Spatial Label Arbitration

• Spatial label arbitration is a set of decision mechanisms that reconcile competing labels by enforcing spatial, geometric, and semantic constraints such as non-overlap and proper ordering.
• Methods employ dynamic programming, geometric de-confliction, and cost-sensitive selection to optimize label placement in diverse scenarios like boundary labeling, cartography, and 3D labeling.
• These techniques are crucial for ensuring clarity and consistency in annotations across applications including AR, road mapping, and weakly supervised segmentation.

Spatial label arbitration denotes a family of decision mechanisms that resolve conflicts among competing labels, label candidates, or label-like supervisory signals under spatial, geometric, or multimodal constraints. Across the literature, it appears in at least three closely related senses: as explicit ordering and grouping policies in boundary labeling for diagrams; as geometric de-confliction and candidate selection in cartographic, road, AR, and 3D point labeling; and as a supervision or representation-level arbitration process in machine learning systems, where spatial structure determines which label, class, or modality should dominate a final assignment (Depian et al., 2024). In all cases, the core operation is the same: a set of potentially incompatible assignments is filtered, ranked, or reconciled so that the result satisfies domain-specific constraints such as non-overlap, non-crossing, contiguity, precedence, or object-level consistency.

1. Conceptual scope and recurring structure

A common formulation treats arbitration as constrained selection over labels or label surrogates. In one-side boundary labeling, labels are indexed by anchor coordinates $y_i$ on a boundary side, with non-overlap enforced by $|y_i-y_j|\ge h$ for uniform-height labels, grouping expressed as contiguity of a group in the boundary permutation, and ordering expressed as $y_i+\delta\le y_j$ for each $(i,j)\in E_{PO}$ (Depian et al., 2024). In dynamic point labeling, arbitration selects at most one candidate per point, subject to pairwise non-overlap constraints, while minimizing a candidate’s “expense,” defined as the sum of the modified values of all neighbors that would be occluded by choosing that candidate (Mote, 2012). In weakly incremental segmentation, arbitration selects a single class per mask by geometry-weighted voting under a candidate restriction rule determined by a novel activation density threshold, then projects that class to all pixels in the mask to enforce “one object, one class” (Wang et al., 2 Jun 2026).

These formulations differ in representation but share a recurring decomposition into four elements. First, there is a discrete or continuous set of candidates: ports on a boundary, corner placements on a map, object-mask labels in segmentation, or per-layer visual/prior token alternatives in VLMs. Second, there is a conflict model, such as leader crossings, rectangle overlaps, contradictory pseudo-labels, or visual-linguistic competition. Third, there is an arbitration policy, ranging from exact feasibility constraints and dynamic programming to weighted voting, geometric de-confliction, or activation steering. Fourth, there is an objective or decision criterion, such as minimizing total leader length, minimizing conflict cost, maximizing labeled road sections, or favoring visual-consistent logits over prior-consistent logits (Nooralahzadeh et al., 10 Apr 2026).

The surveyed work also distinguishes arbitration from adjacent notions. In VLM analysis, perception or encoding refers to whether hidden states carry information about image content, grounding refers to whether the final output reports what is depicted, and arbitration denotes the internal competition between visual evidence and language priors that determines which signal dominates the final readout (Nooralahzadeh et al., 10 Apr 2026). In incremental segmentation, arbitration is not a regularizer on predictions after training but a training-time supervision denoising mechanism applied before gradients are computed (Wang et al., 2 Jun 2026). In map and diagram labeling, arbitration is not merely heuristic decluttering; it is an explicit spatial decision process over candidate placements under combinatorial or geometric constraints (Mote, 2012).

2. Boundary labeling and explicit semantic arbitration

The most formal use of spatial label arbitration appears in constrained boundary labeling. Boundary labeling places text rectangles along an axis-parallel bounding box and connects each site to its label with a non-crossing leader; the 2024 constrained formulation introduces grouping constraints, requiring all labels in a group to be consecutive on the boundary, and ordering constraints, imposing a partial order over labels (Depian et al., 2024). These rules are explicitly described as “spatial label arbitration rules”: grouping decides adjacency and block formation along the boundary, while ordering sets precedence.

