Position-Consistent Accuracy

• Position-consistent accuracy is a framework that measures the invariance in system outputs against positional variations, ensuring uniform reliability across applications.
• It uses techniques like permutation testing, dynamic sampling, and curriculum-driven training to mitigate positional bias and enhance evaluation metrics.
• Empirical assessments in LLM ranking, sensor localization, and biomedical mapping show significant improvements in accuracy consistency and error reduction.

Position-consistent accuracy is a conceptual and quantitative framework for evaluating the stability and reliability of position-dependent outputs in both machine learning and physical measurement systems. It measures the degree to which a system delivers invariant or uniform accuracy with respect to explicit position variables or the arrangement of content, across diverse settings ranging from LLM ranking, sensor localization, rubrics-based evaluation, deep image editing, modular arithmetic under format shift, and biomedical coordinate systems. This notion generalizes beyond average-case accuracy to directly probe invariance under transformations, ordering, spatial displacement, or other positional perturbations.

1. Formal Definitions and Core Metrics

Position-consistent accuracy quantifies the agreement between a system's predictions and reference values (e.g., gold labels, physical ground truth, human judgments), controlling for or averaging over all positional configurations. The specific operationalization depends on the domain:

• Pairwise LLM ranking: Given a prompt ordering $(a,b)$, an LLM returns $J(a,b) \in \{a,b\}$; position-consistent accuracy evaluates the consensus prediction after aggregating results over all orderings (e.g., $(a,b)$ vs $(b,a)$), typically by majoritarian aggregation across repeated and permuted prompts. The metric is generally computed as normalized accuracy versus the consensus standard, $Acc_{\text{method}}/Acc_{\text{consensus}}$ (Vardasbi et al., 23 Jul 2025).
• Rubric-based LLM-as-a-judge: For $K$-option rubrics, per-position accuracy $A_p$ (fraction correct when the correct label occupies position $p$), averaged over all positions, gives the mean position-consistent accuracy $$\mathrm{PCA}_{\rm mean} = \bar A = \frac{1}{K}\sum_p A_p$$, with a consistency-penalized version $\mathrm{PCA}_{\rm cp} = \bar A - \sigma_A$ using the standard deviation $\sigma_A$ of $A_p$ to penalize position bias (Xu et al., 2 Feb 2026).
• Sequential LLM interactions: Position-weighted consistency (PWC) evaluates model correctness across $T$ dialogue rounds using exponentially decaying weights: $$PWC_\gamma(c_1, ..., c_T) = \sum_{i=1}^T \gamma^{i-1} c_i$$ with $0 < \gamma < 0.5$. This captures early-round resilience and rewards recovery/stability (Li et al., 28 Mar 2025).
• ML robustness to input position shift: The position-consistent accuracy $\mathrm{Acc}_B$ is calculated by averaging per-position accuracies over different absolute input shifts: $$\mathrm{Acc}_B = \frac{1}{|P|}\sum_{t\in P} \mathrm{Accuracy}_B(t)$$ (Yudin, 7 Jan 2026).
• Geospatial/sensor systems: In position estimation, position-consistent accuracy is linked to mean error and variance across the spatial domain. Statistical assessment uses root-mean-square error, standard deviation, and confidence regions (e.g., minimum distance of reported points to ground-truth route shapes in real-time transit data (Wong, 6 Jun 2025), or sub-cm error quantiles in mmWave positioning (Chen et al., 2023)).
• Coordinate systems in biology: Transfer and linearity error metrics (e.g., one-way and two-way transfer errors in biventricular coordinates) directly quantify cross-geometry consistency and spatial linearity (Schuler et al., 2021).

Position-consistent accuracy, therefore, subsumes and generalizes several experimental protocols and metrics, with a defining emphasis on invariance under position, order, or spatial configuration.

