Locator Function Mechanisms

• Locator function is a mechanism that converts raw evidence from language, images, or sensor streams into precise positions, probability fields, or stability indicators.
• It operates across domains—such as GUI grounding, dynamic optimization, and inertial localization—by constraining search spaces and ensuring actionable localization decisions.
• Its implementations vary from direct coordinate prediction in multimodal AI to probabilistic uncertainty estimates and physical subsystem functions in high-precision detectors.

In technical literature, “locator function” does not denote a single universal object. In the cited work, it names a family of constructs whose shared role is to determine where something is, where it should be, or which candidate locations are admissible. The term is used for an inference-time mapping from language and images to click coordinates, a scalar function whose roots identify interior steady states, a positive landscape on a Fock-space graph, learned probabilistic mappings from sensor streams to source or trajectory location, precision detector subsystems that reconstruct vertices near an interaction point, and explicit selection structures such as DOM selectors, Layer-2 locators, or pixel-index sequences (Li et al., 29 Sep 2025, Huang et al., 3 Sep 2025, Balasubramanian et al., 2019, Group et al., 2014).

1. Conceptual scope

Across the cited literature, a locator function is best understood as a position-resolving or target-resolving mechanism whose output is directly operational. In some settings it returns a coordinate, in some it returns a likelihood field or a displacement distribution, in some it returns a root set or stability information, and in some it is embodied as a physical subsystem or a metadata structure rather than as a single analytic formula. What remains common is that the locator constrains a search space and converts raw evidence into a localization decision or a localization-compatible representation (Bagagli et al., 14 Jul 2025, Wu et al., 8 Aug 2025, 0803.4311).

Domain	Locator object	Operational role
GUI grounding	$GUIGround(Q,\text{crop})$	Predict click point $(x,y)$
Dynamic optimization	$L(x)$	Roots identify steady states
Many-body localization	$u$ with $(\mathcal H+K)|u\rangle=|1\rangle$	$1/u$ acts as effective potential
Seismology	$XY,XZ,YZ$ likelihood maps	Infer source location
Inertial localization	$f_\theta:X^g\mapsto(\Delta \hat p,\hat b)$	Relative displacement and uncertainty
Networking / testing / steganography	BigMAC, DOM locator, PLS	Select path, element, or pixels

This diversity is not merely terminological. It reflects different mathematical regimes: point prediction, root finding, graph-localization, Bayesian state propagation, and symbolic addressing. In each case, however, the locator is the component that makes localization actionable rather than descriptive.

2. Coordinate-prediction locators in multimodal AI

In GUI grounding, the locator appears as an explicit inference-time function. GMS decomposes the task into a generalist “Scanner” and a specialist “Locator,” with the latter defined by $\hat{p}_l = GUIGround(Q, R^{(l)})$. The Locator takes a natural-language instruction $Q$ and a cropped image region $(x,y)$0, and outputs a pixel coordinate $(x,y)$1. Its formal role is narrow but precise: it does not perform global search, recursive subdivision, or semantic arbitration; it predicts exact coordinates within regions selected by the Scanner. This division is quantitatively consequential on ScreenSpot-Pro: Scanner alone reaches $(x,y)$2, Locator alone $(x,y)$3, while integrated GMS reaches $(x,y)$4, described as a $(x,y)$5 improvement (Li et al., 29 Sep 2025).

The same specialization logic appears in video object removal, though there the Locator is not a standalone segmentation head. In GenEraser, the Locator is the high-noise expert in a decoupled diffusion architecture, trained on multi-source data for fewer optimization steps so that it can “identify and eliminate target objects and effects,” including weakly correlated effects such as smoke, reflections, light, and ripples. Its counterpart, the Preserver, is the low-noise expert responsible for faithful background preservation. The paper explicitly frames this split as a response to an optimization conflict between semantic generalization and pixel-level preservation (Chen et al., 28 May 2026).

In visual semantic localization for autonomous driving, BEV-Locator turns the locator into an end-to-end pose-correction network. Multi-view images are encoded into BEV features, semantic map elements are embedded as map queries, and a cross-model transformer recursively queries localization information through cross-attention. The output is a $(x,y)$6-DoF correction $(x,y)$7, combined with the initial pose as

$(x,y)$8

Reported mean absolute errors reach $(x,y)$9, $L(x)$0, and $L(x)$1 on Qcraft (Zhang et al., 2022).

3. Scalar locator functions and localization landscapes

In dynamic optimization, the locator function is a scalar object constructed directly from the model’s primitives. Its zeros identify interior steady states, and its slope determines local stability. Under strict concavity, it also characterizes basins of attraction; without concavity, it still yields strong shape-based results, including the statement that if the locator function is single crossing from above, its root identifies a globally stable steady state, while if it is inverted-U-shaped with two interior roots, only the higher root can be locally stable. The summarized canonical one-control one-state form is

$L(x)$2

and steady-state comparative statics are obtained from

$L(x)$3

The method is explicitly presented as a way to characterize qualitative dynamics without solving the full dynamic program (Huang et al., 3 Sep 2025).

