PaperlayerView exact revision on arXiv

Source-linked AI summary

GATTA: Graph Active Learning with Test-Time Augmentation

Zsombor Bánfi, András Gézsi, András Formanek

arXiv:2608.15084v1cs.LGcs.AI

TL;DR

TTA’s application to graph data remains largely unexplored, while graph node classification often requires costly labeled data. GATTA aggregates uncertainty across augmented graph views; Least Confidence gains +2.87% average and Entropy +3.03%.

  • Problem

    Although TTA improves uncertainty estimation in computer vision, its application to graph data remains largely unexplored, while graph node classification often requires costly labeled data.

    Source: §1 · pp. 1–2

  • Method

    GATTA integrates graph augmentations into transductive node-classification active learning, aggregating acquisition uncertainty across views and filtering potentially misleading perturbations by consistency.

    Source: §1 · p. 2, §3.2 · p. 5

  • Results

    Least Confidence gains +2.87% average and Entropy +3.03%; GATTA matches or exceeds MC Dropout on 4 of 5 datasets.

    Source: §6 · pp. 10–11

  • Takeaways & Limitations

    Simple uncertainty strategies achieve performance competitive with complex baselines at lower implementation effort, and GATTA-S is recommended when computationally feasible.

    Source: §7 · p. 12

    Evidence is limited to transductive node classification on five homophilic citation and co-purchase networks; extension to other graph settings remains untested.

    Source: §8 · p. 12

Abstract

from arXiv · show

Test-time augmentation (TTA) has proven effective for improving model robustness and uncertainty estimation in computer vision, yet its application to graph-structured data remains largely unexplored. We introduce GATTA (Graph Active Learning with Test-Time Augmentation), a framework for enhancing active learning by aggregating predictions across multiple augmented views to produce more reliable uncertainty estimates. To address the challenge of label-preserving graph augmentations, GATTA incorporates a consistency-based filtering mechanism that discards augmented views yielding unreliable predictions. We systematically evaluate GATTA across multiple graph datasets, GNN architectures, and acquisition strategies. Our results show that simple uncertainty-based methods, such as Entropy and Least Confidence, benefit most from TTA, achieving performance competitive with more sophisticated and computationally expensive approaches. GATTA generalizes across architectures, outperforms model-side ensemble methods such as MC Dropout. We further show that GATTA scales efficiently with both ensemble size and graph size. Extensive analysis of augmentation types, strengths, and filtering strategies provides practical guidelines for effective deployment. Our findings demonstrate that augmenting simple methods with TTA offers a more efficient path to strong active learning performance than engineering complex acquisition functions, enabling practitioners to achieve competitive results with lower computational overhead and reduced implementation complexity.

1 Introduction

GATTA applies test-time augmentation and consistency-based filtering to transductive node-classification active learning, where graph structure complicates uncertainty estimation. Experiments support augmenting simple uncertainty strategies for competitive performance at lower computational cost, with practical guidance on augmentation and ensemble design.

1 Introduction

  • GNN-based node classification often needs substantial labeled data, creating costly annotation bottlenecks in scientific and industrial domains.The introduction identifies expert annotation and experimental validation as particular sources of cost.Source: §1 · p. 1
  • Graph structure makes active-learning uncertainty difficult to estimate because multi-hop dependencies complicate uncertainty distinctions and can bias node selection (Fuchsgruber et al., 2024; Wang et al., 2024).The passage contrasts graph data's non-i.i.d. structure with traditional active-learning settings (Hu et al., 2020).Source: §1 · p. 2
  • Test-time augmentation improves uncertainty estimation in computer vision, but its application to graph data remains largely unexplored (Gaillochet et al., 2022; Wang et al., 2018; Conde et al., 2023; Ju et al., 2023; Bo et al., 2021).The introduction proposes graph-specific augmentations as a way to exploit relational structure for active-learning uncertainty estimates.Source: §1 · p. 2
  • GATTA combines graph-specific augmentations, score or prediction aggregation, and consistency-based filtering for transductive node classification, excluding graph-level classification.Filtering discards potentially misleading perturbations while preserving label-relevant properties; labels are iteratively acquired within a single graph per dataset.Source: §1 · p. 2

Contributions

  • GATTA pairs simple uncertainty methods with consistency-based filtering, matching complex acquisition functions at lower computational cost and providing deployment guidance on augmentation type, strength, and ensemble size.The contributions emphasize practical guidelines for choosing augmentation and ensemble configurations.Source: Contributions · p. 2, §1 · p. 2

2 Related Work

Prior work highlights the challenges of uncertainty estimation on graphs and the risk that graph augmentations change label-relevant information. GATTA positions consistency-filtered test-time augmentation as a framework for graph active learning, with systematic deployment analysis.

