Source-linked AI summary
Residual Calibration via Local Feature-Space Refinement
Xihao Wang, Chengyao Yu, Ruixing Ming
TL;DR
Global score-based calibration can miss spatially varying reliability across feature neighborhoods. LAPE refines a global calibrated anchor using nearby calibration samples, with cross-fitting for shared calibration data and CFAF for multi-view inputs. The paper reports improved calibration across diverse baselines and asymptotic consistency under regular assumptions.
Problem
Global score-based calibrators can give similar corrections to similar scores despite differences in empirical reliability across feature neighborhoods.
Source: §1 · p. 2Method
Results
LAPE improves calibration performance across diverse post-hoc baselines and is asymptotically consistent under regular assumptions.
Source: §1 · p. 2, §5 · p. 7, §7 · p. 10Takeaways & Limitations
LAPE can reuse calibration data for global calibration and residual refinement while serving as a plug-in refinement for existing post-hoc methods.
Source: §1 · p. 2, §7 · p. 10LAPE's effectiveness depends on learned feature geometry: when nearby samples have dissimilar conditional label distributions, local frequency estimates can be biased and the method should rely more on the global anchor.
Source: §5 · p. 7, §7 · p. 10
Abstract
from arXiv · showhide
Deep neural networks often struggle to provide reliable confidence estimates, making post-hoc calibration essential for reliable decision-making. Existing post-hoc methods largely rely on score-based transformations, which improve marginal calibration but often fail to capture spatially varying miscalibration in local regions of the learned feature space. To address this gap, we propose Local Adaptive Probability Estimation (LAPE), a post-hoc residual calibration framework. Specifically, LAPE treats the global calibrator output as a stable anchor and refines the individual prediction by leveraging non-parametric local evidence from neighboring calibration samples. As a plug-and-play refinement module, LAPE effectively utilizes the information about feature-space geometry and can reuse the calibration data that are used to construct the global calibrator. Furthermore, we extend LAPE to the setting with multi-view inputs through Cascade Feature Augmentation Fusion (CFAF). Extensive experiments demonstrate that the proposed methods can improve the calibration quality of various existing baselines and remain efficient and robust across different settings.
1 Introduction
Reliable confidence estimates matter for decisions that depend on prediction uncertainty, but global score-based calibration can miss differing reliability across feature neighborhoods. The paper proposes LAPE to refine a global calibrator with local evidence, and extends it to multi-view inputs with CFAF.
1 Introduction
- Reliable confidence scores can support deferral to experts in clinical diagnostics and improve selection power in robust conformalized selection.The introduction identifies confidence misalignment as overconfidence or underconfidence and cites its importance for downstream tasks relying on prediction uncertainty.Source: §1 · p. 1
- Existing global score-based maps give similar calibrations to similar scores despite differing neighborhood reliability, while local parametric methods risk overfitting and non-parametric methods can discard global logits.These limitations motivate the challenge of correcting residual errors with local feature-space evidence while preserving global predictive stability.Source: §1 · pp. 1–2
- LAPE uses local feature-level evidence to refine existing post-hoc calibrators, with the global calibrator output as an anchor and corrections applied when nearby calibration samples support them.V-fold cross-fitting constructs out-of-sample anchor scores while reusing the calibration set; CFAF extends the approach to multi-view settings.Source: §1 · p. 2
- The paper establishes asymptotic consistency for LAPE and reports improved calibration performance across diverse datasets and architectures for existing post-hoc methods.Reported baselines include TS, Feature Clipping, and Semantic-Aware Grouping; further experiments are described as demonstrating efficiency, robustness, and effectiveness.Source: §1 · p. 2
2 Preliminary
The setup considers estimating class probabilities from a soft classifier, calibration data, and feature embeddings, with LAPE intended to improve existing post-hoc calibrators. The motivation is that learned representations may separate training samples without providing uniform local calibration support for held-out samples.
