CoW-Bench: A Multi-Domain Benchmarking Framework

• CoW-Bench is a flexible benchmarking framework that standardizes evaluations across high-dimensional search, neutron interferometry, quantum cryptography, and anatomical segmentation.
• It converts domain-specific tasks into core stress tests by isolating key observables such as competitive ratios, phase shifts, QBERs, and topological accuracy.
• The framework demonstrates that high primary metrics may mask deeper structural limitations, highlighting the need for comprehensive, multi-faceted evaluation methods.

CoW-Bench is used in the available literature as a benchmark interpretation attached to several technically unrelated “COW” or “CoW” problem families. The label is applied to the $d$-dimensional cow-path problem as a canonical test for high-dimensional online search (Bansal et al., 2022), to Colella–Overhauser–Werner interferometry as a low-energy benchmark for weak-equivalence-principle tests and possible noncommutative or generalized-uncertainty-principle effects [(Saha, 2013); (Krause et al., 2014); (Hamil et al., 2022)], to Coherent One-Way quantum key distribution as a finite-key and entanglement-verification benchmark (Pathania et al., 26 Feb 2026, Rezazadeh et al., 1 Dec 2025), and to topology-aware segmentation of the Circle of Willis through the TopCoW challenge (Yang et al., 2023). This suggests that CoW-Bench functions primarily as a benchmark label: a way of turning a domain-specific COW/CoW task into a standardized stress test for geometry, interference, security, or topology.

1. Benchmark semantics and scope

The available sources do not use CoW-Bench to denote a single standardized benchmark with one protocol, one metric, or one dataset. Instead, the term appears as a benchmark interpretation in multiple research programs. In the geometric-search setting, the $d$-dimensional cow-path problem “can serve as a canonical test for worst-case performance of search strategies” in high-dimensional Euclidean space (Bansal et al., 2022). In neutron interferometry, the COW experiment “functions as a benchmark—or ‘CoW-Bench’—for testing both WEP and possible new physics” (Saha, 2013). In practical quantum cryptography, a “CoW-Bench” is described as a benchmark suite for experimental COW-QKD and finite-key analysis (Pathania et al., 26 Feb 2026). In medical image analysis, TopCoW is explicitly presented as “a first attempt at benchmarking the CoW anatomical segmentation task for MRA and CTA, both morphologically and topologically” (Yang et al., 2023).

A recurring structural feature is the conversion of a domain-specific task into a compact evaluative core. In the cow-path literature, that core is competitive ratio. In COW interferometry, it is a phase-sensitive probe of gravitational or beyond-standard dynamics. In COW-QKD, it is the joint behavior of QBER, phase error, finite-key secure key rate, and effective entanglement. In TopCoW, it is the combination of volumetric overlap and topological correctness. A plausible implication is that CoW-Bench denotes a benchmarking style rather than a single benchmark ontology.

2. High-dimensional online search: the cow-path benchmark

In the $d$-dimensional cow-path problem, the search space is $\mathbb{R}^d$, the searcher starts at the origin, and the hidden target is an unknown $(d-1)$-dimensional hyperplane $H$ at distance $r \ge 1$. A deterministic algorithm is a continuous path $P$ starting at the origin, and if the path travels total length $s$ before first intersecting $H$, the competitive ratio is $d$0. Under the standard normalization $d$1, the problem becomes equivalent, up to constant factors, to viewing every point of the unit sphere $d$2 (Bansal et al., 2022).

The benchmark significance of this model comes from its nearly tight asymptotics. Earlier work gave an upper bound $d$3 and a lower bound $d$4, leaving a $d$5 gap. The note “A Nearly Tight Lower Bound for the $d$6-Dimensional Cow-Path Problem” strengthens the lower bound to

$d$7

or $d$8, using a geometric adversarial argument based on sphere visibility, concentration of measure, short-subpath replacement by a farther point, and iterative orthogonal projections (Bansal et al., 2022). Combined with the $d$9 upper bound, the competitive ratio is nearly determined up to a factor of $d$0.

The sphere-inspection duality is central. A point $d$1 views $d$2 iff $d$3, and a path views all of $d$4 iff $d$5 is contained in the convex hull of the path. The lower bound then measures how much of the sphere can be seen from bounded radius and how projection into successively lower-dimensional subspaces forces cumulative path length. Interpreted as CoW-Bench, this benchmark isolates the intrinsic worst-case cost of deterministic high-dimensional online search: performance substantially better than order $d$6 is impossible in the deterministic regime (Bansal et al., 2022).