The one-side problem admits a precise constraint language. Let $\pi$ be the permutation of labels induced by increasing anchor coordinates. Group contiguity is stated as: for each group $G_g$, there exist indices $a_g\le b_g$ such that $\{\pi(a_g),\ldots,\pi(b_g)\}=G_g$. Ordering is given by $y_i+\delta\le y_j$ for all $(i,j)\in E_{PO}$. Feasibility requires label non-overlap, no leader–leader crossings, and no leader–site crossings, with po-leaders providing a monotone orthogonal model in which order-preserving routing suffices to avoid crossings (Depian et al., 2024).

The paper’s main algorithmic contribution is a dynamic programming framework for one-side instances combined with a PQ-A-Graph encoding of grouping and ordering constraints. Grouping constraints are converted to the consecutive ones property via a sites-versus-groups matrix and represented by a PQ-Tree; ordering constraints are represented as directed arcs on the leaves, producing a PQ-A-Graph $|y_i-y_j|\ge h$0. A consistency check called Reorder decides whether some permissible $|y_i-y_j|\ge h$1-order extends the arc set in $|y_i-y_j|\ge h$2 time. The dynamic program decomposes an instance by routing a leader from the leftmost site to a candidate port, recursively splitting the remaining sites into two independent sub-instances, with recurrence

$|y_i-y_j|\ge h$3

For fixed ports and arbitrary label sizes, this yields runtime $|y_i-y_j|\ge h$4 and space $|y_i-y_j|\ge h$5; for sliding ports with uniform-height labels, canonical-port discretization reduces the problem to the fixed-port case, giving runtime $|y_i-y_j|\ge h$6 and space $|y_i-y_j|\ge h$7 (Depian et al., 2024).

The complexity boundary is sharp. One-side boundary labeling with sliding ports and arbitrarily sized labels is weakly NP-hard by reduction from Partition, even with a constant number of constraints. Two-side constrained boundary labeling on opposite sides is NP-complete even for uniform-height labels and fixed candidate positions, via a reduction from Monotone 1-in-3 SAT in which ordering constraints transmit global assignment through side selection (Depian et al., 2024). This result is significant because it shows that semantic precedence and contiguity constraints are not peripheral embellishments; they fundamentally alter the computational character of labeling.

A closely related extension is orbital boundary labeling on a circular contour. Here labels are circular arcs in an annulus-shaped orbit around a central circle, and the objective is to minimize total leader length under noncrossing constraints (Bonerath et al., 2024). The paper defines a COSA parameterization over candidate ports, label order, label size, and port ratio, and establishes polynomial-time algorithms for several uniform cases, including $|y_i-y_j|\ge h$8 for free order with uniform labels and uniform ratios via reduction to po-Boundary Labeling. It also proves weak NP-hardness for free-order, free-port, nonuniform-size variants via a reduction from Partition (Bonerath et al., 2024). Taken together, these works show that spatial label arbitration in geometric labeling is the explicit mediation between semantic rules and geometric realizability.

3. Cartographic, road, 3D, and AR placement as geometric arbitration

In cartography and visualization, spatial label arbitration is typically realized as geometric de-confliction plus cost-sensitive selection. A prominent dynamic formulation is the trellis-based point-feature label placement method for real-time interactive maps. Each point has four fixed-size corner candidates, conflict pairs are defined by strict interior rectangle overlap, and arbitration proceeds greedily in descending priority order by selecting the least-expensive feasible candidate and removing its conflict partners (Mote, 2012). The expense of candidate $|y_i-y_j|\ge h$9 is

$y_i+\delta\le y_j$0

where $y_i+\delta\le y_j$1 is the set of conflicting neighbor candidates removed if $y_i+\delta\le y_j$2 is chosen. The method uses a screen-space grid whose cells have size $y_i+\delta\le y_j$3, $y_i+\delta\le y_j$4, so any conflict must lie within a bounded $y_i+\delta\le y_j$5 neighborhood. This yields expected $y_i+\delta\le y_j$6 per frame for uniform distributions and supports multiple frames per second for tens of thousands of points (Mote, 2012).

