PWLR: Pairwise Witness Local Rejection for Boundary-Aware Out-of-Distribution Detection

Chengyao Jia; Ruixuan Wang

cs.CVAug 16, 2026

PWLR: Pairwise Witness Local Rejection for Boundary-Aware Out-of-Distribution Detection

Authors: Chengyao Jia, Ruixuan Wang

Organizations: Guanghua School of Stomatology Sun Yat-sen University Guangzhou, China Guangdong Provincial Key Laboratory of Stomatology Guangzhou, China · School of Computer Science and Engineering Sun Yat-sen University Guangzhou, China Key Laboratory of Machine Intelligence and Advanced Computing, MOE Guangzhou, China

Abstract

Out-of-distribution (OOD) detection remains challenging for image classifiers, especially when near-OOD samples lie close to in-distribution (ID) class boundaries. Recent vision-language detectors improve OOD detection through class semantics, local prompting, or LLM-generated outlier concepts, but seldom use language as explicit boundary evidence between confusing ID classes. We propose Pairwise Witness Local Rejection (PWLR), which uses an MLLM offline to describe visible local cues that favor one ID class over a specific rival class. These cue phrases are then screened with ID-only data under a frozen vision-language backbone, so that only reliable local verifiers are kept. At inference, PWLR first retains a small set of globally plausible classes, then checks whether any of them is locally supported against its most relevant rivals, and finally combines this pairwise local evidence with the global class score through calibration. Experiments on ImageNet-100 far-OOD, cleaner/challenging OOD and near-OOD benchmarks show that PWLR consistently improves strong vision-language baselines across multiple backbones. Source code will be released.

Figures & tables

Figure 1. Pipeline for pairwise witness generation. Each ID class is first assigned to a coarse visual family, after which a small set of visually confusable rival classes is selected for each anchor class. For each directed anchor-rival pair, a vision-capable MLLM then generates witness phrases that describe local evidence supporting the anchor against the rival, forming a witness bank for later comparative verification. A workflow diagram showing how ID class names are first grouped into coarse visual families, how rival classes are selected for each anchor class, and how GPT-5.1 generates directed witness phrases that are stored in a witness bank.

Figure 2. Pipeline for graph construction and filtering. Text representations and class prototypes are fused to define both global class plausibility and semantic rival neighborhoods. Each witness is then fitted on ID samples under the frozen VLM backbone, and only those whose local responses remain consistently higher on anchor-class samples than on rival-class and other-class samples are retained as reliable local verifiers.

Figure 3. Pipeline for pairwise local witness verification. Given a test image, the pipeline first retains a small set of globally plausible classes using the fused global score, and then verifies each retained class only against its active rivals in the semantic-pruned witness graph. For each directed class pair, witness responses are computed from local views, standardized, and aggregated into a directed support score.

Method	iNaturalist		SUN		Places		Texture		Average
Method	FPR95 $\downarrow$	AUROC $\uparrow$	FPR95 $\downarrow$	AUROC $\uparrow$	FPR95 $\downarrow$	AUROC $\uparrow$	FPR95 $\downarrow$	AUROC $\uparrow$	FPR95 $\downarrow$	AUROC $\uparrow$
MCM	3.72	98.86	13.94	96.00	18.52	94.85	19.06	95.24	13.81	96.24
GL-MCM	3.00	99.22	13.50	96.53	17.62	95.46	17.20	95.93	12.83	96.78
MSP	13.74	95.99	20.10	94.23	25.62	93.53	11.98	97.43	17.86	95.30
Energy	7.32	97.90	16.44	95.33	22.50	94.77	8.64	97.70	13.73	96.43
ReAct	6.04	98.29	13.94	96.00	20.86	95.14	8.82	97.58	12.41	96.75

Table 1. OOD detection performance (%) on the conventional far-OOD benchmark, with ImageNet-100 as ID dataset and iNaturalist, SUN, Places, Textures as OOD datasets. The best and second-best results are shown in bold and underline, respectively.