2.1 Problem Setup
- LAPE aims to improve an existing post-hoc calibrator by using its calibrated output, the calibration dataset, and feature embeddings to estimate class probabilities.The setup defines a K-class label variable and permits calibrators such as VS and TS.Source: §2.1 · p. 2
2.2 Motivation
- After representation learning, training samples are more clearly separated, while held-out calibration samples can remain mixed near induced boundary regions.The figure contrasts raw and learned features; different colors denote classes, and the highlighted neighborhood illustrates local evidence used by LAPE.Source: §2.2 · p. 3, Figure 1 · p. 3
- For class-1 scores in [0.35, 0.65], 32-neighbor local frequencies better approximate oracle probabilities than softmax predictions (MAE: 0.126 vs. 0.156).The comparison uses the full learned feature space.Source: §2.2 · p. 3
- In a 100-dimensional simulation, learned-feature class separation was larger for training than held-out calibration samples (3.663 ± 0.343 versus 2.981 ± 0.234).Raw-space separation was nearly identical between the training and calibration sets (2.510 ± 0.176 versus 2.519 ± 0.162); results average ten random seeds.Source: §2.2 · p. 3
- In the predefined 100-dimensional boundary region, local-frequency estimation had lower oracle-probability error than softmax (0.0969 ± 0.0055 versus 0.1067 ± 0.0098).The local estimate used Laplace-smoothed frequencies from 50 nearest calibration samples in the learned representation.Source: §2.2 · p. 3
3 Local Adaptive Probability Estimation
LAPE retrieves nearby calibration samples in feature space and blends their local empirical probabilities with a global anchor. It adaptively selects the neighborhood size and correction weight, and uses cross-fitting when the calibration data also fit the global calibrator.
3.1 General Methodology
- LAPE retrieves calibration samples nearest to a test point in feature space and estimates local class probabilities from their labels.The neighborhood is defined by Euclidean distances between feature embeddings, with smoothing used to avoid zero probabilities.Source: §3.1 · p. 4
- The final estimate blends the global anchor with the local probability estimate, using a weight that controls the strength of the local correction.The local evidence is treated as a residual signal rather than a replacement for the anchor, since it can be weakly supported in sparse regions.Source: §3.1 · pp. 4–5
- LAPE selects neighborhood size and correction weight by minimizing density-weighted negative log-likelihood using weighted leave-one-out analysis.The search considers neighborhood sizes of order O(√n), with practical bounds ⌊√n/10⌋ and ⌊√n/2⌋.Source: §3.1 · pp. 4–5
3.2 Applying LAPE to Calibrators with Shared Calibration Dataset
- When a calibrator and LAPE share the calibration dataset, V-fold cross-fitting produces out-of-sample anchor scores for calibration samples.The approach addresses calibration-set overfitting while avoiding the loss of calibration-data utilization associated with a simple disjoint split.Source: §3.2 · p. 5
4 Extension to Multi-View Scenarios
CFAF extends calibration to multi-view inputs, where separate view-specific predictors may miss cross-view interactions. It combines calibrated view predictions and applies LAPE using a rectified joint feature space.
4 Extension to Multi-View Scenarios
- CFAF targets multi-view settings where late fusion of separately calibrated view predictors can miss cross-view interactions.The setting includes objects described by several feature groups or modalities, and retraining an end-to-end interaction model may be costly or infeasible in post-hoc deployment.Source: §4 · p. 6
- CFAF forms an anchor from calibrated view-level predictions, with fusion weights computed using out-of-fold predictions.The method fits view-level calibrators on the full calibration set for test-time predictions.Source: §4 · p. 6
- CFAF concatenates view features, rectifies the joint representation by standardization optionally followed by whitening, then applies LAPE's local estimation and refinement.Neighbor retrieval uses the rectified joint features rather than any single view.Source: §4 · p. 6
5 Theoretical Results
The section establishes LAPE’s asymptotic consistency under regular assumptions and interprets when local refinement is reliable through feature-space geometry.
Theoretical Results
- Under Assumptions B.2–B.7, the LAPE estimator converges in probability to the true conditional probability vector as calibration-set size tends to infinity.The theorem assumes an i.i.d. calibration set and test sample.Source: §5 · p. 7
Geometric Interpretation
- The local estimate acts as a smoothed frequency of labels around a test point, with bias governed by variation in conditional probabilities within its retrieved neighborhood.The geometric interpretation ignores the O(1/m) contribution from Laplace smoothing.Source: §5 · p. 7
- When neighborhood probabilities vary substantially, local estimates can be biased; LAPE can shrink the neighborhood or lower interpolation weight to rely more on the global anchor.The cited appendix table supports the interpretation that LAPE is closer to the true probability in locally smooth neighborhoods and has a larger oracle gap when neighborhood variation is large.Source: §5 · p. 7
6 Experimental Results
The experiments compare existing calibrators with LAPE refinements across tabular and image benchmarks, and examine multi-view calibration and sensitivity to calibration-set size. Results show frequent calibration improvements, with gains depending on the anchor and feature geometry.
