Paper deep dive
From Subtle to Significant: Prompt-Driven Self-Improving Optimization in Test-Time Graph OOD Detection
Luzhi Wang, Xuanshuo Fu, He Zhang, Chuang Liu, Xiaobao Wang, Hongbo Liu
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 7/21/2026, 12:02:14 AM
Summary
The paper introduces SIGOOD, an unsupervised self-improving framework for test-time Graph Out-of-Distribution (OOD) detection. It addresses limitations of one-pass inference by using a prompt-driven approach to iteratively amplify OOD signals. SIGOOD employs an Energy Preference Optimization (EPO) loss to optimize prompts based on energy variations between original and prompt-enhanced graphs, effectively distinguishing ID from OOD graphs without access to training data or labels.
Entities (6)
Relation Signals (5)
SIGOOD → solves → Graph OOD Detection
confidence 95% · we propose a Self-Improving Graph Out-of-Distribution detector (SIGOOD)... for effective graph OOD detection
SIGOOD → uses → Energy Preference Optimization
confidence 95% · To optimize prompts, SIGOOD introduces an Energy Preference Optimization (EPO) loss
SIGOOD → generates → Prompt-Enhanced Graph
confidence 92% · SIGOOD generates a prompt to construct a prompt-enhanced graph that amplifies potential OOD signals
SIGOOD → employs → Test-time Training
confidence 90% · integrates continuous self-learning with test-time training for effective graph OOD detection
Energy Preference Optimization → measures → Energy Variations
confidence 88% · EPO loss, which leverages energy variations between the original test graph and the prompt-enhanced graph
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Graph Out-of-Distribution (OOD) detection aims to identify whether a test graph deviates from the distribution of graphs observed during training, which is critical for ensuring the reliability of Graph Neural Networks (GNNs) when deployed in open-world scenarios. Recent advances in graph OOD detection have focused on test-time training techniques that facilitate OOD detection without accessing potential supervisory information (e.g., training data). However, most of these methods employ a one-pass inference paradigm, which prevents them from progressively correcting erroneous predictions to amplify OOD signals. To this end, we propose a \textbf{S}elf-\textbf{I}mproving \textbf{G}raph \textbf{O}ut-\textbf{o}f-\textbf{D}istribution detector (SIGOOD), which is an unsupervised framework that integrates continuous self-learning with test-time training for effective graph OOD detection. Specifically, SIGOOD generates a prompt to construct a prompt-enhanced graph that amplifies potential OOD signals. To optimize prompts, SIGOOD introduces an Energy Preference Optimization (EPO) loss, which leverages energy variations between the original test graph and the prompt-enhanced graph. By iteratively optimizing the prompt by involving it into the detection model in a self-improving loop, the resulting optimal prompt-enhanced graph is ultimately used for OOD detection. Comprehensive evaluations on 21 real-world datasets confirm the effectiveness and outperformance of our SIGOOD method. The code is at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2602.17342v1
- Canonical: https://arxiv.org/abs/2602.17342v1
Trouble viewing inline? Open PDF directly →
Full Text
53,186 characters extracted from source content.
Expand or collapse full text
From Subtle to Significant: Prompt-Driven Self-Improving Optimization in Test-Time Graph OOD Detection Luzhi Wang1, Xuanshuo Fu2, He Zhang3, Chuang Liu1, Xiaobao Wang4, Hongbo Liu1 Corresponding Author. Abstract Graph Out-of-Distribution (OOD) detection aims to identify whether a test graph deviates from the distribution of graphs observed during training, which is critical for ensuring the reliability of Graph Neural Networks (GNNs) when deployed in open-world scenarios. Recent advances in graph OOD detection have focused on test-time training techniques that facilitate OOD detection without accessing potential supervisory information (e.g., training data). However, most of these methods employ a one-pass inference paradigm, which prevents them from progressively correcting erroneous predictions to amplify OOD signals. To this end, we propose a Self-Improving Graph Out-of-Distribution detector (SIGOOD), which is an unsupervised framework that integrates continuous self-learning with test-time training for effective graph OOD detection. Specifically, SIGOOD generates a prompt to construct a prompt-enhanced graph that amplifies potential OOD signals. To optimize prompts, SIGOOD introduces an Energy Preference Optimization (EPO) loss, which leverages energy variations between the original test graph and the prompt-enhanced graph. By iteratively optimizing the prompt by involving it into the detection model in a self-improving loop, the resulting optimal prompt-enhanced graph is ultimately used for OOD detection. Comprehensive evaluations on 21 real-world datasets confirm the effectiveness and outperformance of our SIGOOD method. Introduction Graph neural networks (GNNs) offer a powerful paradigm for graph representation learning (Wang