Method	OpenImage-O		ImageNet-O		NINCO		Average
Method	FPR95 $\downarrow$	AUROC $\uparrow$	FPR95 $\downarrow$	AUROC $\uparrow$	FPR95 $\downarrow$	AUROC $\uparrow$	FPR95 $\downarrow$	AUROC $\uparrow$
MCM	9.10	97.69	16.16	95.83	29.64	94.90	18.30	96.14
GL-MCM	7.96	98.17	14.76	96.33	25.52	95.44	16.08	96.65
MSP	13.24	96.84	18.36	95.62	24.18	94.69	18.59	95.72
Energy	10.38	97.09	15.38	96.18	26.78	94.23	17.51	95.83
ReAct	8.28	97.45	13.42	96.52	28.16	94.17	16.62	96.05

Table 2. OOD detection performance on the more challenging OOD benchmark, with ImageNet-100 as ID dataset and OpenImage-O, ImageNet-O, NINCO as OOD datasets.

Architecture	Method	Pet18/19		CUB100/100
Architecture	Method	FPR95 $\downarrow$	AUROC $\uparrow$	FPR95 $\downarrow$	AUROC $\uparrow$
SigLIP2-L16-256	MCM	38.01	93.48	63.72	84.85
	PWLR-MCM	27.77	94.99	60.07	85.29
	MSP	27.02	94.07	73.14	78.24
	PWLR-MSP	29.91	94.81	70.61	80.76
	NegLabel	27.19	92.81	63.30	84.40

Table 3. Multi-backbone results of different plug-in variants on the split near-OOD benchmarks. We report FPR95/AUROC on Pet18/19 and CUB100/100 under three frozen vision-language backbones.

(a) Core branches				(b) Local comparison mechanisms
	Global	Pairwise	Full		Reverse	Pruning	Full
Global Prior	✓	–	✓	Reverse Penalty	✓	–	✓
Pairwise Verification	–	✓	✓	Rival Pruning	–	✓	✓
FPR95 $\downarrow$	18.67	31.90	16.42	FPR95 $\downarrow$	16.64	16.90	16.42
AUROC $\uparrow$	95.13	93.69	96.23	AUROC $\uparrow$	96.23	96.20	96.44

Table 4. Ablation on the cleaner/challenging benchmarks. (a) Effect of the global prior branch and the pairwise witness verification branch. (b) Effect of two local comparison mechanisms: reverse-edge penalty and active rival pruning. In (b), both core branches are enabled by default.

	$\lambda_{a}$					$\rho$
	0.2	0.3	0.4	0.5	0.6	0.20	0.35	0.50
NegLabel FPR95 $\downarrow$	18.56	18.26	18.34	18.24	18.74	18.24	18.34	18.36
NegLabel AUROC $\uparrow$	95.73	95.93	96.07	96.16	96.16	96.09	96.07	96.07
CSP FPR95 $\downarrow$	20.82	20.54	20.06	19.56	20.06	20.04	20.06	20.16
CSP AUROC $\uparrow$	95.21	95.48	95.69	95.83	95.89	95.71	95.69	95.68

Table 5. Sensitivity analysis of the hyperparameters for PWLR-NegLabel and PWLR-CSP on the cleaner/challenging OOD benchmarks. Bold: best results within each hyperparameter scan. Underlined: the default settings used in the main experiments.

Statistic	Fixed support images
Directed pairs	500
Classes with outgoing pairs	100
Mean out-degree	5
Min out-degree	5
Max out-degree	5
Reverse-edge coverage	12%

Table 6. Statistics of the raw pairwise witness bank under the three-round prompting and cross-round merge protocol. For each directed pair

(a,r)

, we report the number of raw witness phrases accumulated across the three prompting rounds, the number of merged candidates after cross-round duplicate merging, and the average values of the visibility score and the context-risk score. The table also summarizes the induced directed rival structure at the raw-bank stage.