6.1 Experimental Setup
- The main Figure 2 comparison evaluates four datasets spanning tabular and image data, using MLP, ResNet18, DenseNet121, and ResNet50 backbones.The methods pair TS, VS, FC, and SAG with their corresponding LAPE-adjusted versions; evaluation includes NLL and ECE.Source: §6.1 · pp. 7–8
6.2 Results
- Across tabular MLPs, medical-image transfer models, and natural-image backbones, adding LAPE often reduces NLL and ECE while largely preserving accuracy.The gains vary by anchor and metric, depending on anchor quality and the informativeness of learned feature geometry.Source: §6.2 · p. 8, Figure 2 · p. 8
- CFAF-LAPE consistently improves NLL, ECE, and accuracy across multi-view tasks and is reported as much better than late fusion.CFAF forms a calibrated fusion anchor and applies local refinement in a rectified joint representation without retraining an end-to-end fusion model.Source: §6.2 · p. 9, Table 1 · p. 9
- VS+LAPE lowers NLL at every calibration ratio, while ECE and accuracy show favorable overall trends despite small fluctuations.Its NLL decreases further as more calibration data become available, but improvement in the primary NLL metric does not require a large calibration set.Source: §6.2 · p. 9, Figure 3 · p. 9
- Appendix experiments report practical inference-time runtime, robustness diagnostics, and additional evidence across datasets, backbones, calibration objectives, and multi-view fusion methods.The diagnostics also report reduced gains with less reliable retrieval geometry and greater effectiveness of cross-fitting than direct reuse or simple splitting for flexible anchors with limited calibration data.Source: §6.2 · p. 9
7 Conclusion
The conclusion presents LAPE as a post-hoc residual refinement of global calibrated predictions using local feature-space evidence, with cross-fitting and CFAF extending the framework. It reports theoretical consistency and experimental improvements across diverse post-hoc baselines, while noting dependence on learned feature geometry.
7 Conclusion
- LAPE refines global calibrated predictions with local feature-space evidence, and a V-fold cross-fitting strategy constructs out-of-sample anchor scores from calibration data.The framework also extends to multi-view scenarios through CFAF.Source: §7 · p. 10
- The paper establishes asymptotic consistency under regular assumptions and reports improved calibration quality across diverse post-hoc baselines.It also reports practicality under varying calibration-set sizes and distribution shifts.Source: §7 · p. 10
- Effectiveness depends on learned feature geometry: when nearby samples have dissimilar conditional label distributions, local frequency estimates may be biased and LAPE should rely more on the global anchor.The conclusion identifies more robust retrieval metrics and extension to verbalized confidence in large reasoning models as future directions.Source: §7 · p. 10
A Related Work
This section describes LAPE’s combination of a stable global calibration anchor and feature-space neighbor evidence, alongside preliminary results on the stability and risk behavior of its local refinement.
A Related Work
- Global score-based calibrators can improve average calibration yet leave spatially structured residual errors, while score-level non-parametric methods do not exploit feature-space geometry.Adaptive parametric and group-wise alternatives add fitted components, whereas stand-alone score-level estimators may be data-inefficient and discard information encoded in logits.Source: A Related Work · p. 14
- LAPE uses neighboring reference labels in feature space to refine a fixed global anchor, distinguishing similar-logit examples without adding a learned temperature map or semantic partition.The local empirical distribution acts as a residual correction rather than replacing the classifier output; under regularity conditions, the anchor-preserving refinement consistently estimates conditional class probabilities.Source: A Related Work · pp. 14–15
B.3 Preliminary Lemmas
- Adaptive retrieval-space slices concentrate calibration embeddings around expected counts and control within-slice variation in conditional label frequencies.The number of slices is O(n/an), and conditional-frequency variation within slices vanishes as n grows.Source: §B.3 · pp. 17–18
- Empirical label frequencies aggregated at admissible cut points converge uniformly to their theoretical local expectations, while the local-loss bandwidth shrinks and its neighborhood sample count diverges.The bandwidth result concerns the radius used in the local empirical loss.Source: §B.3 · pp. 19–20
- LAPE’s local weighted leave-one-out objective concentrates uniformly around its conditional mean over the candidate grids.The lemma applies when candidate neighborhoods contain at least mmin,n points, their Gaussian weights are uniformly of order 1/m, and the grid size grows subexponentially in mmin,n.Source: §B.3 · pp. 21–22
- If an estimator’s conditional mean stays a fixed distance from the true conditional probability, its conditional expected negative log-likelihood exceeds the entropy floor.The result assumes positive estimator coordinates, integrable negative log-loss, and conditional independence of the estimator and label given the conditioning sigma-algebra.Source: §B.3 · p. 23
- For candidate neighborhood sizes no larger than mσ,n, the total Gaussian weight outside any fixed feature-space radius vanishes asymptotically.The largest candidate neighborhood radius is bounded by σn, which tends to zero in probability.Source: §B.3 · pp. 23–24
B.4 Proof of Theorem 5.1
For a fixed test point, the proof establishes consistency by controlling local-frequency variance and showing that risk minimization rejects candidates with persistent bias. These controls make the selected estimator converge to the true conditional probability.