et al. 2025b; Fu et al. 2025; Zhang et al. 2025a), which is widely applied in tasks such as sarcasm detection (Wang et al. 2025a, 2023), recommender systems (Jin et al. 2023a), and fraud detection (Pan et al. 2025). Most GNNs assume that training and test graphs are from the same distribution (in-distribution, ID) (Zhang et al. 2024a). However, when these well-trained GNNs are deployed in open-world scenarios, they inevitably encounter out-of-distribution (OOD) graphs (Shen et al. 2024), leading to misprediction risks. For instance, a GNN might misclassify a structurally similar yet distributionally distinct five-membered lactone molecule as aspirin (a typical drug molecule), resulting in prediction errors in drug identification (Wang et al. 2025c). Thus, detecting OOD graphs is essential for ensuring the reliability of GNNs. Figure 1: Comparisons between existing methods and SIGOOD. (a) Existing methods perform a one-pass pattern extraction, aiming to amplify the difference between ID and OOD graphs (Wang et al. 2024). (b) SIGOOD adopts an unsupervised self-improvement strategy, progressively refining OOD signals through iterative optimization. Recent advances in graph OOD detection have attracted increasing attention (Lin et al. 2025). One line of these studies considers that pattern discrepancies exist between ID and OOD graphs (Ding et al. 2024). Consequently, these methods rely on training data to learn ID graph patterns for OOD discrimination (Hou et al. 2025), but such data may be unavailable due to privacy, storage, or deployment constraints (Wang et al. 2024). Therefore, the second line of studies has focused on test-time graph OOD detection, which aims to identify distribution shifts using only the well-trained GNN and test graphs (Jin et al. 2023b). However, these methods are suboptimal when amplifying the difference of ID and OOD core information across graphs, as their one-pass unsupervised rationales neglect to leverage the latent information in test graphs through iterative refinement. The above limitations motivate us to develop a new test-time graph OOD detection method, but it is not trivial due to the following challenges. Particularly, Challenge 1: Distribution overlapping. A well-trained GNN primarily learns the decision boundaries from ID graphs, without explicitly acquiring the ability to recognize unknown patterns. Thus, in the embedding space, the GNN tends to project OOD graph embeddings blindly toward the ID graph embeddings, leading to the distributional overlap between ID and OOD data (Wu et al. 2024). This phenomenon makes it difficult to identify OOD graphs that have subtle structural or feature differences from ID graphs. Challenge 2: Unknown ID/OOD labels. Under the test-time OOD detection setting, the detection model has no access to task labels or other information (e.g., training data), which raises difficulty in effectively extracting OOD patterns with unlabeled test graphs. Challenge 3: Effective optimization strategies. Due to the absence of task labels, it is challenging to design a self-improving optimization strategy that can continuously learn OOD signals from the test graph itself, and effectively enhance the model’s OOD detection performance. To address aforementioned challenges, we propose SIGOOD, an unsupervised self-improving graph OOD detection framework that iteratively enhances OOD signals at test time via energy-based feedback. According to previous studies (Zhang et al. 2024d; Wu et al. 2023), the energy based on prediction logits represents the probability of a graph being OOD, where a high energy score indicates a high likelihood of OOD. Since OOD signals are composed of key nodes that sufficiently represent OOD characteristics, we introduce prompts to amplify OOD signals by capturing the average node-level energy variations before and after prompt enhancement, thus optimizing the objective of OOD detection. Specifically, to address Challenge 1, SIGOOD firstly utilizes the well-trained GNN to obtain the embedding of a test graph, and then injects auxiliary prompt into it, aiming to amplify the differences between OOD patterns and ID patterns in it. To tackle Challenge 2, SIGOOD utilizes the energy variation between the test graph and the prompt-enhanced (PE) graph as a signal to indicate OOD or ID preference, which actively discover and iteratively refine ID and OOD patterns. To address Challenge 3, we propose a tailored loss function that leverages the energy variations induced by prompt injection. This loss amplifies the energy differences between OOD and ID patterns within the PE graph to facilitate more effective OOD detection. Unlike prior one-pass methods, our approach involves the prompt-enhanced graph in an iterative optimization loop, supporting a self-improving process. Fig. 1 illustrates the differences between SIGOOD and other test-time graph OOD detection methods. Our contributions are summarized as follows: • To the best of our knowledge, we propose the first self-improving framework for test-time graph OOD detection, which progressively enhances the OOD detection performance with only relying on test data. • We