Statistic	Value
Evaluated directed edges	500
Surviving directed edges	493
Evaluated witness phrases	10890
Witnesses kept after ID-only screening	10575
Kept witnesses / surviving edge	21.45
Mean $e_{w}$	1.679

Table 7. Statistics of the ID-only witness screening stage under the three-round witness bank. Starting from the directed witness bank constructed by three-round prompting and cross-round merge, the pipeline screens witness phrases on

D_{\mathrm{fit}}

and retains only those that satisfy the mean-separation and tail-separation conditions in Section 2.2.

Statistic	Raw witness graph	Retained witness graph	Semantic-pruned rival subgraph
Directed edges	500	493	103
Classes with outgoing edges	100	100	100
Mean out-degree	5.00	4.93	1.03
Min out-degree	5.00	4.00	1.00
Max out-degree	5.00	5.00	2.00
Reverse-edge coverage	12.00%	10.76%	7.77%

Table 8. Graph statistics before and after the two contraction steps of the proposed pipeline. The retained witness graph is obtained after the ID-only screening stage, and the semantic-pruned rival subgraph is the final local comparison structure used by online pairwise verification.

Rejection reason	Count	Ratio
Tail-separation failure	272	86.35%
Both failures	43	13.65%

Table 9. Causes of witness rejection under the ID-only screening criteria in Section 2.2, based on the three-round witness bank. Rejected witnesses are categorized by whether they fail the mean-separation condition, the tail-separation condition, or both.

Plug-in	Base detector	PWLR	No pairwise local verification	Shuffled rival assignment
NegLabel	95.31 / 20.98	96.28 $\pm$ 0.09 / 16.97 $\pm$ 0.64	95.58 $\pm$ 0.17 / 20.25 $\pm$ 0.71	96.07 $\pm$ 0.05 / 19.15 $\pm$ 0.41
CSP	94.19 / 23.74	95.91 $\pm$ 0.09 / 18.49 $\pm$ 0.60	94.62 $\pm$ 0.27 / 22.01 $\pm$ 0.89	95.69 $\pm$ 0.18 / 20.97 $\pm$ 0.56

Table 10. Role of pairwise local verification on ImageNet-100 (ID) and NINCO (OOD) under fixed support images. Results are reported as AUROC/FPR95 (%).

Method	Generator	AUROC / FPR95
NegLabel	–	95.31 / 20.98
PWLR-NegLabel	GPT-5.1	96.28 $\pm$ 0.09 / 16.97 $\pm$ 0.64
PWLR-NegLabel	GPT-4o	96.37 $\pm$ 0.11 / 15.96 $\pm$ 0.73
PWLR-NegLabel	Qwen2.5-VL-7B	95.97 $\pm$ 0.15 / 17.20 $\pm$ 0.42
CSP	–	94.19 / 23.74
PWLR-CSP	GPT-5.1	95.91 $\pm$ 0.09 / 18.49 $\pm$ 0.60

Table 11. Comparison across different models for pairwise witness generation on ImageNet-100 (ID) and NINCO (OOD) under fixed support images. Results are reported as AUROC/FPR95 (%).

Method	Enc.	Score	Total	Peak Mem.
	(ms/img)	(ms/img)	(ms/img)	(MB)
NegLabel	30.08	6.81	36.89	3978
CSP	56.03	28.97	85.01	4031
NNGuide	35.94	53.92	89.86	14069
SeTAR	154.82	12.70	167.51	14123
PWLR-NegLabel	130.37	33.41	163.78	5468

Table 12. Average inference overhead per image on a T4 GPU.

Stage	Model	Calls	Total tokens	Time
Family assignment	GPT-4o-mini	9	6,279	31.9 s
Rival selection	GPT-5.1	100	47,710	221.7 s
Witness generation	GPT-5.1	1,500	2,075,365	5.64 h
Total	–	1,609	2,129,354	5.71 h

Table 13. Offline construction cost for the ImageNet-100 witness bank.