Road-map labeling imposes a more structured objective: maximize the number of labeled road sections while keeping labels embedded in roads and non-overlapping. The framework of Gemsa, Niedermann, and Nöllenburg preprocesses a road network into a planar graph of road sections and junctions, generates readable label candidates aligned to centerlines, and arbitrates among incompatible placements either with an exact MILP or with dynamic programming on trees (Niedermann et al., 2016). The unweighted MILP maximizes $y_i+\delta\le y_j$7, where $y_i+\delta\le y_j$8 indicates whether road section $y_i+\delta\le y_j$9 is labeled. On 11 major cities, the proposed method labels 31% more road sections on average than Mapnik, and the decomposition-plus-tree heuristic achieves on average 97% of MILP’s optimal coverage while remaining much faster in practice (Niedermann et al., 2016). This use of arbitration is less about semantic ordering than about maximizing informational coverage under strong geometric embedding constraints.

Three-dimensional labeling with leaders introduces continuous displacement rather than discrete candidate choice. The Beams Displacement Method models label centers as nodes in a pruned Delaunay-triangulation graph and treats the graph as an elastic beam network. The optimization is performed in 2D screen space after projection, with forces encoding label-label repulsion, label-point repulsion, leader attachment attraction, and screen-edge constraints (Wei et al., 2024). A representative energy decomposition includes

$(i,j)\in E_{PO}$0

plus penalty terms for minimum distances, edge proximity, leader attachment, and displacement. Solving $(i,j)\in E_{PO}$1 yields cooperative adjustments that arbitrate between overlap removal and preservation of directional relations. On AIS and POI datasets, the method eliminates conflicts and yields substantially lower average directional deviation than a local heuristic, at the cost of higher runtime (Wei et al., 2024).

Augmented-reality labeling in cluttered scenes uses a simpler but task-oriented arbitration language. Three strategies are evaluated: Situated Individual (one label per item), Situated Grouped (one label per spatial group placed above the group), and Centered Grouped (one label per group placed at the centroid), with grouping performed by DBSCAN over same-type items (Park et al., 30 Jun 2025). Arbitration therefore includes the decision of whether to show individual or representative labels, where to anchor them, and how to route leaders. In a study with 15 participants and 50 in-view items, grouped conditions were significantly faster than per-item labeling: SG yielded $(i,j)\in E_{PO}$2, CG $(i,j)\in E_{PO}$3, and SI $(i,j)\in E_{PO}$4, with no significant accuracy difference (Park et al., 30 Jun 2025). This suggests that arbitration can optimize for task performance by reducing label count through controlled grouping, not only by enforcing non-overlap.

A recent LLM-based approach to map labeling reframes automatic label placement as structured data editing. MAPLE pairs map tiles, landmark geometry, and retrieved NGA DPS guideline sections, and instruction-tuned LLMs output a single pixel coordinate for each landmark (Shomer et al., 29 Jul 2025). The paper evaluates zero-shot and QLoRA-tuned models, reporting large reductions in centroid RMSE after tuning; for example, Phi-4 achieves 28.4 RMSE with List formatting, while Gemma2 achieves 30.1 with JSON (Shomer et al., 29 Jul 2025). The method relies on guideline retrieval rather than explicit geometric post-processing. A plausible implication is that this line of work treats arbitration as context-conditioned synthesis of a single coordinate, whereas classical cartographic methods treat it as explicit combinatorial conflict resolution.