B.4 Proof of Theorem 5.1
- The proof combines uniform stability of local empirical frequencies with bias convergence of the risk-minimizing candidate to establish consistency at each fixed test point.The two controls are followed by a decomposition of estimator error into variance and bias terms, both of which vanish in probability.Source: §B.4 · pp. 24, 25, 27
- Uniform risk concentration ensures that a candidate with non-vanishing conditional-mean bias has excess risk, so empirical risk minimization cannot select it asymptotically.The contradiction argument compares such candidates with the consistent pure-local competitor.Source: §B.4 · pp. 25–26
- The pure-local competitor is consistent because its neighborhood radius shrinks while its sample size diverges, making its conditional mean and empirical frequency converge to p(x).The proof uses the fixed-point condition p(x) = pϕ(z) and the stated assumptions.Source: §B.4 · pp. 25–26
C.1 Datasets
The supplementary experiments span diverse datasets, baselines, metrics, and model frameworks, then examine computation and diagnostics of local refinement. Across oracle-variation, local-heterogeneity, and sample-ratio tests, results support the value of local evidence, especially when residual errors vary across neighborhoods.
C.1 Datasets
- The evaluated suites include tabular, medical-image, natural-image, multi-view, and multimodal datasets, alongside synthetic diagnostics designed to isolate specific factors.Examples include Obesity Risk, CIFAR-10/100, MedMNIST, UCI Multiple Features, and HAM10000 image–metadata tasks.Source: §C.1 · p. 27
C.2 Additional Baselines
- Comparisons include classical post-hoc calibrators and newer feature-aware or grouping-based baselines, with LAPE refinement evaluated on eligible anchors using the same reference split.Baselines include TS/VS, Dirichlet calibration, LTS, ETS, IROvA/MIR, FC, and SAG.Source: §C.2 · p. 27
C.3 Additional Metrics
- The main text emphasizes NLL, ECE, and accuracy; appendix tables add ACE where applicable, and relative changes express LAPE's incremental contribution over an anchor.Lower NLL, ECE, and ACE are better, while higher accuracy is better; a positive relative change indicates improvement after LAPE.Source: §C.3 · pp. 27–28
C.4 Backbones and Experimental Frameworks
- Backbones are chosen by experiment type, using transfer-learning models for images and compact MLPs for tabular and synthetic settings; paired comparisons share hardware and software.The supplementary experiments summarize dataset or diagnostic settings alongside their corresponding backbone or calibration framework.Source: §C.4 · p. 28, Table 3
D.1 Computational Cost
- Inference uses indexed neighbor retrieval followed by neighborhood-size and interpolation-weight optimization, with total per-query cost combining both phases.The spatial index is built once and amortized; PCA can keep the effective dimension small, while a moderate maximum neighborhood size and grid search control optimization cost.Source: §D.1 · pp. 28–29
D.2 Oracle Neighborhood-Variation Diagnostic
- LAPE is closer to oracle probabilities than the base predictor in both neighborhood-variation groups, with a larger reduction in the high-variation group, which remains harder.The diagnostic ranks test points by oracle probability variation among a fixed 80 neighbors; the bottom and top 30% form the low- and high-variation groups.Source: §D.2 · p. 30, Table 5 · p. 30
D.3 Local-Heterogeneity Stress Test
- LAPE-side refinements improve calibration in a controlled GMM setting with neighborhood-specific temperature and class-bias distortions that global calibration cannot fully correct.LTS is relatively competitive for temperature variation, but its temperature-only correction does not fully capture class-bias shifts and more general local residuals.Source: §D.3 · pp. 30–31, Table 6 · p. 31
D.4 Sample-Ratio Sensitivity
- The gain over VS is non-monotone as calibration data increase, narrowing at intermediate ratios before rising again at larger ratios.VS benefits from more calibration samples at intermediate ratios, while its fixed global form eventually reaches a ceiling and LAPE's local empirical component continues to improve.Source: §D.4 · p. 31, Table 7 · p. 31
D.5 Covariate-Shift Robustness
LAPE is evaluated under controlled and real-image distribution shifts, high-dimensional and noisy neighborhoods, different backbone depths, and adaptive-weight stress tests. Results show broad calibration gains, while also identifying conditions that constrain their size.