2.1 Active Learning on Graphs

  • Graph dependencies complicate uncertainty estimation because predictions draw on multi-hop neighborhoods, while sophisticated graph active-learning methods can be computationally demanding.Prior approaches include AGE’s combination of embeddings, uncertainty, and centrality; ANRMAB’s bandit formulation; GEEM’s expected-error reduction; JuryGCN’s jackknife estimates; and MP/ESP approximations for epistemic uncertainty.Source: §2.1 · pp. 2–3

2.2 Uncertainty Quantification in Graph Neural Networks

2.3 Test-Time Augmentation

  • Test-time augmentation aggregates predictions across transformed views to capture prediction variance, and its use for active-learning image annotation leaves transfer to graph neural networks open.Prior applications span medical image segmentation (Shanmugam et al., 2021), calibration theory (Conde et al., 2023), text classification (Lu et al., 2022), and medical image annotation (Gaillochet et al., 2022); graph perturbations affect both node attributes and relational structure.Source: §2.3 · p. 3

2.4 Graph Augmentation Techniques

  • Graph augmentations such as edge dropout, subsampling, masking, noise, and shuffling can alter receptive fields and semantic evidence, motivating label-consistency safeguards.Prior approaches constrain perturbations in representation space (Yue et al., 2022) or search for label-invariant policies; GATTA’s filtering aims to distinguish semantic ambiguity from structural instability.Source: §2.4 · p. 4

2.5 Novelty and Positioning

  • GATTA is presented as the first framework to systematically integrate test-time augmentation into graph active learning, addressing non-label-invariant perturbations with a graph-specific consistency mechanism.The work studies acquisition strategies, aggregation, augmentations, and computational trade-offs, and reports that simple uncertainty-based strategies can match or outperform more sophisticated methods with improved uncertainty estimates; earlier graph TTA addressed other objectives, including social influence prediction and degree bias.Source: §2.5 · p. 4

3 Method

GATTA is a plug-and-play method that estimates acquisition uncertainty across augmented graph views. It combines score or prediction aggregation with consistency filtering to reduce the influence of views that may not preserve node semantics.

3.1 Acquisition Strategies

  • Acquisition strategies score unlabeled nodes from predicted class probabilities; the paper groups them into simple uncertainty-based, complex uncertainty-based, and other methods.The categories include Least Confidence and Entropy; MP and ESP; and AGE, ANRMAB, and GEEM, respectively.Source: §3.1 · p. 4

3.2 Graph Active Learning with Test-Time Augmentation

  • GATTA is a plug-and-play module for transductive node classification that estimates acquisition uncertainty over stochastic graph augmentations using Monte Carlo predictions from augmented views and the original graph.The framework generates perturbed views and aggregates predictions to estimate uncertainty-based acquisition scores.Source: §3.2 · p. 5

3.3 Aggregation Methods: GATTA-S and GATTA-P

  • GATTA-S averages acquisition scores computed per view, whereas GATTA-P applies the acquisition function after averaging predictions; P is cheaper but more sensitive to disagreement from non-label-preserving views.GATTA-P requires only one acquisition-function evaluation, and its sensitivity motivates consistency filtering.Source: §3.3 · p. 5

3.4 Graph Augmentations

  • The method uses feature masking, Gaussian feature noising, and edge dropout to perturb feature reliance or graph structure while aiming to preserve label-relevant information.Masking zeros node features with probability pmask, while edge dropout removes edges with probability pdrop.Source: §3.4 · p. 6

3.5 Consistency Filtering

  • Because augmentations can induce semantic drift, consistency filtering restricts uncertainty estimates to model-consistent perturbations and averages scores over consistent views as a conditional-expectation estimate.Prediction consistency with the original graph serves as a practical proxy for local semantic consistency.Source: §3.5 · p. 6

4 Active Learning Protocol

The protocol retrains the GNN from scratch after each acquired label, then evaluates the final model on a held-out test set. Augmentation-strength effects vary by GATTA variant and filtering, while comparisons keep training hyperparameters matched.