4. Supervision denoising and object-level consistency in segmentation and geospatial correction

In weakly incremental semantic segmentation, spatial label arbitration is a geometry-aware supervision denoising module. SASA introduces Semantic Anchors and Spatial Arbitration for Weakly Incremental Learning for Semantic Segmentation, where contradictory pseudo-labels arise because CAM seeds for novel classes cover only discriminative parts while a frozen old model fills uncovered regions with old-class predictions (Wang et al., 2 Jun 2026). The arbitration mechanism operates on class-agnostic object masks $(i,j)\in E_{PO}$5, computes a Gaussian reliability weight

$(i,j)\in E_{PO}$6

and forms a novel activation density

$(i,j)\in E_{PO}$7

If $(i,j)\in E_{PO}$8, candidate classes are restricted to $(i,j)\in E_{PO}$9; otherwise they are restricted to $\pi$0. A weighted majority vote

$\pi$1

assigns a single label to the mask, and all pixels in the mask are projected to that label, enforcing “one object, one class” (Wang et al., 2 Jun 2026).

This is an explicit arbitration policy between old and novel supervision sources. It neither adds a separate loss nor modifies inference; rather, it rectifies pseudo-labels before training. The empirical effect is substantial: on VOC 10–5 with ViT-B/16, adding SLA alone improves All mIoU from 64.1 to 67.6, and full SASA reaches 71.7; the threshold ablation reports the best overall mIoU at $\pi$2 (Wang et al., 2 Jun 2026). The method’s central claim is that geometry-aware arbitration reduces contradictory gradients at the source.

A distinct but related use appears in historical map annotation correction. The Label Correction Algorithm resolves conflicts between raster-map color evidence and vector-derived priors when automatic overlap generates misaligned and false labels (Duan et al., 2021). The method combines a Chan–Vese energy,

$\pi$3

with a shape-overlap penalty,

$\pi$4

where $\pi$5 is an affine-transformed prior shape derived from external vector data (Duan et al., 2021). Arbitration here consists of accepting foreground labels when color and shape terms agree, rejecting false annotations when no suitable foreground is found, and adjusting misaligned labels through affine transformation parameters. The reported results state that annotation precision is 10% higher than a state-of-the-art algorithm and that recognition correctness improves by 9% (Duan et al., 2021).

Both cases treat spatial label arbitration as adjudication among conflicting supervision sources rather than among display labels. This broadens the concept: a “label” need not be a visible annotation; it can be a pseudo-label, mask label, or training target whose spatial coherence must be enforced.

5. Arbitration inside vision-language and image-labeling systems

In multimodal models, spatial label arbitration denotes competition between encoded evidence and prior-driven outputs. The VLM study on “Arbitration Failure, Not Perceptual Blindness” formalizes this distinction by defining visual-consistent and prior-consistent token sets and measuring the layerwise logit gap

$\pi$6

with the Multimodal Arbitration Crossover (MAC) defined as the first stable layer where $\pi$7 becomes positive and remains positive at the next layer (Nooralahzadeh et al., 10 Apr 2026). The central empirical result is an Encoding–Grounding Dissociation: visual attributes are linearly decodable very early, with AUC $\pi$8 at about 10% depth and near 1.0 by mid-depth, even for samples that ultimately answer incorrectly. By contrast, the final-layer visual–prior logit gap predicts grounding outcomes with Spearman $\pi$9, while encoding strength shows $G_g$0 (Nooralahzadeh et al., 10 Apr 2026).

Causal analysis reinforces the arbitration interpretation. Full-sequence activation patching at MAC-identified layers flips 60–84% of outputs, whereas last-token patching flips only 0–1%; image-only patching retains 97–101% of the full effect in several 7–8B models and 100% in scaled variants, while text-only patching has essentially no effect (Nooralahzadeh et al., 10 Apr 2026). Training-free steering in early layers improves visual grounding by up to +3.8%, especially when guided by a sparse autoencoder. The paper does not directly test spatial relations, but it proposes a concrete extension to spatial labels such as left/right and above/below, using the same MAC, patching, and steering protocol (Nooralahzadeh et al., 10 Apr 2026). This suggests that spatial label arbitration in VLMs is not primarily about whether relations are represented, but about whether late-layer decision dynamics favor spatial evidence over priors.