3. Matter-wave interferometry: COW as a precision benchmark

In the Colella–Overhauser–Werner experiment, a monochromatic beam of thermal or cold neutrons enters a silicon perfect-crystal interferometer, Bragg diffraction coherently splits it into two paths, and the recombined beams produce a gravity-induced phase shift because the paths traverse different gravitational potentials. In the standard semiclassical treatment,

$d$7

with $d$8 the enclosed area and $d$9 the de Broglie wavelength. The benchmark role of the experiment is tied to the weak equivalence principle through $\mathbb{R}^d$0: agreement between the gravitational acceleration inferred from the phase and the macroscopic value is interpreted as a quantum-regime test of WEP (Saha, 2013).

One CoW-Bench strand uses this apparatus to probe noncommutative spacetime. With canonical noncommutative algebra $\mathbb{R}^d$1, the physically relevant first-order effect arises from the time-space sector $\mathbb{R}^d$2, yielding an effective Schrödinger equation with observed gravitational acceleration

$\mathbb{R}^d$3

and a corrected phase shift containing an explicit time-space noncommutative contribution. Using a reported $\mathbb{R}^d$4 discrepancy in a COW-type experiment and assuming that WEP holds to significantly better accuracy than $\mathbb{R}^d$5, the resulting upper bound is

$\mathbb{R}^d$6

which is stronger by about four orders of magnitude than an earlier GRANIT-based bound (Saha, 2013). The conceptual point is that an apparent WEP violation can arise even if $\mathbb{R}^d$7 remains universal, because the spacetime structure itself produces a mass-dependent correction.

A second strand studies quantum undecayed unstable particles in COW interferometry. There the relevant complex phase difference decomposes as $\mathbb{R}^d$8, with

$\mathbb{R}^d$9

The output intensity becomes

$(d-1)$0

so the visibility and a priori path predictability are

$(d-1)$1

with $(d-1)$2 (Krause et al., 2014). The benchmark lesson is that loss of contrast can originate from intrinsic instability and path-dependent survival probability, not only from measurement-induced decoherence.

A third strand embeds the COW phase in higher-order generalized uncertainty principle models. In the exponential model, the effective gravitational acceleration is

$(d-1)$3

while in the minimal-length/maximal-momentum model it is

$(d-1)$4

The gravitational phase shift is then written as

$(d-1)$5

so the benchmark observable is a momentum-dependent deviation from the standard COW phase (Hamil et al., 2022). In the paper’s synthesis, COW interferometry, Einstein–Bohr’s photon box, and black-hole thermodynamics are linked by the same deformation functions.

4. Quantum cryptography: Coherent One-Way QKD as CoW-Bench

In practical quantum cryptography, CoW-Bench refers to the Coherent One-Way protocol and its evaluative framework. The experimentally realized variant is the two-decoy, two-vacuum-pulse COW-QKD protocol, in which two consecutive time bins form one round, logical $(d-1)$6-basis states are $(d-1)$7 and $(d-1)$8, and the decoys are $(d-1)$9 and $H$0. Bob uses a 90:10 beam splitter, a data line with a single SPAD detector for time-of-arrival key generation, and a monitoring line with an unbalanced Mach–Zehnder interferometer and two SPAD detectors for coherence monitoring (Pathania et al., 26 Feb 2026).

The finite-key framework reports QBER, phase error rate, and secure key rate under composable security with $H$1 and $H$2. For the experimentally realized system at $H$3 km, $H$4, and $H$5, the setup was kept running for $H$6 hours, with stable secure key rates in the range $H$7–$H$8 kbps and QBER stabilizing around $H$9. Under temperature variation from $r \ge 1$0 to $r \ge 1$1C, QBER remained between $r \ge 1$2 and $r \ge 1$3, and SKR remained between $r \ge 1$4 and $r \ge 1$5 kbps (Pathania et al., 26 Feb 2026).

The benchmark also distinguishes between different operational horizons. Using the analytical gain model and a $r \ge 1$6 abort threshold for QBER, the paper reports QBER-based limits of about $r \ge 1$7 km for detector efficiency $r \ge 1$8 and about $r \ge 1$9 km for detector efficiency $P$0. In the finite-key regime with block size $P$1, however, the secure key rate drops to zero at about $P$2–$P$3 km for $P$4 and about $P$5 km for $P$6 (Pathania et al., 26 Feb 2026). The benchmark significance lies in this separation between a feasibility distance inferred from raw error rates and a shorter distance compatible with finite-key secrecy.