investigate the role of energy variation in capturing OOD signals and propose an energy preference optimization (EPO) loss to enhance the energy contrast between OOD and ID signals. • Extensive experiments on 2121 benchmark datasets and comparisons with 1212 state-of-the-art methods demonstrate the superior performance of SIGOOD. Related Work Graph OOD Detection Graph OOD detection aims to identify test graphs whose distributions deviate from the ID graphs (Lin et al. 2024; Zhang et al. 2025b). Existing methods generally use pre-trained encoders with post-hoc detectors, such as distance or energy-based scores, which are static detection methods that lack any training-time or test-time optimization process (Fuchsgruber et al. 2024). Several studies focus on improving ID recognition rather than detecting OOD graphs (Cao et al. 2025). Recent test-time methods amplify ID-OOD differences but lack iterative refinement mechanisms to progressively enhance OOD signals (Zhang et al. 2024c). In contrast, our SIGOOD model introduces a novel energy-based self-improving mechanism to iteratively mine OOD patterns without label or training supervision. Test-time Training Test-Time Training (T) is a general approach designed to enhance the performance of predictive models when there is a distribution shift between training and test data (Liang et al. 2025). In addition to the computer vision (Dalal et al. 2025), natural language processing (Zhang et al. 2024b), and multimodal foundation models (Bi et al. 2025), T technology is also widely used to improve graph OOD generalization (Zheng et al. 2024). Nevertheless, the application of T techniques to graph OOD detection remains largely underexplored. In our work, we leverage T in a self-improving manner to identify OOD signals of graphs, thereby enabling effective test-time OOD detection. Preliminaries Graphs. Given a graph G=(V,E,X)G=(V,E,X), the node set is indicated by V=v1,v2,…,vnV=\v_1,v_2,…,v_n\, E=e1,e2,…,emE=\e_1,e_2,…,e_m\ denotes the set of edges, and X∈ℝn×dX ^n× d represents the d-dimensional node feature matrix. Each graph is associated with an adjacency matrix A∈ℝn×nA ^n× n, where Aij=1A_ij=1 indicates the presence of an edge between nodes viv_i and vjv_j, and Aij=0A_ij=0 indicates that no edge exists between viv_i and vjv_j. Figure 2: Overview of SIGOOD. Step 1: Given a test graph GtG_t as an input, SIGOOD first encodes it using the well-trained GNN. The obtained embedding is then passed to the prompt generator (PG), which produces a prompt PmP_m to enhance OOD signals of the graph. Step 2: The prompt PmP_m is applied to GtG_t, yielding a prompt-enhanced (PE) graph GpG_p with amplified OOD tendency. Step 3: Calculate node-wise energy variations between GpG_p and GtG_t to locate sensitive nodes of OOD signals. Step 4: Calculate the global energy variations between GpG_p and GtG_t to evaluate the overall OOD tendency of GpG_p. These energy variations are used as OOD signals to guide the optimization of the prompt generator. The updated GpG_p replaces GtG_t as the input for the next iteration. After convergence, the final GpG_p is used to calculate the OOD score for detection. Test-time Graph OOD Detection. During the inference stage of a GNN model, the test-time graph out-of-distribution detection task aims to determine whether an input graph is drawn from a distribution different from that of the training data. At test time, neither the original training data nor auxiliary information such as task-specific labels for OOD detection is accessible. Formally, the test-time graph OOD detection can be defined as: Detection label=1(OOD),ifD(Gt)⩾τ0(ID),otherwise Detection label= cases1\ (OOD),&if\ D(G_t) τ\\ 0\ (ID),&otherwise\\ cases (1) where D(Gt) D(G_t) denotes a scoring function that measures the discrepancy between the test graph GtG_t and the in-distribution data, and τ is a predefined detection threshold. Method Overview According to existing studies (Liu et al. 2020), energy scores (e.g., the negative log-likelihood of sample predictions) reflect model confidence during the inference phase. For example, a higher energy value (i.e., a lower likelihood) indicates lower confidence in model predictions, suggesting a greater probability that the test sample is OOD data (Chen et al. 2023). In this paper, our intuition is that the subtle difference between ID and OOD data can be amplified by introducing additional prompts, thereby increasing the energy-based distinction for OOD detection. Due to the inductive bias of a well-trained GNN on ID graphs (Wu et al. 2020), introducing prompts into a test graph generally leads to a greater energy change in its OOD components (e.g., nodes) than in its ID components. This difference in energy change is then used to iteratively identify potential OOD signals in the test graph and guide the prompt generator toward optimal detection performance. Fig. 2 illustrates the overview of our SIGOOD method. Given a test graph, it is first embedded by a well-trained GNN, and then an optimizable prompt is utilized to generate a prompt-enhanced graph that amplifies the OOD signal. Under the guidance of energy-based