Squrve: Text-to-SQL, Curve Analysis, & Squircle

• Squrve is a term spanning three domains: a modular Text-to-SQL framework, SRV-based elastic shape analysis, and the Fernandez–Guasti squircle in algebraic geometry.
• The Text-to-SQL framework version of Squrve modularizes query processing into seven atomic actors, achieving high execution accuracy on benchmarks like Spider-dev and Bird-dev.
• In geometric contexts, Squrve refers either to the SRV framework for elastic curve analysis or to a quartic squircle, an intermediate shape between a circle and a square with practical mapping applications.

Searching arXiv for the term and closely related uses to ground the article in the cited literature. “Squrve” is an ambiguous term used in at least three distinct technical contexts in the arXiv literature: as the name of a modular Text-to-SQL framework, as an informal label associated with square-root-velocity-based curve analysis, and as a variant spelling for the Fernandez–Guasti squircle, a circle–square hybrid algebraic curve. In current research usage, the explicit proper-name usage “Squrve” refers to the Text-to-SQL framework introduced in 2025 (Wang et al., 28 Oct 2025). In geometric analysis, closely related usage appears around the square root velocity (SRV) or square-root-velocity function (SRVF) framework for elastic shape analysis of curves (Hartman et al., 2021, Lahiri et al., 2015). In planar geometry, “Squrve” is also used informally for the Fernandez–Guasti squircle, a low-degree quartic interpolant between a circle and a square (Fong, 2016). The term therefore requires domain-specific disambiguation.

1. Lexical scope and domain-specific meanings

In the most explicit contemporary usage, Squrve denotes “a unified, modular, and extensible framework for real-world Text-to-SQL,” designed to integrate heterogeneous academic methods and deployment-oriented system components (Wang et al., 28 Oct 2025). This usage treats Squrve as a software and systems framework rather than as a geometric object or a mathematical transform.

A second usage appears in discussions of elastic shape analysis, where the square root velocity representation is described as central to “Squrve”-style analysis. In that setting, the object of study is not a framework named Squrve, but the SRV or SRVF representation used to compare curves modulo transformations such as translation, rotation, and reparametrization (Hartman et al., 2021). This suggests that “Squrve” can function as an informal shorthand for SRV-based curve analysis, although the formal paper titles use SRV or SRVF terminology rather than Squrve as a canonical name.

A third usage is geometric and algebraic: if “Squrve” is intended to denote a rounded-square curve, the relevant object is the Fernandez–Guasti squircle. That curve is a mathematically precise intermediate shape between a circle and a square, defined by a quartic biquadratic equation and studied for its area, polar form, parametrizations, and three-dimensional extension (Fong, 2016).

These meanings are not interchangeable. The Text-to-SQL framework concerns modular orchestration of LLM-based database querying; the SRV literature concerns infinite-dimensional quotient geometry of curves; the squircle literature concerns algebraic plane curves and associated mappings. A plausible implication is that any unqualified use of “Squrve” in technical writing benefits from immediate contextual specification.

2. Squrve as a Text-to-SQL framework

The 2025 framework named Squrve was introduced to address a gap between high-performing academic Text-to-SQL methods and deployable real-world systems (Wang et al., 28 Oct 2025). Its stated motivation is that strong methods for schema linking, decomposition, decoding, self-correction, optimization, and selection remain fragmented because they use incompatible interfaces, task assumptions, data formats, and execution pipelines. Squrve addresses this fragmentation through a universal execution paradigm and a modular collaboration mechanism built from reusable actor abstractions.

The framework is organized into four major modules: Data module, Actor module, Task module, and Execution module (Wang et al., 28 Oct 2025). The Data module standardizes schema and instance representations across benchmarks such as Spider and BIRD and supports user-provided private data. The Actor module abstracts seven reusable capabilities from prior literature. The Task module binds together an actor interface, a dataset interface, and a runtime controller. The Execution module provides a configuration manager, an execution controller, and an evaluation unit.

A central formalization in the framework defines an actor as a black-box mapping

$SQL = f_{\text{actor}}(Q, S, C \mid M),$

where $Q$ is the user query, $S$ the database schema, $C$ valuable contextual information specific to the query, and $M$ the LLM backbone (Wang et al., 28 Oct 2025). This standardizes invocation across otherwise heterogeneous methods such as DIN-SQL, CHESS, MAC-SQL, RSL-SQL, LinkAlign, CHASE-SQL, OpenSearch-SQL, and MCS-SQL.