Dataset	Method	Base	Main	B (mean $\pm$ std)	C (mean $\pm$ std)	All (mean $\pm$ std)
OpenImage-O	PWLR-NegLabel	97.93 / 10.64	98.60 / 6.72	98.79 $\pm$ 0.02 / 6.06 $\pm$ 0.05	98.75 $\pm$ 0.02 / 6.26 $\pm$ 0.14	98.74 $\pm$ 0.07 / 6.24 $\pm$ 0.25
OpenImage-O	PWLR-CSP	97.41 / 12.20	98.47 / 7.28	98.67 $\pm$ 0.01 / 6.57 $\pm$ 0.10	98.64 $\pm$ 0.02 / 6.91 $\pm$ 0.29	98.63 $\pm$ 0.07 / 6.82 $\pm$ 0.32
ImageNet-O	PWLR-NegLabel	96.36 / 16.46	97.16 / 12.56	97.45 $\pm$ 0.03 / 10.59 $\pm$ 0.55	97.41 $\pm$ 0.06 / 11.36 $\pm$ 0.51	97.39 $\pm$ 0.11 / 11.20 $\pm$ 0.83
ImageNet-O	PWLR-CSP	95.98 / 17.42	97.17 / 12.66	97.46 $\pm$ 0.03 / 10.63 $\pm$ 0.51	97.42 $\pm$ 0.04 / 11.23 $\pm$ 0.35	97.40 $\pm$ 0.11 / 11.18 $\pm$ 0.80
NINCO	PWLR-NegLabel	95.31 / 20.98	96.08 / 18.30	96.48 $\pm$ 0.06 / 15.72 $\pm$ 0.29	96.28 $\pm$ 0.09 / 16.97 $\pm$ 0.64	96.34 $\pm$ 0.16 / 16.62 $\pm$ 1.05
NINCO	PWLR-CSP	94.19 / 23.74	95.70 / 20.18	96.11 $\pm$ 0.06 / 17.12 $\pm$ 0.45	95.91 $\pm$ 0.09 / 18.49 $\pm$ 0.60	95.97 $\pm$ 0.16 / 18.15 $\pm$ 1.21

Table 14. Stability under repeated witness generation on the more challenging benchmark with ImageNet-100 as the ID dataset. Entries are reported as AUROC/FPR95 (%). “Base” denotes the fixed base detector from the main paper. “Main” denotes the single instantiated witness bank reported in the main paper. “B” denotes three independent full witness-bank regenerations, while “C” denotes three independent regenerations with fixed support images. Repeated results are reported as mean

\pm

std over three runs. “All” summarizes the main-paper witness bank together with B1–B3 and C1–C3.

Dataset	Method	Base	Main	B (mean $\pm$ std)	C (mean $\pm$ std)	All (mean $\pm$ std)
CUB100/100	PWLR-NegLabel	84.40 / 63.30	85.09 / 61.97	85.60 $\pm$ 0.55 / 61.14 $\pm$ 2.51	85.63 $\pm$ 0.10 / 60.53 $\pm$ 0.38	85.54 $\pm$ 0.38 / 61.00 $\pm$ 1.56
CUB100/100	PWLR-CSP	84.34 / 64.21	85.19 / 61.89	86.11 $\pm$ 0.54 / 58.92 $\pm$ 2.57	86.16 $\pm$ 0.11 / 59.75 $\pm$ 0.77	86.00 $\pm$ 0.48 / 59.70 $\pm$ 1.87

Table 15. Stability under repeated witness generation on the CUB100/100 split under SigLIP2-L16-256. Entries are reported as AUROC/FPR95 (%). “Base” denotes the fixed base detector from the main manuscript. “Main” denotes the single instantiated witness bank reported in the main manuscript, using the corrected PWLR-CSP AUROC value of 85.19. “B” denotes three independent full witness-bank regenerations, while “C” denotes three independent regenerations with fixed support images. Repeated results are reported as mean

\pm

std over three runs. “All” summarizes the corrected main-paper witness bank together with B1–B3 and C1–C3.