D.5 Covariate-Shift Robustness
- LAPE improves NLL across nearly all anchor–severity combinations on CIFAR-10-C, with gains generally increasing at stronger Gaussian-noise corruption.Accuracy also shows broad improvements; the ResNet-18 linear probe is trained on clean CIFAR-10, with corrupted target samples used for calibration and evaluation.Source: §D.5 · p. 32, Table 9 · p. 32
D.6 High-Dimensional Stress Test
- At 200 feature dimensions, LAPE's relative gains shrink to around one percent but remain positive for most anchors; structural noise also reduces gains while leaving most positive.The high-dimensional result is consistent with more difficult local neighborhood estimation; under moderate degradation of neighborhood information, LAPE retains positive calibration improvements for most anchors.Source: §D.6 · pp. 33–34, Table 10 · p. 33, Table 11 · p. 33
D.7 Architecture-Depth Diagnostic
- LAPE improves calibration across underfitted, best-matched, and mildly overfitted Obesity backbones, though the improvement magnitude depends on the reliable local information remaining in the representation.Base accuracies are 87.50% for shallow, 93.94% for medium, and 91.29% for deep MLPs; the medium model has the highest accuracy and lowest Base NLL.Source: §D.7 · p. 34, Table 12 · p. 34
D.8 Adaptive-Strategy Ablation
- Under deliberately heterogeneous anchor reliability, adaptive weighting outperforms fixed correction coefficients overall and within each test region.The two zones differ in anchor fit and local residual strength, requiring different correction strengths.Source: §D.8 · pp. 34–35, Table 13 · p. 34
D.9 Evaluations on Modern Vision-Language Models
Prompt tuning changes CLIP prediction prototypes and boundaries while retaining frozen image neighborhoods for retrieval. On DTD, LAPE improves over CoOp and CoOp+DAC, including a 25.87% NLL reduction over CoOp.
D.9 Evaluations on Modern Vision-Language Models
- The DTD evaluation uses a fixed CoOp checkpoint and labeled novel-class reference/test splits, averaged over five stratified splits.CoOp uses 16-shot base-class training, prompt length 16, and 100 epochs; DAC is included as a CLIP-specific calibration anchor.Source: §D.9 · p. 35
- LAPE reduces NLL by 25.87% over CoOp and 13.45% over CoOp+DAC, while improving accuracy by more than 20% in both settings.The experiment uses frozen OpenAI CLIP ViT-B/32 encoders with CoOp prompt tuning; the image embeddings and LAPE retrieval neighborhoods remain unchanged.Source: §D.9 · p. 35, Table 14 · p. 35
- LAPE's gains beyond CoOp+DAC indicate that it can complement a CLIP-specific calibration method in this evaluation.The reported results also support the usefulness of frozen CLIP image neighborhoods for refining prompt-adapted predictions.Source: §D.9 · p. 35
D.10 Compatibility with Training-Time Calibration Objectives
The supplementary studies test LAPE alongside training-time calibration, alternative cross-fitting choices, and multi-view fusion settings. Results report additional calibration improvements, with CFAF-LAPE also improving over real-data fusion baselines and learned dynamic-fusion anchors.