Active learning protocol

  • Each round scores unlabeled nodes with the selected strategy, annotates the highest-scoring node, and retrains the GNN from scratch before the next acquisition.The process begins with one randomly selected node per class and continues until the acquisition budget is exhausted, followed by held-out test evaluation.Source: §4 · p. 6
  • Unless otherwise stated, the protocol acquires one node per iteration for a budget of 4C, matching prior graph active-learning work and leaving a small labeled set.C denotes the number of classes.Source: §4 · p. 6

Evaluation protocol

  • Results average 25 trials, and GATTA changes only inference-time acquisition scores while keeping training hyperparameters identical to each baseline.The trials combine five random initial labeled pools with five random model initializations.Source: §4 · p. 7

Augmentation strength

  • For FN+ED, gains concentrate at augmentation strengths around 0.3–0.5; filtering broadens effective strengths for GATTA-S, while GATTA-P improves only with filtering.The heatmaps compare noise variance and dropout probability, with and without filtering.Source: Figure 1 · p. 7

5 Sensitivity and Configuration Analysis

Across CoraML and PubMed, GCN and SGC, and Entropy and Least Confidence, this section examines augmentation choices, filtering, strength, ensemble size, and runtime to inform GATTA configuration.

5.1 Augmentation Type Selection

  • FN+ED delivered the largest average improvement (+1.12 ± 1.05%), while FN exceeded FM (0.69 ± 0.67% vs. 0.03 ± 0.83%).FM+ED showed +0.53±0.96%, comparable to single augmentations; the study adopted FN+ED for subsequent experiments. The paper attributes FN's advantage to preserving feature scale, while masking creates information voids that propagate through message passing.Source: Table 1 · p. 7, §5.1 · p. 8

5.2 Augmentation Strength and Filtering

  • GATTA-S peaked without filtering (2.49 ± 1.42%) and stayed robust, whereas GATTA-P failed unfiltered (0.12 ± 1.04%) but improved with filtering (1.96 ± 1.27%).With filtering, stronger augmentation strengths—noise σ ∈ [0.3, 0.5] and edge-dropout probability pdrop ∈ [0.3, 0.5]—achieved improvements up to 3.0 over baseline. Filtering is essential for GATTA-P because label-inconsistent views corrupt prediction-level averages; GATTA-S averages acquisition scores across views and can balance such signals without explicit filtering.Source: §5.2 · p. 8

5.3 Ensemble Size, Runtime, and Scalability Analysis

  • Accuracy gains followed Performance ≈ baseline + 0.37×log(N+1), reaching +2.65±0.93% at N ≈200 before diminishing.Beyond N = 500, each additional 100 views yielded less than 0.1% improvement, suggesting moderate ensembles capture most uncertainty-signal diversity.Source: Figure 2 · p. 8, §5.3 · p. 9
  • At N = 500, both variants took ∼2× baseline time for Entropy; GATTA-S scaled poorly for AGE while GATTA-P remained efficient.Runtime scaling was consistent across CoraML (2,810 nodes, 15,962 edges) and PubMed (19,717 nodes, 88,648 edges), indicating overhead depends on ensemble size and acquisition complexity rather than graph scale.Source: Figure 2 · p. 8, §5.3 · p. 9

6 Results

Across five graph datasets, GATTA improves simple uncertainty-based active learning most, while gains vary by dataset and extend across architectures. It also matches or exceeds MC Dropout on most datasets.

R1. GATTA selectively benefits uncertainty-based methods

  • Least Confidence gains +2.87% on average and Entropy +3.03%, with GATTA-S generally outperforming GATTA-P on these simple uncertainty methods.GATTA-S peaks above +5% on multiple datasets and can exceed baseline GEEM; MP gains +2.11% on average but falls by 1.88% on Citeseer, while ESP, ANRMAB, AGE, and GEEM show inconsistent or minimal benefit.Source: §6 · p. 10, Figure 3 · p. 10

R2. GATTA’s benefits emerge early and persist throughout learning

  • GATTA improves sample efficiency from early acquisition rounds, and its performance gap over baseline methods persists throughout learning.The early benefit matters when labeling budgets end before model saturation.Source: §6 · p. 10, Figure 3 · p. 10