ID dataset	Method	iNaturalist	SUN	Places	Textures	Average
CUB-200-2011	MCM	99.79/0.78	99.85/0.59	99.67/0.74	99.87/0.50	99.80/0.65
	GL-MCM	99.86/ 0.54	99.90 / 0.36	99.71/ 0.57	99.93 / 0.32	99.85/ 0.45
	MSP	99.78/0.91	99.85/0.70	99.67/0.86	99.87/0.57	99.79/0.76
	NegLabel	99.99 / 0.00	100.00 / 0.00	99.90 / 0.00	100.00 / 0.00	99.97 / 0.00
	CSP	99.99 / 0.00	100.00 / 0.00	99.90 / 0.00	100.00 / 0.00	99.97 / 0.00
	PWLR-CSP	100.00 / 0.00	100.00 / 0.00	99.96 / 0.00	100.00 / 0.00	99.99 / 0.00

Table 16. OOD detection performance on additional fine-grained ID datasets under the conventional far-OOD benchmark. Results are reported as AUROC/FPR95 (%). The best and second-best results are shown in bold and underline, respectively.

OOD Detection

Out-of-distribution (OOD) detection identifies inputs unlike a model’s training data so systems can flag uncertainty, reject inputs, or adapt safely.

See all 223 papers

DynProto: Dynamic Prototype Evolution for Out-of-Distribution Detection
Apr 26, 2026Yanqi Wu, Xinhua Lu, Runhe Lai +4Prototype-Based Classification OOD Detection
School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou 510275, China · Peng Cheng Laboratory, Shenzhen 518066, China · Key Laboratory of Machine Intelligence and Advanced Computing, MOE, Guangzhou 510275, China +2
Respecting Modality Gap in Post-hoc Out-of-distribution Detection with Pre-trained Vision-Language Models
May 26, 2026Yuanwei Hu, Bo Peng, Yadan Luo +3Vision-Language Models Test-Time Adaptation
University of Technology Sydney · The University of Queensland
A Theoretical Analysis of Environment-Aware Out-of-Distribution Detection for Vision-Language Models
Oct 19, 2026Salimeh Sekeh, Mary WisellOOD Detection Continual Learning for VLMs
San Diego State University, San Diego, CA, USA

Apr 26, 2026J/K move · Enter open · S save

DynProto: Dynamic Prototype Evolution for Out-of-Distribution Detection

Yanqi Wu, Xinhua Lu, Runhe Lai, Qichao Chen+3

School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou 510275, China · Peng Cheng Laboratory, Shenzhen 518066, China · Key Laboratory of Machine Intelligence and Advanced Computing, MOE, Guangzhou 510275, China+2

Prototype-Based Classification OOD Detection

Open paper AskPDF arXiv

Recent studies show that using potential out-of-distribution (OOD) labels from large corpora as auxiliary information can improve OOD detection in vision-language models (VLMs). However, these methods often fail when real-world OOD samples fall outside the predefined OOD label set. To address this limitation, we propose DynProto, a novel approach that learns OOD prototypes dynamically during testing using only in-distribution (ID) information. DynProto is inspired by a key observation: OOD samples predicted as the same ID class tend to cluster in the feature space. With this insight, we leverage easy-to-detect OOD samples as ``anchors'' to find their harder-to-detect, similar counterparts. To this end, DynProto introduces two modules: Coarse OOD Pattern Capturing Module caches OOD patterns that are easily confused with each ID class during testing, and Fine-grained OOD Pattern Refinement Module subsequently clusters these patterns within each cache and aggregates them into representative OOD prototypes. By measuring similarity to ID and dynamic OOD prototypes, DynProto enables accurate OOD detection. DynProto significantly outperforms prior methods across multiple benchmarks. On ImageNet OOD benchmark, DynProto reduces FPR95 by 11.60% and improves AUROC by 4.70%. Moreover, the framework is architecture-agnostic and can be integrated into various backbones.

Abstract

OOD Detection

Explore similar work

DynProto: Dynamic Prototype Evolution for Out-of-Distribution Detection

Respecting Modality Gap in Post-hoc Out-of-distribution Detection with Pre-trained Vision-Language Models

A Theoretical Analysis of Environment-Aware Out-of-Distribution Detection for Vision-Language Models