REVIEW 24 cited by
A Simple Fix to Mahalanobis Distance for Improving Near-OOD Detection
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
A Simple Fix to Mahalanobis Distance for Improving Near-OOD Detection
read the original abstract
Mahalanobis distance (MD) is a simple and popular post-processing method for detecting out-of-distribution (OOD) inputs in neural networks. We analyze its failure modes for near-OOD detection and propose a simple fix called relative Mahalanobis distance (RMD) which improves performance and is more robust to hyperparameter choice. On a wide selection of challenging vision, language, and biology OOD benchmarks (CIFAR-100 vs CIFAR-10, CLINC OOD intent detection, Genomics OOD), we show that RMD meaningfully improves upon MD performance (by up to 15% AUROC on genomics OOD).
Forward citations
Cited by 24 Pith papers
-
Decodable but Not Detectable: A Leakage Fingerprint for Near-OOD Benchmarks
A trained class left in the detector's fit set and reused as 'OOD' flips AUROC from 0.326 to 0.911 when removed, and a supervised-decodability plus unsupervised-collapse fingerprint detects such leaks (sensitivity 18/...
-
Comparing Linear Probes with Mahalanobis Cosine Similarity
For balanced Gaussian class projections, OOD AUROC is a linear function of MCS to the reference probe because both are sigmoid-shaped functions of the probe SNR on test data.
-
Assessing Sample Quality in Conditional Generation under Compositional Shift
Introduces a per-sample trust score combining global realism and attribute-wise faithfulness, estimable from training data alone, for assessing conditional generations under compositional shift.
-
Beyond Binary Out-of-Distribution Detection: Characterizing Distributional Shifts with Multi-Statistic Diffusion Trajectories
DISC extracts multi-statistic trajectories from diffusion denoising to both detect and classify types of distributional shifts in OOD data.
-
Activation Subspaces for Out-of-Distribution Detection
ActSub decomposes activations via SVD of the classifier head into decisive and insignificant subspaces, using cosine similarity on the insignificant part for far-OOD and shaped energy on the decisive part for near-OOD...
-
Integrating Contextual Embeddings into Evaluation of Expressive MIDI Piano Performances
Kernel distances on CLaMP3/Aria embeddings score expressive MIDI performances about as well as human listeners and catch contextual corruptions invisible to attribute statistics.
-
VLA-FAIL: Efficient Task Failure Detection for Finetuned Vision-Language-Action Models
VLA-FAIL introduces last-layer Mahalanobis distance and action chunk consistency detectors that together enable early, reliable failure detection in finetuned VLAs without failure data or expensive sampling.
-
From Local Geometry to Global Pseudo Labeling for Robust Positive Unlabeled Learning under Covariate Shift
SPUNA leverages spectral neighborhood annotation on visual feature manifolds to enable robust PU learning for covariate shift detection, matching fully supervised performance.
-
Temporal Reasoning Is Not the Bottleneck: A Probabilistic Inconsistency Framework for Neuro-Symbolic QA
Temporal reasoning is not the core bottleneck for LLMs on time-based QA; the real issue is unstructured text-to-event mapping, addressed by a neuro-symbolic system with PIS that reaches 100% accuracy on benchmarks whe...
-
Component-Based Out-of-Distribution Detection
CoOD decomposes inputs into components and applies Component Shift Score plus Compositional Consistency Score to improve detection of both standard and compositional out-of-distribution data.
-
HyCal: A Training-Free Prototype Calibration Method for Cross-Discipline Few-Shot Class-Incremental Learning
HyCal mitigates Domain Gravity in cross-discipline imbalanced few-shot class-incremental learning by calibrating prototypes with complementary directional and covariance-aware distances on frozen CLIP embeddings.
-
Hidden Failures in Robustness: Why Supervised Uncertainty Quantification Needs Better Evaluation
Supervised uncertainty probes for LLMs show poor robustness under distribution shift, with middle-layer representations and multi-token aggregation proving more reliable than final-layer or single-token features.
-
Mind the Way You Select Negative Texts: Pursuing the Distance Consistency in OOD Detection with VLMs
InterNeg improves OOD detection in VLMs by using inter-modal distance criteria for negative text selection and by inverting high-confidence OOD images into additional negative text embeddings.
-
Language Models (Mostly) Know What They Know
Language models show good calibration when asked to estimate the probability that their own answers are correct, with performance improving as models get larger.
-
Dual Feature Decoupling for Fine-Grained OOD Detection
DFDNet disentangles content from style via dual modules to boost fine-grained OOD detection performance on multiple datasets.
-
Evaluating the Relevance of Uncertainty Estimators for LLM Hallucination
Empirical study across multiple benchmarks finds the link between uncertainty estimators and LLM hallucinations is highly variable and often weak.
-
MahaVar: OOD Detection via Class-wise Mahalanobis Distance Variance under Neural Collapse
MahaVar augments the Mahalanobis OOD score with class-wise distance variance, which is theoretically higher for in-distribution samples under relaxed Neural Collapse geometry.
-
RankOOD -- Class Ranking-based Out-of-Distribution Detection
RankOOD detects out-of-distribution samples by training a model to predict fixed class-specific ranking permutations via the Plackett-Luce loss, achieving a 4.3% FPR95 reduction on near-OOD TinyImageNet.
-
Medix: Out-of-Distribution Detection from Unlabeled Wild Data via Robust Gradient Statistics
Medix uses element-wise median gradients to pick outliers from unlabeled wild data and trains an OOD detector, claiming theoretical bounds that do not actually cover the greedy algorithm.
-
DCV-ROOD Evaluation Framework: Dual Cross-Validation for Robust Out-of-Distribution Detection
DCV-ROOD is a dual cross-validation framework for OOD detection that splits ID data by stratified folds and OOD data by class groups, reproducing benchmark statistical comparisons at lower cost.
-
Rapidly Adapting to New Voice Spoofing: Few-Shot Detection of Synthesized Speech Under Distribution Shifts
Few-shot adaptation with a self-attentive prototypical network reduces spoofed-speech detection errors by up to 32% relative under distribution shifts.
-
Uncertainty-Driven Reliability: Selective Prediction and Trustworthy Deployment in Modern Machine Learning
A training-dynamics abstention method matches deep ensembles at a fraction of the training cost, and a five-term error budget explains why selective classifiers still fall short of the oracle.
-
MM++: Unsupervised Scale-Invariant Multilayer OOD Detection via Top-K Gated Feature Fusion
MM++ fuses entropy-selected intermediate layers with terminal features via Ledoit-Wolf regularized covariance for scale-invariant unsupervised OOD detection across architectures.
-
Designing Practical Models for Isolated Word Visual Speech Recognition
The submission is internally mismatched: the VSR abstract has no supporting text because the body is PGTuner, a different cs.DB paper (arXiv:2508.17886).
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.