R3. Performance depends on unique dataset characteristics, not broader graph category

  • GATTA’s gains vary substantially within citation and co-purchase networks, with simple-method improvements exceeding +5% on AmazonComputers using SGC.Citation-network results range from strong gains on PubMed to modest ones on Citeseer; the paper recommends pilot testing on a data subset because dataset characteristics may matter more than broad domain category.Source: §6 · pp. 10–11

R5. GATTA outperforms MC-Dropout

  • GATTA matches or exceeds MC Dropout on 4 of 5 datasets, with gains up to +3.73% on PubMed.Combining the methods gives no consistent improvement; GATTA-P + MCD falls 8.39% below baseline on AmazonComputers, suggesting overlapping or conflicting uncertainty signals.Source: Table 3 · p. 11, §6 · p. 11

R4. GATTA generalizes across architectures

  • GATTA generalizes across GCN, SGC, GAT, and GraphSAGE, with improvements on most datasets for GAT and GraphSAGE.GATTA-S is stronger for GAT, while GraphSAGE shows more dataset-dependent results between variants; GATTA operates at the input level without model modifications.Source: §6 · p. 11, Table 3 · p. 11

7 Discussion

GATTA improves the predictive distribution rather than the acquisition objective, so methods that rely most on predictive uncertainty benefit most. The discussion explains how consistency filtering affects augmentation strategies and offers practical deployment guidelines.

7 Discussion

  • GATTA’s impact is greatest for Entropy and Least Confidence, which rely almost entirely on predictive uncertainty, and smaller or potentially negative for methods combining uncertainty with structural information or model-improvement approximations.This may explain smaller gains for ESP and GEEM and MC Dropout’s degradation on Amazon Computers.Source: §7 · p. 11
  • GATTA-P is sensitive to semantic drift because inconsistent predictions affect its aggregated distribution, while GATTA-S averages acquisition scores across views and is more robust.Consistency filtering mitigates the effect for GATTA-P; empirically, GATTA-S performs well without filtering, while GATTA-P benefits substantially from it.Source: §7 · pp. 11–12
  • Practitioners should prioritize Entropy and Least Confidence for lower-effort performance competitive with complex baselines, and use GATTA-S when feasible or filtered GATTA-P when acquisition functions are expensive.GATTA-S consistently outperforms GATTA-P and requires no filtering.Source: §7 · p. 12
  • Moderate-to-high augmentation strengths (noise ∈[0.4, 0.5], pdrop ∈[0.3, 0.5]) and ensemble size N = 500 offer a favorable performance-runtime trade-off.GATTA is robust across a broad range of hyperparameter settings, as noted in section 5.Source: §7 · p. 12
  • Because augmentation settings interact with acquisition strategies differently across datasets, practitioners should validate them on a representative subset of the target graph before large-scale deployment.The recommendation follows the reported dataset-dependent interaction between test-time augmentation and acquisition strategies.Source: §7 · p. 12

8 Conclusions and Limitations … A.3 Model Details

GATTA integrates test-time augmentation into graph active learning, improving uncertainty estimation especially for simple acquisition strategies. The paper reports the setup and limits its evaluation to transductive node classification on five homophilic networks.

8 Conclusions and Limitations

  • GATTA improves graph active learning by enhancing predictive uncertainty, with simple acquisition strategies often matching or outperforming more sophisticated methods without retraining or architectural changes.The study also derives deployment guidelines by examining aggregation, augmentations, and computational trade-offs.Source: §8 · p. 12
  • Evaluation is limited to transductive node classification on five homophilic citation and co-purchase networks; broader graph domains and tasks remain untested.The filtering mechanism assumes reliable predictions on the original graph, and formal analysis of convergence and sample complexity remains future work.Source: §8 · p. 12
  • Future work includes confidence- or structure-adaptive augmentation, theory linking graph properties to uncertainty, and acquisition strategies designed to exploit TTA.Interactions with ESP, GEEM, and MC Dropout also warrant further study.Source: §8 · p. 12