A different formalization appears in many-body localization theory. There, the object playing the role of a locator function is the many-body localization landscape $L(x)$4, defined on the nodes of the Fock-space graph by

$L(x)$5

This function is positive, its inverse $L(x)$6 acts as an effective potential on the Fock-space graph, and it yields the bound

$L(x)$7

The paper further shows that the landscape admits a locator expansion over Fock-space paths and emphasizes a key distinction from the conventional Green-function locator expansion: the landscape expansion contains no resonances. In this setting, the locator function is not a coordinate predictor but a graph-defined field that delimits classically allowed and forbidden regions in Fock space and supports Agmon-type decay bounds (Balasubramanian et al., 2019).

4. Probabilistic locators in seismology and inertial navigation

HEIMDALL embeds localization inside a joint detector–picker–associator–locator architecture operating directly on continuous multi-station waveform windows. The model ingests $L(x)$8 windows with $L(x)$9 stride on a static station graph, and outputs station-specific event, P, and S probability traces together with three dense $u$0D source-likelihood images on the $u$1, $u$2, and $u$3 planes. At inference time, the coordinate head is not used directly; the event location is taken from the peak of the likelihood images, and the image itself is retained as a representation of epistemic uncertainty. Window-level hypotheses are filtered by a $u$4 3D consistency criterion, a minimum plane-peak threshold of $u$5, and a compactness threshold of $u$6, then stitched across consecutive windows. On the February 3, 2019 sequence, HEIMDALL detected $u$7 events versus $u$8 in the manual catalog, a $u$9 increase, with mean and median hypocentral error both about $(\mathcal H+K)|u\rangle=|1\rangle$0 (Bagagli et al., 14 Jul 2025).

ReNiL treats localization as relative motion inference between Inertial Positioning Demand Points rather than dense fixed-rate tracking. For an aligned IMU segment $(\mathcal H+K)|u\rangle=|1\rangle$1, the locator is the mapping

$(\mathcal H+K)|u\rangle=|1\rangle$2

with the displacement modeled as

$(\mathcal H+K)|u\rangle=|1\rangle$3

Any-scale inference is enabled by patchification with

$(\mathcal H+K)|u\rangle=|1\rangle$4

followed by pooling over a variable number of patches. A Bayesian recursion then chains successive IPDP estimates into a trajectory. The framework emphasizes “homogeneous Euclidean uncertainty,” because the Laplace scale parameter lives in the same coordinate space as the displacement. On WUDataset, ASLE-20s reaches MAE $(\mathcal H+K)|u\rangle=|1\rangle$5 on seen subjects and ASLE-10s/20s reaches $(\mathcal H+K)|u\rangle=|1\rangle$6 on unseen subjects, while the motion-aware orientation filter also outperforms Mahony and Madgwick in QAE (Wu et al., 8 Aug 2025).

These two systems illustrate a common probabilistic pattern. The locator does not merely output a point; it outputs a structured uncertainty-bearing object—likelihood images in one case, Laplace displacement parameters in the other—and localization is completed by an explicit temporal consistency or Bayesian chaining stage.

5. Physical locator systems in particle physics and safety engineering

In high-energy physics, “locator” can name a detector subsystem whose practical function is precise spatial reconstruction. The LHCb Vertex Locator (VELO) is the silicon microstrip detector surrounding the proton–proton interaction region. Its role is to measure charged-particle trajectories close to the beam so that primary vertices, displaced decay vertices, and impact parameters can be reconstructed with high precision. During physics running, the sensors move to $(\mathcal H+K)|u\rangle=|1\rangle$7 from the beam; the first active silicon is at $(\mathcal H+K)|u\rangle=|1\rangle$8. Reported performance includes a signal-to-noise ratio of approximately $(\mathcal H+K)|u\rangle=|1\rangle$9, a best hit resolution of $1/u$0, track finding efficiency above $1/u$1, alignment precision of $1/u$2 for transverse translations, primary-vertex resolution of $1/u$3 in the transverse plane and $1/u$4 along the beam axis for vertices with $1/u$5 tracks, and impact-parameter resolution below $1/u$6 for particles with transverse momentum greater than $1/u$7 (Group et al., 2014).

That locator function is dynamic rather than static, because radiation damage changes the operating conditions. Radiation studies report an average sensor-current increase of $1/u$8 per $1/u$9 at approximately $XY,XZ,YZ$0, an effective silicon bandgap of $XY,XZ,YZ$1, and the first observation at the LHC of $XY,XZ,YZ$2-on-$XY,XZ,YZ$3 sensor type inversion around $XY,XZ,YZ$4. Material-localization studies then refine the detector’s effective geometry by reconstructing secondary hadronic interactions and computing a compatibility statistic

$XY,XZ,YZ$5

which is converted into a $XY,XZ,YZ$6-value for the hypothesis that a secondary vertex originated in material rather than in a genuine displaced decay (Affolder et al., 2013, Alexander et al., 2018).