loss function (EPO loss), the prompt generator is optimized to enhance OOD signals and the corresponding PE graph is regarded as the updated test graph in the self-improving loop. The specific steps are described in detail below. Step 1: Graph Prompt Generator Given a test graph GtG_t, SIGOOD leverages a well-trained GNN to obtain its node embedding h. However, due to the distribution overlapping between ID and OOD data, it is not trivial to conduct test-time OOD detection by directly using the raw embeddings of test graph. Therefore, we use a graph prompt generator to dynamically guides and improves the distribution distinguishability of ID and OOD data. Specifically, the prompt generator PG(⋅)PG(·) is implemented as a lightweight three-layer MLP modle (Rumelhart et al. 1986), which rapidly generates OOD prompts during the testing phase of GNNs. It is formally defined as : v∗=ReLU(2⋅ReLU(1v+b1)+b2), v^*= ReLU(W_2· ReLU(W_1v+b_1)+b_2), (2) PG(v)=3(γ⋅v∗−μσ2+ϵ+λ)+b3, PG(v)=W_3 (γ· v^*-μ σ^2+ε+λ )+b_3, where v is the node embedding, 1W_1, 2W_2, 3W_3 are trainable weight matrices, and γ, λ, μ, σ represent learnable parameters for normalization. The prompt PmP_m is composed of all nodes re-embedded by the prompt generator PG(v)PG(v). These parameters allow the generator to effectively guide the distinguishing on graphs. Step 2: Prompt Injection Previous studies have shown that graph components, such as certain nodes or subgraphs, typically contain specific signals that represent their ID or OOD distribution patterns (Yu et al. 2023). Motivated by this, SIGOOD uses the generated prompt to facilitate the recognition of OOD signals included in a test graph GtG_t. Specifically, SIGOOD integrates the generated prompt PmP_m with GtG_t to construct a prompt-enhanced graph GpG_p: Gp=Gt⊕Pm,G_p=G_t P_m, (3) where ⊕ denotes element-wise addition between the corresponding node embeddings of GtG_t and its prompt PmP_m. By optimizing the prompt generator, GpG_p is obtained by improving the OOD signals while suppressing the ID signals in GtG_t, thus enhancing the separability between ID and OOD data. Step 3: OOD Signal Recognition A well-trained GNN exhibits strong inductive bias toward ID graphs (Wu et al. 2020). Therefore, if a test graph is ID, the representation generated by the well-trained GNN has a high probability to align with the training distribution. Otherwise, its embeddings are less aligned with the training distribution. Given that energy is a transformed form of negative log-probability, existing OOD detection methods often leverage energy as a scoring function to distinguish ID and OOD graphs (Jiang et al. 2025). However, directly observing the energy provides only a static information and fails to capture the model’s dynamic response after prompt injection. To further exploit energy scores, SIGOOD analyzes node-wise energy variations between the original test graph GtG_t and the prompt-enhanced graph GpG_p. The calculation of energy variations begins by computing the energy scores of each node in both GtG_t and GpG_p. SIGOOD adds a lightweight scoring head on the well-trained GNN to produce a 2-dimensional logit for representing the ID and OOD preference. For a node embedding v, let f(v)∈ℝ2f(v) ^2 denote logit function, the energy of node v is defined as: E^(v)=−log∑i=12exp(fi(v)). E(v)=- log _i=1^2 exp(f_i(v)). (4) After computing the energy of each node in both the original graph GtG_t and the PE graph GpG_p, we quantify the energy variation of each corresponding node to emphasize relative changes. Specifically, the energy variation is defined as: ΔEv E_v =logE^(v;Gp)−logE^(u;Gt) = log E(v;G_p)- log E(u;G_t) (5) =logE^(v;Gp)E^(u;Gt), = log E(v;G_p) E(u;G_t), where u∈Gtu∈ G_t is the corresponding node of v∈Gpv∈ G_p. The energy variation ΔEv E_v provides a directional indicator for identifying and amplifying potential OOD signals. Following the intuition in the overview, SIGOOD identifies nodes with ΔE>0 E>0 as OOD nodes voodv_ood, whereas nodes with ΔE<0 E<0 are regarded as ID nodes vidv_id. Step 4: Feedback Optimization To optimize SIGOOD, we amplify the energy variation between the prompt-enhanced graph and the original test graph to strengthen the OOD signal. We proposed an energy preference optimization (EPO) loss ℒEPOL_EPO, which is defined as: ℒEPO=−logσ(βlogvood∼Gp[E^(vood;Gp)E^(uood;Gt)] _EPO\ =- logσ (β logE_v_ood G_p [ E(v_ood;G_p) E(u_ood;G_t) ] (6) −βlogvid∼Gp[E^(vid;Gp)E^(uid;Gt)]). -β logE_v_id G_p [ E(v_id;G_p) E(u_id;G_t) ] ). The EPO loss consists with an energy-based Bradley-Terry model (Bradley and Terry 1952) and Kullback–Leibler (KL) divergence (Kullback and Leibler 1951). The implementation of EPO is as follows. Given the ID/OOD region signal in Step 33, SIGOOD optimizes the prompt generator to enhance the distinguishability of OOD data. An expected loss is that which can increase the difference between ID and OOD signals to facilitate the detection of OOD data. An intuitive optimization strategy is the probability P of OOD signals is more preferred by the model than ID signals. Inspired by the Bradley-Terry model, the probability P is formulated as (vood≻vid)=exp(r(Pm,vood))exp(r(Pm,vood))+exp(r(Pm,vid))P(v_ood v_id)= (r(P_m,v_ood)) exp(r(P_m,v_ood))+ exp(r(P_m,v_id)), where r(Pm,v)r(P_m,v) represents a reward function modeling the joint effect of the prompt PmP_m and node v, designed to enhance OOD signals in graphs. Then we maximize the log-likelihood of such pairwise comparisons, the optimization goal is defined as: Max(logσ(r(Pm,vood)−r(Pm,vid))). Max( logσ(r(P_m,v_ood)-r(P_m,v_id))). (7) SIGOOD aims to optimize the OOD signals in the original graph GpG_p under the guidance of a reward function r(⋅)r(·), while ensuring that the PE graph GpG_p does not deviate excessively from GtG_t. To achieve this, SIGOOD introduces the Kullback–Leibler (KL) divergence constraint to regularize the difference between GpG_p and GtG_t. Specifically, SIGOOD defines an energy-based KL-divergence as follows: DKL=(E^(v;Gp)||E^(u;Gt))D_KL=( E(v;G_p)|| E(u;G_t)). Inspired by DPO (Rafailov et al. 2023) reward mechanisms, the reward function of SIGOOD can be defined as: r(Pm,v) r(P_m,v) =βlogE^(v;Gp)E^(u;Gt), =β log E(v;G_p) E(u;G_t), (8) where u is the correspond node of v in graph GtG_t. The detailed derivation is provided in the Appendix. Combined with Eq.(5), the reward function r(Pm,v)r(P_m,v) can be interpreted as the energy variation induced by parameter amplification. This observation suggests that energy variations can serve as an effective optimization objective for enhancing OOD signals. Furthermore, by incorporating the reward function into Eq.(7), the optimization goal can be formulated as: Max(logσ(βlogE^(vood;Gp)E^(uood;Gt)−βlogE^(vid;Gp)E^(uid;Gt))). Max ( logσ (β log E(v_ood;G_p) E(u_ood;G_t)-β log E(v_id;G_p) E(u_id;G_t) ) ). (9) To optimize the prompt generator, we contrast the energy variation of OOD and ID nodes between the PE graph GpG_p and the input graph GtG_t. The first term encourages energy of OOD nodes to increase after prompt injection, while the second term penalizes excessive energy variation in ID nodes. By maximizing their difference, SIGOOD promotes OOD-signal responses while maintaining stability for ID nodes. While the aforementioned optimization objectives primarily focus on identifying and amplifying local OOD signals, the core task of SIGOOD is graph-level OOD detection. Therefore, relying solely on local signals is insufficient. To comprehensively assess whether a test graph deviates from the training distribution, we further incorporate graph-level energy variation as a global optimization signal, aligning with the model’s preference for global energy shifts. Specifically, we compute the average energy increase and decrease across all nodes in graphs, capturing the overall energy variation to reflect global OOD tendencies. The final EPO loss, is shown in Eq.(6). Guided by energy-based feedback, the prompt generator is iteratively updated to amplify the OOD signal in GpG_p. The refined GpG_p replaces GtG_t as the new input for subsequent iterations, forming a self-improving loop that continuously strengthens SIGOOD’s ability to detect OOD patterns. After training, the input graph GtG_t is combined with the optimized prompt to produce a prompt-enhanced graph. This PE graph is then evaluated using the EPO loss, which serves as the OOD detection scores. A predefined threshold τ is applied to determine whether the graph is ID or OOD. Why SIGOOD is effective? Well-trained GNNs exhibit strong expressive power in modeling ID patterns, resulting in high confidence when classifying ID graphs. However, they often show low confidence and unstable predictions for OOD graphs. Energy, which serves as a transformed representation of the GNNs output probability, is widely used as an indicator for OOD detection, where higher energy values suggest a greater likelihood of being OOD. Leveraging this property, SIGOOD initially identifies potential OOD signals through energy variations. Guided by energy-based feedback, the prompt generator is iteratively optimized to produce prompt-enhanced graphs that better expose OOD patterns. These enhanced graphs are then used as new test inputs GtG_t in the next iteration. Through this iterative refinement, SIGOOD progressively corrects prior mispredictions, forming a self-improving optimization loop that enhances OOD detection performance over time. Experiments We evaluate the effectiveness of SIGOOD on both graph OOD detection and anomaly detection tasks. The following sections present the datasets, baseline methods, implementation details, and experimental results. Experimental Setups Table 1: OOD detection results in terms of AUC score (%\%). The best results are highlighted with bold. ID dataset BZR PTC-MR AIDS ENZYMES Tox21 FreeSolv ClinTox Esol Avg. OOD dataset COX2 MUTAG DHFR PROTEIN SIDER ToxCast LIPO MUV Rank PK-LOF 42.22±8.3942.22 ± 8.39 51.04±6.0451.04 ± 6.04 50.15±3.2950.15 ± 3.29 50.47±2.8750.47 ± 2.87 51.33±1.8151.33 ± 1.81 49.16±3.7049.16 ± 3.70 50.00±2.1750.00 ± 2.17 50.82±1.4850.82 ± 1.48 11.111.1 PK-OCSVM 42.55±8.2642.55 ± 8.26 49.71±6.5849.71 ± 6.58 50.17±3.3050.17 ± 3.30 50.46±2.7850.46 ± 2.78 51.33±1.8151.33 ± 1.81 48.82±3.2948.82 ± 3.29 50.06±2.1950.06 ± 2.19 51.00±1.3351.00 ± 1.33 11.011.0 PK-iF 51.46±1.6251.46 ± 1.62 54.29±4.3354.29 ± 4.33 51.10±1.4351.10 ± 1.43 51.67±2.6951.67 ± 2.69 49.87±0.8249.87 ± 0.82 52.28±1.8752.28 ± 1.87 50.81±1.1050.81 ± 1.10 50.85±3.5150.85 ± 3.51 8.58.5 WL-LOF 48.99±6.2048.99 ± 6.20 53.31±8.9853.31 ± 8.98 50.77±2.8750.77 ± 2.87 52.66±2.4752.66 ± 2.47 51.92±1.5851.92 ± 1.58 51.47±4.2351.47 ± 4.23 51.29±3.4051.29 ± 3.40 51.26±1.3151.26 ± 1.31 8.08.0 WL-OCSVM 49.16±4.5149.16 ± 4.51 53.31±7.5753.31 ± 7.57 50.98±2.7150.98 ± 2.71 51.77±2.2151.77 ± 2.21 51.08±1.4651.08 ± 1.46 50.38±3.8150.38 ± 3.81 50.77±3.6950.77 ± 3.69 50.97±1.6550.97 ± 1.65 9.09.0 WL-iF 50.24±2.4950.24 ± 2.49 51.43±2.0251.43 ± 2.02 50.10±0.4450.10 ± 0.44 51.17±2.0151.17 ± 2.01 50.25±0.9650.25 ± 0.96 52.60±2.3852.60 ± 2.38 50.41±2.1750.41 ± 2.17 50.61±1.9650.61 ± 1.96 10.010.0 InfoGraph-iF 63.17±9.7463.17 ± 9.74 51.43±5.1951.43 ± 5.19 93.10±1.3593.10 ± 1.35 60.00±1.8360.00 ± 1.83 56.28±0.8156.28 ± 0.81 56.92±1.6956.92 ± 1.69 48.51±1.8748.51 ± 1.87 54.16±5.1454.16 ± 5.14 6.46.4 InfoGraph-MD 86.14±6.7786.14 ± 6.77 50.79±8.4950.79 ± 8.49 69.02±11.6769.02 ± 11.67 55.25±3.5155.25 ± 3.51 59.97±2.0659.97 ± 2.06 58.05±5.4658.05 ± 5.46 48.12±5.7248.12 ± 5.72 77.57±1.6977.57 ± 1.69 6.16.1 GraphCL-iF 60.00±3.8160.00 ± 3.81 50.86±4.3050.86 ± 4.30 92.90±1.2192.90 ± 1.21 61.33±2.2761.33 ± 2.27 56.81±0.9756.81 ± 0.97 55.55±2.7155.55 ± 2.71 47.84±0.9247.84 ± 0.92 62.12±4.0162.12 ± 4.01 7.07.0 GraphCL-MD 83.64±6.0083.64 ± 6.00 73.03±2.3873.03 ± 2.38 93.75±2.1393.75 ± 2.13 52.87±6.1152.87 ± 6.11 58.30±1.5258.30 ± 1.52 60.31±5.2460.31 ± 5.24 51.58±3.6451.58 ± 3.64 78.73±1.4078.73 ± 1.40 3.83.8 GTrans 55.17±5.0455.17 ± 5.04 62.38±2.3662.38 ± 2.36 60.12±1.9860.12 ± 1.98 49.94±5.6749.94 ± 5.67 61.67±0.7361.67 ± 0.73 50.81±3.0350.81 ± 3.03 58.54±2.3858.54 ± 2.38 76.31±3.8576.31 ± 3.85 6.56.5 GOODAT 82.16±0.1582.16 ± 0.15 81.84±0.5781.84 ± 0.57 96.43±0.2596.43 ± 0.25 66.29±1.5466.29 ± 1.54 68.92±0.0168.92 ± 0.01 68.83±0.0268.83 ± 0.02 62.46±0.5462.46 ± 0.54 85.91±0.2785.91 ± 0.27 2.32.3 Ours 87.36±0.1787.36 ± 0.17 85.70±0.0385.70 ± 0.03 97.38±0.0197.38 ± 0.01 67.88±0.0467.88 ± 0.04 69.97±0.6869.97 ± 0.68 68.89±0.2468.89 ± 0.24 71.33±0.1471.33 ± 0.14 87.72±0.0587.72 ± 0.05 1 Improve 1.41%1.41\% 4.72%4.72\% 0.99%0.99\% 2.40%2.40\% 1.52%1.52\% 0.09%0.09\% 14.20%14.20\% 2.11%2.11\% −- Table 2: Anomaly detection results in terms of AUC score (%\%). The best results are highlighted with bold. Method PK-OCSVM PK-iF WL-OCSVM WL-iF InfoGraph-iF GraphCL-iF GTrans GOODAT Ours Improve PROTEINS-full 50.49±4.9250.49 ± 4.92 60.70±2.5560.70 ± 2.55 51.35±4.3551.35 ± 4.35 61.36±2.5461.36 ± 2.54 57.47±3.0357.47 ± 3.03 60.18±2.5360.18 ± 2.53 60.16±5.0660.16 ± 5.06 77.92±2.3777.92 ± 2.37 79.54±0.6079.54 ± 0.60 2.07%2.07\% ENZYMES 53.67±2.6653.67 ± 2.66 51.30±2.0151.30 ± 2.01 55.24±2.6655.24 ± 2.66 51.60±3.8151.60 ± 3.81 53.80±4.5053.80 ± 4.50 53.60±4.8853.60 ± 4.88 38.02±6.2438.02 ± 6.24 52.33±4.7452.33 ± 4.74 76.80±0.8976.80 ± 0.89 39.02%39.02\% DHFR 47.91±3.7647.91 ± 3.76 52.11±3.9652.11 ± 3.96 50.24±3.1350.24 ± 3.13 50.29±2.7750.29 ± 2.77 52.68±3.2152.68 ± 3.21 51.10±2.3551.10 ± 2.35 61.15±2.8761.15 ± 2.87 61.52±2.8661.52 ± 2.86 65.17±0.4565.17 ± 0.45 5.93%5.93\% BZR 46.85±5.3146.85 ± 5.31 55.32±6.1855.32 ± 6.18 50.56±5.8750.56 ± 5.87 52.46±3.3052.46 ± 3.30 63.31±8.5263.31 ± 8.52 60.24±5.3760.24 ± 5.37 51.97±8.1551.97 ± 8.15 64.77±3.8764.77 ± 3.87 75.42±3.8775.42 ± 3.87 16.44%16.44\% COX2 50.27±7.9150.27 ± 7.91 50.05±2.0650.05 ± 2.06 49.86±7.4349.86 ± 7.43 50.27±0.3450.27 ± 0.34 53.36±8.8653.36 ± 8.86 52.01±3.1752.01 ± 3.17 53.56±3.4753.56 ± 3.47 59.99±9.7659.99 ± 9.76 77.78±1.8577.78 ± 1.85 29.65%29.65\% D 48.30±3.9848.30 ± 3.98 71.32±2.4171.32 ± 2.41 47.99±4.0947.99 ± 4.09 70.31±1.0970.31 ± 1.09 55.80±1.7755.80 ± 1.77 59.32±3.9259.32 ± 3.92 76.73±2.8376.73 ± 2.83 77.62±2.8877.62 ± 2.88 72.59±1.8472.59 ± 1.84 −- NCI1 49.90±1.1849.90 ± 1.18 50.58±1.3850.58 ± 1.38 50.63±1.2250.63 ± 1.22 50.74±1.7050.74 ± 1.70 50.10±0.8750.10 ± 0.87 49.88±0.5349.88 ± 0.53 41.42±2.1641.42 ± 2.16 45.96±2.4245.96 ± 2.42 59.07±0.4759.07 ± 0.47 16.41%16.41\% IMDB-B 50.75±3.1050.75 ± 3.10 50.80±3.1750.80 ± 3.17 54.08±5.1954.08 ± 5.19 50.20±0.4050.20 ± 0.40 56.50±3.5856.50 ± 3.58 56.50±4.9056.50 ± 4.90 45.34±3.7545.34 ± 3.75 65.46±4.3465.46 ± 4.34 68.96±0.0568.96 ± 0.05 5.07%5.07\% REDDIT-B 45.68±2.2445.68 ± 2.24 46.72±3.4246.72 ± 3.42 49.31±2.3349.31 ± 2.33 48.26±0.3248.26 ± 0.32 68.50±5.5668.50 ± 5.56 71.80±4.3871.80 ± 4.38 69.71±2.2169.71 ± 2.21 80.31±0.8580.31 ± 0.85 86.64±1.9786.64 ± 1.97 7.88%7.88\% HSE 57.02±8.4257.02 ± 8.42 56.87±10.5156.87 ± 10.51 62.72±10.1362.72 ± 10.13 53.02±5.1253.02 ± 5.12 53.56±3.9853.56 ± 3.98 51.18±2.7151.18 ± 2.71 58.49±2.6858.49 ± 2.68 63.05±0.9063.05 ± 0.90 64.68±0.7264.68 ± 0.72 2.58%2.58\% MMP 46.65±6.3146.65 ± 6.31 50.06±3.7350.06 ± 3.73 55.24±3.2655.24 ± 3.26 52.68±3.3452.68 ± 3.34 54.59±2.0154.59 ± 2.01 54.54±1.8654.54 ± 1.86 48.19±3.7448.19 ± 3.74 69.41±0.0469.41 ± 0.04 70.17±0.1170.17 ± 0.11 1.09%1.09\% p53 46.74±4.8846.74 ± 4.88 50.69±2.0250.69 ± 2.02 54.59±4.4654.59 ± 4.46 50.85±2.1650.85 ± 2.16 52.66±1.9552.66 ± 1.95 53.29±2.3253.29 ± 2.32 53.74±2.9853.74 ± 2.98 