A Experimental Setup

A.1 Active Learning Protocol

  • The protocol starts with one randomly selected labeled node per class, trains to convergence, acquires an unlabeled node, adds its label, and repeats until the budget is exhausted.The acquisition loop retrains from scratch after each iteration.Source: §A.1 · p. 17
  • After the final acquisition, the model is retrained on all labeled nodes and evaluated by held-out test accuracy; the default budget is 4C labels, one per iteration.Final training pools are smaller than standard semi-supervised benchmark splits.Source: §A.1 · p. 17

A.2 Datasets

  • The benchmark comprises CoraML, Citeseer, PubMed, AmazonComputers, and AmazonPhotos: three citation networks and two co-purchase networks.Nodes represent papers linked by citations in the citation datasets, and products linked by frequent co-purchases in the co-purchase datasets.Source: §8 · p. 12, §A.2 · p. 17

A.3 Model Details

  • Without validation-based hyperparameter optimization, the study applies literature-reported hyperparameters uniformly across datasets for SGC and GCN.The model configurations are listed in Table 5.Source: §A.3 · p. 17, Table 5 · p. 18

A.4 Training and Evaluation Details · B Information-Theoretic Interpretation of GATTA Aggregation · C Filtering variants

The sections specify the training and evaluation protocol, explain why prediction aggregation can be more sensitive to graph augmentations than score aggregation, and report that confidence-weighted Firm filtering did not outperform Hard filtering.

A.4 Training and Evaluation Details

  • Models use binary cross-entropy with Adam, a 10^-3 learning rate and weight decay, and early stopping on validation loss after 100 iterations without improvement.The optimizer and loss settings apply to all models.Source: §A.4 · p. 18
  • Evaluation averages 25 runs per dataset, acquisition function, and model, reports test accuracy, and computes gains against the matching non-TTA baseline.The protocol uses five dataset splits and five independent initializations per split; a shared 20% test set is excluded from acquisition, and each split has a reused 20% validation set.Source: §A.4 · p. 18

B Information-Theoretic Interpretation of GATTA Aggregation

  • For entropy acquisition, GATTA-S estimates expected view entropy, whereas GATTA-P estimates averaged-prediction entropy, which adds mutual information between predictions and augmentations.The mutual-information term captures disagreement across augmented graph views.Source: B Information-Theoretic Interpretation… · p. 18
  • Augmentation disagreement can reflect semantic drift as well as local instability, so GATTA-P may overestimate uncertainty while GATTA-S is less sensitive to cross-view disagreement.Consistency filtering mitigates this risk by approximately restricting aggregation to model-consistent perturbations.Source: B Information-Theoretic Interpretation… · pp. 18–19

C Filtering variants

  • Hard filtering discards inconsistent views, while Firm filtering also weights consistent views by confidence; soft weighting alone proved ineffective.Firm filtering combines hard exclusion with confidence-based weighting because inconsistent views can be spuriously confident for the wrong class.Source: C Filtering variants · p. 19
  • Firm filtering did not outperform Hard filtering in the reported experiments.Table 6 compares filtering strategies using Entropy and Least Confidence across CoraML and PubMed with GCN and SGC models.Source: Table 6 · p. 19, C Filtering variants · p. 19

D Impact of Data Augmentation

Stronger augmentations generally improve performance, but the value of consistency-based filtering depends on the aggregation strategy. These trends hold across SGC and GCN, with no substantial difference between Firm and Hard filtering.

D Impact of Data Augmentation

  • The strength analysis covers Feature Noising and Edge Drop on CoraML and PubMed with Entropy and LC acquisition strategies.Figures 11–14 report accuracy gains relative to the corresponding configuration without test-time augmentation.Source: D Impact of Data Augmentation · p. 19
  • Stronger augmentations generally improve performance, but filtering is essential for GATTA-P at higher strengths and unnecessary for optimal GATTA-S results.These strategy-specific patterns are consistent across both SGC and GCN architectures.Source: D Impact of Data Augmentation · p. 19
  • The analysis finds no substantial performance difference between Firm and Hard filtering variants.The reported trends are consistent across SGC and GCN architectures.Source: D Impact of Data Augmentation · p. 19

E Learning Dynamics and Performance Comparison · F Batch Active Learning and Extended Labeling Budgets · G Computational complexity

GATTA improves simple acquisition strategies most, with effectiveness varying by dataset and acquisition batch size; GATTA-P is more computationally efficient than GATTA-S for complex acquisition functions. Larger batches and larger labeling budgets reduce GATTA’s relative benefit.