The seven atomic actors are Reduce, Parse, Generate, Decompose, Scale, Optimize, and Select (Wang et al., 28 Oct 2025). Reduce filters large schemas; Parse performs schema linking; Generate produces SQL directly; Decompose breaks a query into progressive reasoning steps; Scale broadens candidate generation; Optimize performs execution- or environment-driven refinement; and Select chooses among candidates. These actor abstractions are the framework’s core conceptual contribution, because they factorize monolithic methods into reusable capabilities.

Squrve supports two composition strategies for combining actors into complex workflows: Pipeline and Tree (Wang et al., 28 Oct 2025). Pipeline composition supports staged refinement, while Tree composition supports parallel ensembling and later merging or selection. The framework explicitly treats these as recursive “plug-and-socket templates,” so composite actors can themselves be reused as higher-level components.

Empirically, the paper evaluates reproduced baselines and collaborative ensembles on Spider-dev and Bird-dev using Qwen-3 as the LLM backbone (Wang et al., 28 Oct 2025). The strongest reported ensemble, Ensemble-2, attains 90.8 execution accuracy on Spider-dev and 70.0 on Bird-dev, outperforming all reproduced individual baselines. The two identified collaborative workflows are:

Workflow	Composition
Ensemble-1	LinkAlignParser + RSLSQLScaler + CHESSSelector + MACSQLOptimizer
Ensemble-2	RSLSQLBiDirParser + MACSQLDecompose + ChessScaler + CHESSSelector + LinkAlignOptimizer

The framework also reports a schema-linking parallel analysis on AmbiDB and Bird-dev, where parallel actor composition improves recall and where two-actor parallelization provides the best trade-off relative to adding a third actor (Wang et al., 28 Oct 2025). This supports the broader claim that modular collaboration can strengthen intermediate stages as well as end-to-end performance.

At the implementation level, Squrve uses a configuration-driven startup pattern based on core.base.Router and core.engine.Engine, with unified task definitions and execution order (Wang et al., 28 Oct 2025). This positions the framework primarily as a systems and abstraction contribution rather than as a single new Text-to-SQL model.

3. “Squrve” in square-root-velocity curve analysis

In geometric data analysis, the term appears in connection with the square root velocity representation, which is central to elastic shape analysis of curves (Hartman et al., 2021). In this setting one studies parametrized curves $c$ in $\mathbb{R}^d$, while quotienting out transformations that preserve shape, especially translation, rotation, and reparametrization. Curves are modeled as absolutely continuous maps

$c \in \operatorname{AC}(M,\mathbb{R}^d),$

with $M=[0,1]$ for open curves and $M=S^1$ for closed curves (Hartman et al., 2021).

The SRV transform is given by

$Q$0

with

$Q$1

and in standard SRVF notation

$Q$2

with $Q$3 when $Q$4 (Hartman et al., 2021, Lahiri et al., 2015). The transform pulls back the $Q$5 metric to an elastic metric, so the distance between parametrized curves becomes an $Q$6 norm in SRV space:

$Q$7

The actual shape distance must quotient out nuisance actions. Reparametrization acts by composition $Q$8, rotations by $Q$9, and the quotient distance is

$S$0

(Hartman et al., 2021). In equivalent SRVF form,

$S$1

The foundational SRVF theory for piecewise linear curves in $S$2 was developed rigorously in “Precise Matching of PL Curves in $S$3 in the Square Root Velocity Framework” (Lahiri et al., 2015). That work establishes the quotient construction carefully, proves closed-orbit results using the semigroup of weakly increasing reparametrizations, shows that piecewise linear curves correspond to step functions in SRVF space, and gives an exact optimal matching algorithm for PL curves. The optimization is reduced to maximizing inner products over reparametrizations, with local compatibility encoded by the weight matrix

$S$4

for step-function SRVFs $S$5 and $S$6 (Lahiri et al., 2015).