63.27±0.0463.27 ± 0.04 60.51±1.9560.51 ± 1.95 −- PPAR-gamma 53.94±6.9453.94 ± 6.94 45.51±2.5845.51 ± 2.58 57.91±6.1357.91 ± 6.13 49.60±0.2249.60 ± 0.22 51.40±2.5351.40 ± 2.53 50.30±1.5650.30 ± 1.56 56.20±1.5756.20 ± 1.57 68.23±1.5468.23 ± 1.54 72.59±0.0472.59 ± 0.04 6.39%6.39\% Avg. Rank 7.37.3 6.36.3 5.35.3 6.26.2 4.94.9 5.45.4 5.55.5 2.62.6 1.21.2 −- Datasets. We selected multiple datasets from diverse domains included in the widely-used UB-GOLD benchmark (Wang et al. 2025c). The distribution of OOD detection datasets comes from drug chemical formulas, protein structures, and etc. During testing, ID and OOD samples are mixed in a 1:11:1 ratio. The anomaly detection datasets span biological, chemical, and social domains. Instances belonging to the minority or ground-truth anomaly class are designated as anomalies, whereas others are treated as normal. Baselines. We compare SIGOOD with 1212 competitive baseline methods, detailed as follows: • Traditional OOD Detectors. These methods employ pre-trained encoders to extract representations, followed by classical OOD or anomaly detectors. Common encoders include the Weisfeiler-Lehman (WL) kernel (Shervashidze et al. 2011) and the propagation kernel (PK) (Neumann et al. 2016). The downstream OOD detectors include Local Outlier Factor (LOF) (Breunig et al. 2000), One-Class SVM (OCSVM) (Manevitz and Yousef 2001), Isolation Forest (iF) (Liu et al. 2008), and Mahalanobis Distance-based detector (MD) (Sehwag et al. 2021). • GNNs with Post-hoc Detectors. These methods generate graph representations using GNNs and perform detection in a post-hoc detectors. The GNN encoders include InfoGraph (Sun et al. 2020) and GraphCL (You et al. 2020), while the post-hoc detector typically involves Isolation Forest. • Test-time Training Methods. These methods perform OOD detection during test-time without training datasets. Notable methods include GTrans (Jin et al. 2023b) and GOODAT (Wang et al. 2024). Implementation Details. Following prior studies, we adopt the Area Under the ROC Curve (AUC) as the primary evaluation metric (Wang et al. 2024). All experiments are conducted on an NVIDIA RTX 4090 GPU with 24GB of memory. Each experiment is repeated five times to ensure stability. For baseline methods, we use the results reported in their original papers, such as GOODAT. Performance on OOD Detection Table 1 reports the AUC scores (%) for graph OOD detection across eight ID/OOD dataset pairs. Overall, SIGOOD consistently achieves the best performance, outperforming 1212 state-of-the-art (SOTA) baselines across all benchmarks. In particular, (1) Compared to traditional OOD detectors, SIGOOD demonstrates a clear advantage over traditional graph OOD methods such as PK-iF and WL-OCSVM. This illustrates the limitations of traditional shallow OOD detectors in capturing complex distribution shifts in graphs. (2) Compared to GNN-based baselines, SIGOOD exhibits steady improvements, highlighting the effectiveness of test-time enhancement. This indicates that refining the OOD signal at inference time enhances the model’s ability to capture the patterns of OOD graphs. (3) Compared with other test-time OOD detection methods, SIGOOD maintains robust performance on biochemical datasets. On Tox21/ToxCast and ClinTox/LIPO, it achieves AUCs of 69.97%69.97\% and 71.33%71.33\%, respectively, surpassing the previous best method GOODAT by 1.52%1.52\% and 14.20%14.20\%. These results confirm the effectiveness of SIGOOD, particularly in different domains graph OOD detection scenarios. Performance on Anomaly Detection Table 2 reports the AUC scores for anomaly detection on 1313 benchmark datasets with 88 baselines. Overall, SIGOOD consistently delivers the top performance, ranking first on 1111 out of 1313 datasets and achieving the best average rank (1.21.2) among 88 competing methods. Specifically, (1) SIGOOD significantly outperforms traditional baselines. These results indicate that SIGOOD is also well-suited for graph anomaly detection. (2) SIGOOD also surpasses GNN-based detectors. On BZR dataset, SIGOOD improves over the strongest neural baseline (e.g., GraphCL-iF) by 19.12%19.12\%. It is suggests that, beyond pretraining or contrastive learning, the energy-based self-improvment mechanism enhances the SIGOOD to distinguish subtle OOD signals that not captured by existing detectors. (3) Compared to SOTA test-time methods, SIGOOD shows notable gains, with improvements of 39.02%39.02\% on ENZYMES and 29.65%29.65\% on COX2, demonstrating the effectiveness of self-improvement mechanism across biochemical and social graphs. (4) While SIGOOD does not achieve the top score on D and p53, these deviations likely stem from dataset-specific variance rather than fundamental limitations. Overall, although SIGOOD fails to achieve the highest score on a few datasets, it achieves highly competitive average rankings across a variety of benchmarks. Ablation Study The Fig. 3 (a) visualizes the performance (AUC) of three model variants across different dataset pairs. As shown, SIGOOD consistently achieves the highest AUC on all three tasks (BZR/COX2, PTC-MR/MUTAG, and Esol/MUV), demonstrating the