E Learning Dynamics and Performance Comparison

  • GATTA’s effectiveness varies by dataset: Amazon Computers gains at least 5% with GATTA-S, while Amazon Photos improves more moderately and Citeseer shows modest gains.Within citation networks, PubMed shows strong gains; graph properties such as feature informativeness, homophily, or class balance may matter more than broad dataset categories.Source: E Learning Dynamics and Performance… · p. 20
  • For AGE and ANRMAB, GATTA’s gains are modest or may reverse, while GEEM shows slight degradation with GATTA.The text suggests that uncertainty-aware mechanisms in complex strategies and interference with structure-aware acquisition may limit augmentation’s benefit.Source: E Learning Dynamics and Performance… · p. 20
  • GATTA-S reduces standard deviation in 18 of 20 simple-method dataset–architecture combinations, with exceptions on PubMed using SGC.The passage attributes the more consistent outcomes to prediction aggregation across augmented views stabilizing node selection.Source: E Learning Dynamics and Performance… · p. 21

F Batch Active Learning and Extended Labeling Budgets

  • Larger acquisition batches degrade both baseline and GATTA performance and reduce GATTA’s relative advantage as decisions rely on more of the uncertainty ranking.Figure 4 compares batch sizes of 1, 5, 10, 50, and 100 queried nodes per iteration on CoraML and PubMed with GCN and SGC; larger batches may yield redundant acquisitions.Source: Figure 4 · p. 20, F Batch Active Learning and Extended… · p. 21
  • GATTA’s advantage over the baseline is largest early in active learning and diminishes as the labeling budget grows.Figure 4 aligns learning curves by total acquired labels; the text attributes the pattern to uncertainty being most valuable when labeled data are scarce.Source: Figure 4 · p. 20, F Batch Active Learning and Extended… · p. 21

E Learning Dynamics and Performance Comparison

  • GATTA variants improve simple strategies, closing the gap with GEEM and sometimes surpassing it on Citeseer, CoraML, and PubMed; gains are especially pronounced for simple uncertainty methods and MP.Table 7 reports final test accuracy across datasets, strategies, and architectures; the text identifies the comparative outcomes.Source: E Learning Dynamics and Performance… · p. 20, Table 7 · p. 24

G Computational complexity

  • GATTA-P has complexity O((N × I) + Q), whereas GATTA-S has O(N × (I + Q)); with 10,000 augmentations and AGE on PubMed, their runtimes exceed 5,000 and approximate 50 seconds per iteration, respectively.GATTA-P evaluates the acquisition function once on aggregated predictions; GATTA-S requires N + 1 evaluations.Source: G Computational complexity · p. 25
  • When acquisition evaluation is negligible, as with Entropy or LC, GATTA-P and GATTA-S have similar runtimes; when it dominates, GATTA-S incurs multiplicative overhead.GATTA-P is preferable for complex acquisition functions, while GATTA-S may suit simple strategies and offer marginal performance benefits.Source: G Computational complexity · p. 25

H Confidence Analysis

For Entropy on Cora-ML, GATTA concentrates prediction confidences toward the mid-range, while consistency filtering shifts them higher by removing inconsistent augmented views. These changes suggest more reliable uncertainty estimates for active learning.

H Confidence Analysis

  • Across 500 augmented views, GATTA concentrates confidence in the 0.4–0.6 range, moderating overconfident and underconfident predictions relative to the baseline.The distributions are averaged over 25 runs across 25 active-learning iterations.Source: H Confidence Analysis · p. 25, Figure 10 · p. 26
  • In both baseline and GATTA distributions, mean confidence stays above the densest region, reflecting a high-confidence-skewed unlabeled pool as uncertain nodes are queried.This asymmetry is characteristic of uncertainty-based active learning.Source: H Confidence Analysis · p. 25, Figure 10 · p. 26
  • Consistency filtering reduces predictions below 0.4 and shifts the GATTA confidence distribution upward by discarding augmented views with inconsistent predictions.The filtered distribution is both more concentrated and shifted toward higher confidence values.Source: H Confidence Analysis · p. 25, Figure 10 · p. 26
  • The concentration and elevation effects suggest GATTA can provide more reliable uncertainty estimates, potentially improving node selection and label efficiency.Source: H Confidence Analysis · p. 26

Related papers

Loading 2608.15084v1…