Later work extends SRV methodology beyond Euclidean curves. “Comparing Curves in Homogeneous Spaces” generalizes the framework to curves in $S$7, using horizontal lifts and Lie-algebra-valued square-root-velocity representations (Su et al., 2017). “Elastic Analysis of Augmented Curves and Constrained Surfaces” studies how the SRV transform affects speed and curvature of plane curves and applies the framework to augmented curves and constrained surfaces arising in tubes, ruled surfaces, spherical strips, protein molecules, and hurricane tracks (Nava-Yazdani, 2024).

This literature indicates that, in geometric contexts, “Squrve” is best understood not as a standard formal term but as a colloquial pointer to SRV- or SRVF-based elastic curve analysis.

4. Learned surrogates for SRV quotient distance

A significant recent development in SRV-based analysis is the replacement of expensive test-time variational optimization with supervised learning surrogates. “Supervised deep learning of elastic SRV distances on the shape space of curves” introduces a supervised deep-learning framework for direct computation of SRV distances between curves, thereby avoiding explicit alignment over the reparametrization group and rotations at inference time (Hartman et al., 2021).

The problem addressed is the central computational bottleneck in SRV analysis: minimizing over the infinite-dimensional reparametrization group $S$8, with dynamic programming typically restricted to discretized reparametrizations and thus usually returning an over-estimate of the true quotient distance (Hartman et al., 2021). The exact Lahiri–Robinson–Klassen algorithm computes precise distances for piecewise linear curves but has high polynomial complexity on large datasets (Hartman et al., 2021, Lahiri et al., 2015).

The proposed surrogate takes as input pairs of discretized curves, each sampled at $S$9 vertices and flattened into vectors of length $C$0, and predicts a single scalar: the SRV quotient distance between the corresponding geometric curves (Hartman et al., 2021). The architecture is a Siamese CNN in which each branch uses several 1D convolutional layers with kernel size $C$1, each followed by batch normalization, ReLU, and max pooling of size $C$2, after which the branch outputs are concatenated and passed through four dense layers before a final scalar output (Hartman et al., 2021).

Training uses exact SRV distances for functions and 2D curves, computed with the precise piecewise-linear algorithm, and dynamic-programming distances for 3D curves, where exact labels are too expensive (Hartman et al., 2021). The loss is mean squared error, and training uses Adam. An especially domain-specific feature is shape-preserving data augmentation: because the quotient SRV distance is invariant under reparametrization and rotation, additional training examples are generated by resampling and rotating curves without changing the target label (Hartman et al., 2021).

Reported runtime differences are large. For 1D data, per-distance computation times are $C$3 ms for Exact, $C$4 ms for DP, $C$5 ms for DL on CPU, and $C$6 ms for DL on GPU. For 2D data, the corresponding values are $C$7 ms, $C$8 ms, $C$9 ms, and $M$0 ms (Hartman et al., 2021).

The experiments show that on matched train/test distributions the learned predictor clearly outperforms dynamic programming in 1D mean relative error and correlation against exact distances, and that in adversarially reparametrized 2D data the learned method is more robust than DP (Hartman et al., 2021). However, the paper explicitly notes the absence of approximation bounds, consistency theorems, or guarantees that the learned scalar satisfies metric axioms such as the triangle inequality. The model is therefore a regression surrogate for SRV distance rather than a replacement for SRV geometry itself (Hartman et al., 2021).

5. Squrve as Fernandez–Guasti squircle

In a geometric-design and algebraic-curve context, “Squrve” most commonly refers to the Fernandez–Guasti squircle, a planar algebraic curve that interpolates between a circle and a square (Fong, 2016). It is defined by two parameters: $M$1, a scale or radius parameter, and $M$2, a squareness parameter. When $M$3, the curve is a circle of radius $M$4; when $M$5, it becomes a square of side length $M$6; and for $M$7, it is a smooth intermediate shape (Fong, 2016).

Its canonical Cartesian equation is

$M$8

or equivalently

$M$9

(Fong, 2016). In normalized coordinates $c$0, $c$1, this becomes

$c$2

The paper emphasizes that one usually studies only the central closed component inside

$c$3

because the full quartic curve also has four extraneous infinite branches outside that square (Fong, 2016).

The FG-squircle is algebraically notable because it is a quartic, more specifically a biquadratic, and therefore easier to manipulate than the Lamé curve or superellipse family, which depends on variable exponents (Fong, 2016). The contrast is central: the Lamé family

$c$4

approaches a square only as $c$5, whereas the FG-squircle reaches the square exactly at the finite value $c$6 (Fong, 2016).

