and the final answer must cite references that resolve to active evidence. In course terms this is structured note-taking with source checks and an explicit way to retire a note, aimed at the conflict failure mode, where a context holds two contradicting facts.

The paper's own case study shows why this matters. The task is the geometric mean of 80 monthly Treasury expenditure values. The Codex harness (running GLM-5.3-Flash) found the right table in 22 command calls, left some missing values implicit, and then made 103 consecutive backward reads. It computed 4958.09 using 4100 for May 1942, later recognized the corrected value 4400, and still kept the old result in its final answer. A geometric mean over 80 values takes the 80th root of the product, so changing one operand multiplies the result by (4400 / 4100)^(1/80), about 1.00088, which moves 4958.09 to about 4962.5. The paper states that replacing this operand yields the reference answer, 4962.46; my one-line check agrees only to rounding. The mistake is only about 0.088%, so it passes a 5% tolerance and still fails exact match. LongHarness recorded 4400 in the EvidenceBank, tracked cross-issue consistency as an open gap, corrected two other entries, and returned 4962.46 in 10 policy actions.

Every ablation costs accuracy, and all five cost more on V2, the split with a larger corpus whose questions combine values from several sources. Append-only evidence loses 10.0 points on V2 even though it keeps the conclusion feedback, so being able to retire a fact matters as much as being able to store one. The generic-notes row is the one to remember for builders: free-text notes that can hold conclusions, open questions and references still lose 6.8 and 7.8 points to the structured version. Effort grew far less for LongHarness. Between V1 and V2, median model turns per question rose from 22 to 30 for LongHarness, from 21 to 116 for Claude Code and from 19 to 43 for Codex, and Codex exhausted its recorded turn budget on 3.8% of V1 questions and 21.1% of V2 questions. All numbers are the authors' own, from one run per harness and setting, so treat the gaps as strong evidence for the design and not as a settled leaderboard.

Goes deeper in: AI Agents → Context Engineering → The 4 Fixes

Related explainers

• Harness-1 — externalizing search state into the harness — the same idea of keeping state outside the transcript, aimed at keeping context flat
• LedgerAgent — a state ledger with policy validation — another agent that keeps task facts in a checked ledger