Survey, not a system. No original benchmarks -- do not plot. Its Table VII ("Classification of representative ZKML systems by verification objective") CLASSIFIES 30 systems (2017.6 SafetyNets -> 2025.08 VeriLoRA) by Training / Testing / Inference, with Date and Model columns. The TIMELINE is Figure 6, not Table VII -- CORRECTED 2026-07-13; the two were fused, and content/surveys/index.md and content/papers/zkml-survey.md inherit the error. Tables IV, V and VI give proving time / verification time / proof size for training, testing and inference systems respectively, each with the hardware used. This file's numbers_source: survey entries are drawn from those tables.

What is new

The taxonomy, and it is the reason this repo is organized the way it is. The survey's decisive move is to split what is being proven into three objectives, inference ("this model produced this output on this input"), testing ("this model achieves this accuracy on this dataset"), and training ("this model is the result of correctly training on this data"), and then to insist that a system's cost is only interpretable relative to which one it targets. That distinction is what makes it possible to notice that zkDT and zkCNN are doing something categorically different from zkGPT, and it is a better organizing principle than "by proof system," which is how most of the field talks about itself.

Its Table VII, a timeline of thirty systems, from SafetyNets in 2017 through to 2025, is the single most useful artifact in the paper, and is where most of the older entries in papers.yml came from.

Beyond that: it is a survey. There are no original benchmarks, no reimplementations, no reproductions. It is excluded from the plots for that reason, and it should be.

What it actually proves

Nothing, in the technical sense, but the question worth asking of a survey is what did the authors actually verify?, and the answer is: they read abstracts.

Its quantitative tables (IV for training, V for testing, VI for inference) are compilations of numbers as reported by the primary papers. They are not re-measured, not normalized for hardware, not normalized for bit width, not normalized for sequence length, and, most consequentially, not checked against the primary papers' own tables.

The coverage window is worth noting too: the abstract states the review covers "ZKML research published from June 2017 to August 2025," and the version in references/ is a March 2026 revision that did not extend it. So DeepProve, Jolt Atlas and NANOZK, the entire 2026 frontier this repo is built around, are outside its scope. It is a good map of a field that has since moved.

What to distrust

Table VI, specifically and concretely. Three of this repo's known conflicts trace to it, and reading the primary papers has now resolved the most important one against the survey.

That single row tells you the failure mode. The survey is transcribing headline claims, and headline claims in this field are systematically the most flattering number an author could print. The other two recorded conflicts have the same shape, the survey's figures for ZKML and Lu et al. on GPT-2 disagree with zkGPT's direct re-measurement of those same systems, and zkGPT is the only party that actually ran them.

The 33-byte proof size for zkDL (Table IV) is smaller than a single pair of group elements and cannot be right for any pairing-based or FRI-based scheme. It is still in papers.yml flagged as implausible, and it should stay flagged until someone reads the zkDL paper.

Hardware is recorded but never used. Table VI diligently lists the platform for each system, a large cloud instance here, an A100 there, an 8-core desktop somewhere else, and then presents the proving times side by side as though they were commensurable. The survey's own column shows they are not, and the survey never says so.

Quantization is absent. The word does not organize anything in this survey, and there is no column for bit width in any of the three tables. This is the field's central confounder, SpaGKR measures ternary quantization alone buying a further several-fold prover speedup, ZEN measures proof-friendly quantization cutting constraint counts by an order of magnitude, and a survey whose entire purpose is to make systems comparable does not track the variable that makes them incomparable. That omission is the gap this repo exists to fill.

Use it for the map, not the measurements. The taxonomy is good and the timeline is genuinely useful. Every number sourced from it in papers.yml carries numbers_source: survey for a reason, and the reason has now been demonstrated rather than assumed.