D.10 Compatibility with Training-Time Calibration Objectives
- Adding LAPE improves NLL over CE-, MDCA-, and MbLS-trained predictors, with the MbLS NLL improvement reaching 22.54%.Accuracy also improves when LAPE is applied post hoc using the same calibration protocol.Source: §D.10 · p. 35, Table 15 · p. 35
D.11 Cross-Fitting versus Data Splitting
- Five-fold cross-fitting performs best across all four metrics while retaining all labeled calibration samples, unlike the 50/50 split that sacrifices half the data at each stage.The comparison uses a flexible tree-based anchor and 200 labeled calibration samples, averaged over 20 seeds.Source: §D.11 · p. 36, Table 16 · p. 36
D.12 CFAF Experiments
- CFAF uses an augmented multi-view representation for local correction, addressing cross-view reliability patterns and nonlinear fusion structure without an end-to-end fusion network.The noisy-view stress test includes two weakly informative XOR-related views and three random views; results suggest CFAF can still recover useful cross-view calibration information.Source: §D.12 · p. 36, Table 17 · p. 36
- On UCI Multiple Features and HAM10000 tasks, CFAF-LAPE improves over single-view and late-fusion baselines on the reported real-data comparisons.The HAM10000 experiments combine image and metadata views for two binary tasks.Source: §D.12 · pp. 36–37, Table 18 · p. 37
D.13 Compatibility with Stronger Dynamic-Fusion Anchors
- CFAF-LAPE improves NLL and accuracy over QMF and PDF, with NLL reductions of 44.67% and 25.49%, respectively.Mean ECE and ACE also improve for both anchors; CFAF-LAPE is applied post hoc without modifying the fusion model.Source: §D.13 · p. 37, Table 19 · p. 37
D.14 Additional Tabular and Visual Benchmarks
This section reports supplementary LAPE benchmarks across tabular and visual settings, where NLL improvement generally accompanies preserved accuracy, and provides checklist details on proofs, limitations, reproducibility, research practices, ethical-review variation, and LLM use.
D.14 Additional Tabular and Visual Benchmarks
- Across tabular datasets, LAPE improves NLL for most anchor–dataset pairs, with limited gains on financial fraud; Table 20 reports relative changes versus each anchor.The results indicate additional refinement rather than a standalone ranking across different anchors.Source: §D.14 · p. 37, Table 20 · p. 39
- LAPE serves as a plug-in refinement layer across visual backbones, evaluated with ResNet18 and DenseNet121 features for medical data and frozen ResNet50 features for CIFAR.Table 21 reports relative changes versus the corresponding anchor.Source: §D.14 · pp. 37–38, Table 21 · p. 40
- Across tabular and visual results, LAPE most consistently improves NLL while largely preserving accuracy; ECE improves in many cases or remains comparable when anchors are strong or retrieval geometry is less informative.This pattern is consistent with residual probability refinement that does not substantially change the underlying classifier’s decision rule.Source: §D.14 · p. 38
NeurIPS Paper Checklist
- The checklist states that the paper discusses limitations and provides the assumptions and proofs for Theorem 5.1 in Appendix B.The checklist also reports that the abstract and introduction’s claims match the theoretical and experimental results.Source: NeurIPS Paper Checklist · pp. 41–42
- The checklist points to Sections 6 and Appendices C and D for reproducibility details, Appendix C for data access and experimental settings, and reports non-deterministic statistics averaged across random seeds.It says reproducible code was submitted as supplementary material and would be made public on GitHub after acceptance.Source: NeurIPS Paper Checklist · pp. 42–43
- The checklist says the paper discusses societal impacts of confidence calibration, poses no high-risk release concerns, credits original asset creators, and involves no human-subject or crowdsourcing research.It also states that code is included as a new asset in the supplementary material.Source: NeurIPS Paper Checklist · pp. 44–46
- Ethical-review procedures may vary across institutions and locations, and authors are expected to follow the NeurIPS Code of Ethics and institutional guidelines while preserving anonymity in initial submissions.The checklist advises omitting information that could break anonymity, such as the reviewing institution.Source: NeurIPS Paper Checklist · p. 47
- The paper marks LLM usage N/A because core method development does not use LLMs as an important, original, or non-standard component.The checklist says LLM use limited to writing, editing, or formatting that does not affect methodology, rigor, or originality need not be declared, and directs authors to the NeurIPS handbook policy.Source: NeurIPS Paper Checklist · p. 47