2. Position Bias, Permutation and Ordering Effects

Position bias is a fundamental driver of position-consistent accuracy failures across domains:

• In LLM ranking and rubric evaluation, position bias is empirically demonstrated when model choice frequencies depend on presentation order, as evidenced by per-position accuracy spread (e.g., primacy/recency effect: higher accuracy for rubric positions 1 and $K$) (Xu et al., 2 Feb 2026, Vardasbi et al., 23 Jul 2025).
• In sequence modeling (e.g., modular arithmetic by Transformers), catastrophic failure can occur under format shifts and positional displacement if the model has not been exposed to a sufficient range of positional variability during training (Yudin, 7 Jan 2026).
• In sensor localization and computer vision tasks, geometric layout (e.g., anchor placement, input region) can induce spatially varying estimation uncertainty; geometric dilution of precision (GDOP) and coverage holes manifest as spatial dips in position-consistent accuracy (Pucci et al., 2 Jul 2025, Yogesh et al., 29 Sep 2025, Chen et al., 2023).

Mitigating these effects typically requires aggregating over permutations, explicit curriculum on position, or optimizing topology and spatial layout.

3. Methodologies for Ensuring and Measuring Consistency

The realization and measurement of position-consistent accuracy arise from both experimental protocol design and system architecture:

Domain	Protocol/Methodology	Key Metric/Operation
LLM ranking/judging	Swap orders, dynamic early-stopping, permuted rubrics, confidence adaptation	Majority or mean over permutations, PWC scoring, $A_p$, $\bar A$, $\sigma_A$
Modular arithmetic	Position curriculum, template diversity, consistency loss across shifted formats	Eval-B mean over input positions
UWB/mmWave positioning	Swarm least-squares/global optimization, overdetermined geometries, reduced-complexity tracking	RMSE, mean/stddev over all tags, confidence regions
Geospatial data/GTFS-RT	Geodesic distance to route, KD-tree, ellipsoidal calculations	Distance threshold consistency, mean/var statistics
Cardiac coordinate systems	Laplacian PDEs, trajectory normalization, polygonal mappings	Transfer error, spatial linearity error

Key innovations include: dynamic repetition budgets (Vardasbi et al., 23 Jul 2025); confidence-calibrated early stopping; balanced permutation for rubric ordering (Xu et al., 2 Feb 2026); explicit multi-variant and template diversity loss terms (Yudin, 7 Jan 2026); attention-based correction over temporal sequences in mmWave tracking (Chen et al., 2023); and spatially-aware parameterizations for cardiac geometry (Schuler et al., 2021).

4. Empirical Outcomes and Quantitative Performance

Position-consistent accuracy metrics have been shown to distinguish and benchmark robustness in diverse contexts:

• LLM-based Ranking and Judging: Consensus aggregation and early stopping achieve 100% of position-consistent accuracy (vs. 60–75% for swap-once), with 81–87% fewer LLM calls; confidence-adaptive methods attain nearly constant accuracy while sharply reducing repetitions (Vardasbi et al., 23 Jul 2025). Balanced-permutation scoring improves correlation with human judgment by +0.02–0.09 across datasets, while reducing position bias from 8% spread to ≤1% (Xu et al., 2 Feb 2026).
• Input-Displacement Robustness: Transformers trained with curriculum, anchor tokens, and variant-consistency loss achieve position-consistent accuracy (Eval-B) ≈74%, vs. 15% for baseline models, with >95% in-distribution accuracy maintained (Yudin, 7 Jan 2026).
• Sensor and Localization Systems: UWB swarm optimization reduces error mean up to 40% and variance up to 30% compared to trilateration, maintaining 95% confidence bounds under realistic noise (Yogesh et al., 29 Sep 2025). mmWave V-ChATNet achieves sub-20 cm 2D error at the 95th percentile, an order of magnitude gain over geometric or EKF baselines (Chen et al., 2023). Geodesic GTFS-RT post-processed feeds maintain ~80% of reported positions within 35 m of the scheduled route during the day (Wong, 6 Jun 2025).
• Biosciences/Coordinate Consistency: Cobiveco delivers one-way transfer errors 4–7× lower than legacy methods and rotational/apicobasal linearity errors reduced by factors of 4–10, ensuring high-precision spatial mapping between anatomical geometries (Schuler et al., 2021).