The upgrade path extends the same locator function into a 4D regime. For the 2030s LHCb Upgrade II, a new VELO based on 4D hybrid silicon pixels is proposed, with a target timing resolution of $XY,XZ,YZ$7 per track and $XY,XZ,YZ$8 per hit. Two scenario anchors are given: $XY,XZ,YZ$9 inner radius at $f_\theta:X^g\mapsto(\Delta \hat p,\hat b)$0, and $f_\theta:X^g\mapsto(\Delta \hat p,\hat b)$1 at $f_\theta:X^g\mapsto(\Delta \hat p,\hat b)$2 (Gkougkousis, 2023).

A different physical interpretation appears in aviation safety. There the locator function is the post-accident ability to detect, identify, and localize an aircraft for search and rescue. Emergency Locator Transmitters implement this as autonomous radio beacons with five subsystems: autonomous power supply, sensing and activation unit, control and processing unit, RF transmission module, and antenna system. The review emphasizes the transition from legacy $f_\theta:X^g\mapsto(\Delta \hat p,\hat b)$3 analog beaconing to digitally encoded $f_\theta:X^g\mapsto(\Delta \hat p,\hat b)$4 transmission, with $f_\theta:X^g\mapsto(\Delta \hat p,\hat b)$5 retained as a homing signal in combined units. It also notes that Cospas-Sarsat satellite processing of $f_\theta:X^g\mapsto(\Delta \hat p,\hat b)$6 ended on 1 February 2009, and discusses LEOSAR, GEOSAR, and MEOSAR as the space segments that mediate localization performance. Endurance requirements are generally $f_\theta:X^g\mapsto(\Delta \hat p,\hat b)$7 to $f_\theta:X^g\mapsto(\Delta \hat p,\hat b)$8 hours, while second-generation distress-tracking beacons have a minimum operating lifetime of $f_\theta:X^g\mapsto(\Delta \hat p,\hat b)$9 (Martínez-Heredia et al., 10 Mar 2026).

6. Addressing, querying, and search metadata

In networking, the locator function is an addressing role. The Layer-2 locator/identifier split proposal redefines MAC addresses as pure topology-dependent locators while IP addresses become topology-independent identifiers. A “BigMAC” address of the form

$\hat{p}_l = GUIGround(Q, R^{(l)})$0

encodes a downward path from a top-tier switch to a target device, and the next byte $\hat{p}_l = GUIGround(Q, R^{(l)})$1 selects a downlink branch. The proposal even allocates $\hat{p}_l = GUIGround(Q, R^{(l)})$2 bits for downlink port number and $\hat{p}_l = GUIGround(Q, R^{(l)})$3 bits for uplink number when switches have $\hat{p}_l = GUIGround(Q, R^{(l)})$4–$\hat{p}_l = GUIGround(Q, R^{(l)})$5 ports and fewer than $\hat{p}_l = GUIGround(Q, R^{(l)})$6 uplinks. In this sense, the locator is neither a coordinate nor a probability field; it is a path-coded attachment point within a tiered Ethernet fabric (0803.4311).

In automated web GUI testing, a locator is the mechanism a test uses to find a web element in the DOM so that it can inspect or interact with it. ReproBreak formalizes a locator change as $\hat{p}_l = GUIGround(Q, R^{(l)})$7, a locator break as $\hat{p}_l = GUIGround(Q, R^{(l)})$8, and a non-breaking locator change as $\hat{p}_l = GUIGround(Q, R^{(l)})$9, with

$Q$0

Across $Q$1 repositories, the study identified $Q$2 locator changes, and in the four projects that were made reproducible it validated $Q$3 changes and confirmed $Q$4 locator breaks. Here the locator function is operationally the string argument supplied to APIs such as cy.get(...) or page.locator(...) (Moura et al., 12 May 2026).

In image steganography, the locator is an explicit side structure. The Pixel Locator Sequence is a randomly generated or manually created sequence of pixel numbers specifying where encrypted payload bits are written. It is described as “a randomly generated pixel sequence” that “will act as a key during the decoding process,” and the scheme uses

$Q$5

pixels for $Q$6 encrypted characters. The PLS itself is then encrypted with AES and sent alongside the stego image, so the locator function is an encrypted list of pixel positions rather than an embedded estimator (Gangurde et al., 2020).

In gravitational-lens discovery, the term shifts again from object to survey pipeline. MNELLS uses MUSE integral-field spectroscopy to locate background emission-line sources behind nearby massive early-type galaxies by subtracting continuum with a $Q$7 running median, smoothing the residual cube, and searching $Q$8 spectral slices with connected-component labeling. The survey reports one confirmed galaxy-scale strong lens, three cluster-scale lensing systems, and nine further galaxies with singly imaged but closely projected sources within $Q$9 arcsec. Injection-and-recovery tests predict that one in twenty-five observations is expected to yield a new strong-lens system (Collier et al., 2020).

Taken together, these usages show that “locator function” is a role-class rather than a single formula. It may be a coordinate regressor, a scalar equilibrium locator, a graph-defined landscape, a probabilistic source localizer, a detector subsystem, or a symbolic address or selector. The unifying feature is not mathematical form but operational purpose: the locator is the component that turns raw structure, evidence, or topology into a usable statement of where a target, state, event, packet, or payload is to be found.