The paper derives an area formula using elliptic integrals:

$c$7

where $c$8 and $c$9 are complete Legendre elliptic integrals of the first and second kind, respectively (Fong, 2016). At the endpoints, the area tends to $\mathbb{R}^d$0 for the circle and equals $\mathbb{R}^d$1 for the square.

It also derives a polar form by substituting $\mathbb{R}^d$2, $\mathbb{R}^d$3, yielding

$\mathbb{R}^d$4

(Fong, 2016). Several exact parametrizations are given, including closed parametrizations derived from elliptical-grid mappings and open parametrizations based on “squelch” mappings (Fong, 2016).

The paper makes an explicit caveat concerning arc length: there is no closed-form elementary arc-length formula given; arc length is expressed as an integral and evaluated numerically (Fong, 2016). This distinguishes the curve from cases where low-degree algebraicity yields elementary closed forms for all classical geometric quantities.

6. Mappings, higher-dimensional extensions, and broader significance

A major applied theme of the FG-squircle literature is explicit mapping between canonical domains. The paper derives disc-to-square and square-to-disc transformations, including the closed FG-squircular mapping

$\mathbb{R}^d$5

for normalized variables, together with explicit inverse formulas (Fong, 2016). It also gives an exact elliptical-grid mapping

$\mathbb{R}^d$6

with explicit inverse formulas (Fong, 2016).

These constructions matter because concentric circles in the disc map to concentric squircles in the square, making the curve useful for geometry processing, image warping, texture mapping, coordinate transforms, and explicit homeomorphisms between disc and square (Fong, 2016). The paper extends the same philosophy to three dimensions through the sphube, a sphere–cube hybrid surface defined by

$\mathbb{R}^d$7

(Fong, 2016). As in the planar case, the intended central surface must be clipped to $\mathbb{R}^d$8 because the full algebraic surface contains extraneous lobes.

The same paper derives a radial cube-to-sphere map based on a “squircularity condition” and discusses explicit but algebraically unwieldy inverse formulas for the sphere-to-cube map (Fong, 2016). It further constructs a family of related “squircular quadrics,” including the squircular ellipsoid, sqylinder, squircular cone, squircular cushion, and sphylinder (Fong, 2016).

Taken together, these results show that the Fernandez–Guasti usage of “Squrve” belongs to a family of algebraically tractable intermediates between smooth round domains and axis-aligned box-like domains. By contrast with the SRV and Text-to-SQL usages, this branch of the literature is concerned with exact implicit equations, parametrizations, geometric integrals, and coordinate mappings rather than statistical learning or modular LLM systems.

7. Disambiguation and scholarly usage

Across the arXiv literature, Squrve is therefore not a single stable technical concept. In systems research it is a proper name for a Text-to-SQL framework (Wang et al., 28 Oct 2025). In shape analysis it appears as an informal label associated with the SRV or SRVF approach to elastic curve comparison (Hartman et al., 2021), whose theoretical foundations were formalized for absolutely continuous and piecewise linear curves (Lahiri et al., 2015) and later generalized to homogeneous spaces and augmented curves (Su et al., 2017, Nava-Yazdani, 2024). In algebraic geometry and graphics-oriented literature it denotes, or is used to point toward, the Fernandez–Guasti squircle and its three-dimensional generalizations (Fong, 2016).

A common misconception is that these usages share a substantive conceptual connection because of the similar spelling. The available evidence indicates otherwise. The SRV literature and the Fernandez–Guasti squircle literature are mathematically unrelated beyond the accidental resemblance of names, while the Text-to-SQL framework is a software architecture contribution in NLP and database systems rather than a curve-analysis method or a geometric shape.

For precise scholarly communication, the term is best expanded on first use. In practice this means specifying Squrve (Text-to-SQL framework) (Wang et al., 28 Oct 2025), SRV/SRVF elastic shape analysis (Hartman et al., 2021, Lahiri et al., 2015), or Fernandez–Guasti squircle (Fong, 2016), depending on context. This suggests that “Squrve” functions more as a homographic label spanning independent research traditions than as a unified technical term.