A second COW-QKD strand turns CoW-Bench into an entanglement-verification problem. In the effective-entanglement formulation of COW, the single-photon sector is represented by $P$7, $P$8, $P$9, and $s$0. The paper “Witnessing the Effective Entanglement in the COW Protocol” introduces the two-parameter family

$s$1

derives the admissible $s$2 region where these are valid entanglement witnesses, and shows that negative expectation values on experimental data reveal effective entanglement (Rezazadeh et al., 1 Dec 2025). Since effective entanglement is a necessary precondition for secret key distillation in the prepare-and-measure setting, this witness family supplies a benchmark that is orthogonal to raw key-rate reporting: it tests whether the observed COW statistics can still be explained by a separable effective state.

5. Topology-aware anatomical segmentation: the TopCoW benchmark

In medical image analysis, CoW-Bench refers to TopCoW, a benchmark for topology-aware anatomical segmentation of the Circle of Willis in CTA and MRA. The dataset contains $s$3 unique patients, each with exactly one CTA and one MRA, making $s$4 paired CTA–MRA cases. It is described as the first public dataset with voxel-level annotations for thirteen possible CoW vessel components and the first large dataset with paired MRA and CTA from the same patients (Yang et al., 2023).

The label space contains $s$5 classes: basilar artery, right and left posterior cerebral arteries, right and left internal carotid arteries, right and left middle cerebral arteries, right and left posterior communicating arteries, anterior communicating artery, right and left anterior cerebral arteries, and third A2. Annotation used virtual-reality technology, specifically syGlass with a Meta Quest head-mounted display, and a typical CTA+MRA pair could usually be annotated within about two hours. The challenge data split was $s$6 training patients, $s$7 validation patients, and $s$8 hidden test patients, with an additional $s$9 CROWN MRAs re-annotated for training (Yang et al., 2023).

TopCoW formalized the problem as both binary vessel segmentation and $H$0-class multiclass segmentation, with evaluation restricted to a manually defined CoW ROI. Official metrics combined morphological and topological criteria: Dice, clDice, and $H$1 error for binary segmentation; per-case class-average Dice, clDice on the merged binary mask, and per-case class-average $H$2 error for multiclass segmentation (Yang et al., 2023). Beyond the official leaderboard, the authors also performed component detection analysis and variant topology matching, requiring correct presence or absence, correct neighbors, and no $H$3 errors.

The benchmark outcomes emphasize the gap between overlap accuracy and anatomical correctness. The challenge attracted over $H$4 registered participants from four continents. Top-performing teams segmented many CoW components to Dice scores around $H$5, but scores were lower for communicating arteries and rare variants. Additional topological analysis showed that predictions with high Dice scores could still contain topological mistakes, particularly for Acom, Pcoms, third A2, fetal PCA configurations, and hypoplastic segments (Yang et al., 2023). In benchmark terms, TopCoW makes explicit that CoW characterization is not exhausted by voxel overlap.

6. Cross-domain logic, misconceptions, and recurrent limitations

Across these literatures, CoW-Bench repeatedly isolates a small set of observables and then asks whether they are sufficient for a deeper structural guarantee. In the cow-path problem, the observable is competitive ratio, but the deeper issue is worst-case geometric coverage of all hyperplanes or, equivalently, all sphere directions (Bansal et al., 2022). In COW interferometry, the observable is a phase shift or fringe contrast, but the deeper issue may be WEP, noncommutative dynamics, GUP corrections, or intrinsic which-way information [(Saha, 2013); (Krause et al., 2014); (Hamil et al., 2022)]. In COW-QKD, QBER alone is not the full object of interest; phase error, finite-size corrections, and effective entanglement remain decisive (Pathania et al., 26 Feb 2026, Rezazadeh et al., 1 Dec 2025). In TopCoW, high Dice does not guarantee correct vascular topology (Yang et al., 2023).

Several recurrent misconceptions are therefore directly contradicted by the benchmark evidence. One is that CoW-Bench names one benchmark; the sources instead use the label in multiple unrelated domains. Another is that a high primary metric is automatically sufficient. In TopCoW, high Dice can coexist with topological mistakes (Yang et al., 2023). In COW-QKD, a favorable QBER-based reach can exceed the finite-key secure distance by a large margin (Pathania et al., 26 Feb 2026). In COW interferometry, an apparent WEP violation can arise from time-space noncommutativity even if $H$6 is universal (Saha, 2013), and visibility loss can arise from unstable-particle survival asymmetry without measurement-induced decoherence (Krause et al., 2014).

A plausible implication is that CoW-Bench, as used in these sources, marks a family of benchmarks in which the operationally measured quantity is intentionally simple, while the interpretive target is structurally richer. That pattern explains why the term recurs across worst-case search, matter-wave precision tests, quantum cryptography, and topology-aware segmentation, despite the absence of any shared data format, shared hardware, or shared mathematical domain.