A mathematically different, but structurally related, formulation appears in geometric image labeling. “Image Labeling by Assignment” defines an assignment matrix $G_g$1 whose rows lie in the simplex interior and optimizes

$G_g$2

where $G_g$3 is a Riemannian mean of neighbor likelihoods on the simplex manifold (Åström et al., 2016). The induced Riemannian gradient flow yields row-wise replicator equations

$G_g$4

and the practical multiplicative update

$G_g$5

Here arbitration is the reconciliation of competing per-site label supports through geometric averaging and multiplicative competition. The method is not about visible label placement, but it still fits the general pattern: spatial structure couples local decisions so that a coherent labeling emerges (Åström et al., 2016).

6. Interaction graphs, coded paths, and broader extensions

Several additional literatures use the same concept in more specialized ways. In labeled multi-object tracking with the GLMB filter, spatial label arbitration is the process of deciding which object labels statistically interact during the measurement update. Labels interact when their predicted measurement regions overlap, forming an interaction graph

$G_g$6

whose connected components define update groups (Lee et al., 2023). The two-layer partitioning method introduces a coarse grid plus four secondary tile classes to enumerate intersecting label pairs without duplicate generation, enabling scalable graph construction on datasets with 10K–100K labels while leaving tracking accuracy unchanged (Lee et al., 2023). Arbitration here is group formation for conditional independence rather than spatial layout.

In RIS-based wireless sensing, labeling refers to deliberate temporal coding of propagation paths via reconfigurable intelligent surfaces. Paths traversing RIS $G_g$7 receive gains

$G_g$8

and spatial label arbitration is the association step that assigns extracted paths to the most likely RIS code by least-squares residual minimization over known sequences $G_g$9 (Wang et al., 2021). This is a non-visual use of the term, but it preserves the same structure: labels are competing identifiers tied to spatial routes, and arbitration resolves which route carries which label.

A more weakly supervised classification setting uses temporal and spatial consistency as arbitration between noisy single-label supervision and latent multi-label evidence. Spatial Consistency Loss maintains per-image, per-class running-average heatmaps and enforces consistency with current spatial outputs after geometric augmentation transforms:

$a_g\le b_g$0

Combined with Expected-Negative masking and temporal consistency, this improves multi-label mAP over binary cross-entropy baselines on MS-COCO, Pascal VOC, and ImageNet ReaL (Verelst et al., 2022). Although the paper does not use the phrase for visible label placement, its mechanism plainly arbitrates between region-level evidence accumulated across crops and epochs.

Across these domains, several limitations recur. In constrained boundary labeling, tractability hinges on one-side structure, uniform height, or fixed ports, while two-side constrained instances are NP-complete (Depian et al., 2024). In segmentation, arbitration quality depends on the quality of class-agnostic masks; over-merged masks can propagate incorrect decisions across an object (Wang et al., 2 Jun 2026). In VLMs, last-token interventions are ineffective because visual evidence is distributed across image tokens, so successful arbitration control requires full-sequence or image-token interventions (Nooralahzadeh et al., 10 Apr 2026). In AR and 3D labeling, grouping or block processing improves efficiency but can introduce trade-offs between leader length, local clutter, and preservation of adjacency relations (Park et al., 30 Jun 2025).

This body of work suggests a unifying interpretation. Spatial label arbitration is not a single algorithm but a recurrent design problem that arises whenever labels, label surrogates, or label-conditioned decisions must be made coherent under spatial structure. In computational geometry, this means contiguity, precedence, and crossing-free routing; in cartography and AR, non-overlap, visibility, and task effectiveness; in segmentation and weak supervision, object-level consistency and geometry-aware denoising; and in VLMs, the causal selection of spatial evidence over priors. The surveyed papers therefore define a technically diverse but conceptually consistent field of methods for resolving competing labeled interpretations in space (Depian et al., 2024).