5. Best Practices for Achieving Position-Consistent Accuracy

A consistent set of design and training practices emerges across application domains:

• Aggregate across all (or a balanced subset of) possible position/orderings. Balanced permutation or dynamic sampling guarantees all labels or items are equitably tested for positional artifacts (Xu et al., 2 Feb 2026, Vardasbi et al., 23 Jul 2025).
• Incorporate explicit positional variability and diversity during training. Curriculum-driven or template-diverse training improves robustness against shift and OOD format (Yudin, 7 Jan 2026).
• Employ overdetermined, all-pair optimization (in sensor nets) or redundant geometric constraints. Swarm-based optimization and robust path estimation lower both error means and variances (Yogesh et al., 29 Sep 2025, Chen et al., 2023).
• Calibrate using transfer/linearity error metrics in geometric mapping. Evaluate and enforce linearity and bijection to control for cross-sample and cross-space biases (Schuler et al., 2021).
• Use Fisher information analysis and network-level resource allocation to flatten spatial PEB heatmaps. Optimize array geometry, spatial diversity, and sensing resource usage for homogeneous accuracy coverage (Pucci et al., 2 Jul 2025).
• Estimate and monitor per-position or per-region accuracy/variance. Deploy per-position accuracy reporting, error-threshold distributions, and per-metric standard deviations to localize and correct consistency failures (Wong, 6 Jun 2025, Xu et al., 2 Feb 2026).

6. Interpretation, Limitations, and Contextual Considerations

Although position-consistent accuracy is a principled and rigorous standard, its attainment can demand significant resource or architectural overhead (e.g., repeated LLM calls for all rubrics, or comprehensive coverage in sensor topology). There exist trade-offs between computational cost, statistical efficiency, and the depth of positional invariance achieved.

Systematic sources of residual bias include:

• Data/modeling artifacts: Input format not seen during training, model position sensitivity, or lack of explicit invariance can cause failures under position/displacement shift (Yudin, 7 Jan 2026).
• Physical/geometric layout: Anchor geometry or spatial placement in wireless/robotic applications can give rise to blind spots or high-variance “GDOP” regions (Pucci et al., 2 Jul 2025).
• Prompt or rubric design: In LLM evaluation, default orderings or overuse of a single template can mask or exacerbate position effects, necessitating balanced permutation and metric decomposition (Xu et al., 2 Feb 2026).

Position-consistent accuracy thus functions both as an evaluation metric and as an actionable design principle, enabling robust, bias-corrected, and equitable validation across rapidly proliferating domains.

7. Connections and Application Domains

The concept of position-consistent accuracy integrates advances from multiple research communities:

• LLM evaluation and ranking (Vardasbi et al., 23 Jul 2025, Li et al., 28 Mar 2025, Xu et al., 2 Feb 2026)
• Robust and invariant learning (Yudin, 7 Jan 2026)
• Sensor fusion and optimization in localization (Yogesh et al., 29 Sep 2025, Chen et al., 2023, Pucci et al., 2 Jul 2025)
• Precise spatial calibration in biomedical imaging and physics (Schuler et al., 2021, Hong et al., 2016)
• Transport and geospatial informatics (Wong, 6 Jun 2025)
• Hand pose reconstruction from sEMG (Hadidi et al., 9 Mar 2026)

Uniformity and reliability under positional changes are now considered essential for the deployment of both high-stakes machine learning models and mission-critical physical systems. The consistent application of rigorously defined position-consistent accuracy metrics enables meaningful comparison, principled improvement, and scientific reproducibility across these diverse fields.