What happened
The ParseBench maintainers released v2.0.2 on November 14, 2026. The GitHub release notes say the update adds a Tencent Hy4 pipeline and benchmark results. It also publishes Reducto r-1 ParseBench 2 outputs and verifies that system’s row on the leaderboard. Those are concrete additions to the systems represented in the benchmark.
The release announcement does not state the scores for either Tencent Hy4 or Reducto r-1, and it does not describe their positions relative to other entries. That distinction matters. A benchmark entry can make a system visible to evaluators without establishing that it is a broadly superior choice. The release nevertheless gives teams using ParseBench another pair of named systems to inspect when forming a candidate set for document extraction or document-understanding work.
Version 2.0.2 also changes the operating machinery around the ParseBench 2 leaderboard. The notes list results-v2.yaml, a --board v2 option, and workflows for the v2 board under “leaderboard accountability.” They additionally record run dates in result maps and add a per-subset missing-pages check. These changes make it more feasible to see which result artifacts belong to the versioned board and whether coverage gaps exist within a subset, rather than treating a leaderboard row as a self-contained claim.
Another listed change removes the leaderboard output coverage threshold. The release notes name that change but do not give the threshold’s former value, explain why it was removed, or quantify the effect on any rankings. Readers should therefore avoid inferring that the update tightened every eligibility condition. It combines new checks and metadata with one removal in the display or scoring workflow, and the practical consequences require inspection of the repository’s benchmark implementation.
Two fixes address scorer behavior. Markdown table cells are now normalized like HTML cells, and escaped dollar signs are preserved in the v2 text and formatting scorers. Both are narrow but consequential details for document-parsing evaluation. A parser may express an equivalent table through different markup, while financial or technical documents can contain literal dollar signs whose escaping should not be lost before comparison. If the scorer treats such representation differences inconsistently, it can attribute an evaluation discrepancy to a parser when the disagreement is actually in the normalization path.
The release follows v2.0.1; GitHub presents a full changelog comparison between those versions and reports 28 commits to the main branch since this release. The announcement describes a benchmark maintenance and coverage update rather than a new parser product, a pricing revision, or an access change. It also does not announce a change in ParseBench’s stated purpose. The immediate value is in the added result artifacts and the revised mechanisms used to organize and score them.
What the archive adds
New Runtime’s ParseBench-2: Severity-aware document parsing benchmark covered the earlier ParseBench-2 introduction on November 12, 2026. This is a new fact on a known story: v2.0.2 adds Reducto r-1 outputs, Tencent Hy4 results, and specific leaderboard and scorer changes.
Trend signal
This update supports scientific-agents-need-artifact-evidence: evaluation claims are more useful when a reader can inspect the result artifacts and the machinery around them, rather than seeing only a leaderboard number. Versioned result data, run dates, per-subset missing-page checks, and explicit scorer fixes all move in that direction. The release does not establish independent reproducibility or prove that every listed system was evaluated under equivalent operational conditions.
Why it matters
Document-parsing teams often face a gap between a benchmark headline and production behavior. Contracts, invoices, presentations, scanned forms, tables, and text containing special characters can stress different parts of an extraction pipeline. New results can help teams decide what to test, but scorekeeping details determine whether a comparison is interpretable in the first place.
The scorer changes are therefore more than routine cleanup. If equivalent table representations are normalized differently, or escaped characters are handled inconsistently, an evaluation can reward formatting coincidence rather than extraction quality. The added result maps and coverage checks offer a better basis for auditing a row before relying on it. That is especially useful when a team is comparing vendors or models whose public claims are otherwise difficult to reconcile.
Still, benchmark coverage is not a deployment decision. Teams should use ParseBench as a source of hypotheses: identify candidate systems, inspect the evaluation path, and then run a local test set that reflects their own documents, acceptable error types, latency budget, and operating cost. A verified leaderboard row cannot answer those workload-specific questions.
Limits and uncertainty
The release notes do not state the Reducto r-1 or Tencent Hy4 scores, their relative leaderboard positions, or their cost and latency. They do not report before-and-after accuracy changes from the scorer fixes, error rates by subset, confidence intervals, or the effect of removing the output coverage threshold. The notes identify a Tencent Hy4 pipeline but do not detail its configuration or the evaluation protocol in the release announcement. Independent reproduction, comparability across systems, and suitability for a particular document corpus remain to be checked.
Skill to build now
Create a small regression pack from the document features that matter in your workflow: Markdown-like and HTML-like tables, pages with literal currency symbols, missing or sparse pages, and the document classes that cause downstream failures. Run candidate parsers against the pack, retain both extracted outputs and scoring artifacts, and review mismatches by failure type. Use public benchmark rows to choose candidates, but make acceptance criteria explicit for the documents your team actually processes.
Reader takeaway
ParseBench v2.0.2 broadens its document-parsing comparisons with Reducto r-1 and Tencent Hy4 while making several tracking and scoring details more inspectable. Treat the release as a better starting point for a local evaluation, not as a final answer on which parser to deploy.