effectiveness of the full model. In contrast, removing the energy preference optimization loss (W/O LEPOL_EPO) leads to the most severe performance drop, particularly on PTC-MR/MUTAG and Esol/MUV, indicating that LEPOL_EPO is crucial for enhancing OOD sensitivity. Replacing the prompt generator with an optimizable parameter matrix (W/O PG) also results in a significant performance degradation. This suggests that the prompt generator plays a crucial role in amplifying ood signals. These results collectively confirm that both the energy-guided optimization and prompt-based modulation are essential to SIGOOD’s superior generalization capability. (a) Ablation Study on AUC. (b) Effect of PG Depth. Figure 3: Ablation Study and PG Depth Analysis. (a) Effect of Iteration Number. (b) Effect of the Parameter β. Figure 4: Parameter Sensitivity Analysis. Parameter Sensitivity Analysis Effect of Prompt Generator (PG) Depth. To evaluate the impact of the prompt generator depth on OOD detection performance, we experiment with 11-layer, 22-layer, and 33-layer MLPs on three datasets: BZR/COX2, PTC-MR/MUTAG, and Esol/MUV. As shown in Fig. 3 (b), using a 33-layer MLP significantly improves detection accuracy, aligning with common empirical practices in deep learning. In contrast, shallow configurations yield noticeably lower AUCs, especially on BZR/COX2 and PTC-MR/MUTAG. These results suggest that a deeper prompt generator (with-in 33-layers) is more effective in capturing complex semantic patterns and refining energy-based representations for reliable OOD detection. The effects of Self-improving Interations. To intuitively demonstrate the impact of self-improvement iterations on the OOD detection performance, we conducted experiments on the Esol/MUV dataset with large-scale and high-density features. As shown in Fig. 4 (a), increasing the number of iterations from 10001000 to 1000010000 consistently improves the performance of SIGOOD, with the AUC reaching a peak of 87.72%87.72\%. This trend demonstrates the effectiveness of iterative on self-learning. However, when the iterations exceed 1000010000, the performance gain diminishes slightly, indicating that SIGOOD has approached convergence. Therefore, a moderate iterations offers a favorable trade-off between performance and computational efficiency. The effects of parameter β. In SIGOOD, β serves as a weighting factor in the energy preference optimization (EPO) objective, controlling the trade-off between preserving original graph semantics and encouraging energy-based differentiation between in-distribution and OOD samples. We investigate the sensitivity of our model to the hyperparameter β on the BZR/COX2 dataset. As shown in Fig. 4 (b), the highest value of 87.36%87.36\% is observed at β=80β=80. This indicates that moderate values of β yield slightly better performance. A small β may under-emphasize the energy gap, resulting in weaker OOD separation, while a large β may distort the embedding space. Figure 5: Visualization of Graph Distribution. Visualization To further illustrate the effectiveness of our method in distinguishing OOD samples, we visualize the normalized entropy distribution of the BZR/COX2 dataset in Fig. 5. The figure clearly shows a separation between ID and OOD samples, with OOD instances exhibiting higher entropy values. The kernel density estimation (KDE) curves further reveal that the OOD distribution is skewed towards the high-entropy region, while the ID distribution remains concentrated in the low-entropy area. The distinct distributional shift highlights the reliability of entropy as a scoring metric for OOD detection in our framework. Conclusion In this paper, we propose SIGOOD, a self-improving framework for test-time graph OOD detection, which leverages energy-based feedback to iteratively refine OOD detection. By integrating a lightweight prompt generator with a well-trained GNN, SIGOOD constructs PE graphs that amplify potential OOD signals. Through energy variations between both the PE graph and the original test graph, the model identifies and enhances OOD-relevant signals. SIGOOD also proposes a novel energy preference optimization loss to guide prompt updates, enabling the framework to form a closed-loop self-improving process without requiring additional labels at test time. Extensive experiments demonstrate that SIGOOD achieves superior OOD detection performance across various graph datasets, highlighting its effectiveness in real-world scenarios. Acknowledgments This work is supported by National Natural Science Foundation of China (62502065, 62176036, 62302333); the Beatriu de Pinós del Departament de Recerca i Universitats de la Generalitat de Catalunya (2022 BP-00256); the predoctoral program AGAUR-FI ajuts (2025 FI-200470) Joan Oró, which is backed by the Secretariat of Universities and Research of the Department of Research and Universities of the Generalitat of Catalonia, as well as the European Social Plus Fund. Appendix Derivation of the Energy-based Reward Function We begin with the general form of the optimization objective: maxE^(v|Pm)[r(Pm,v)]−β⋅DKL(E^(v|Gp)||E^(u|Gt)), _ E(v|P_m)\;E [r(P_m,v) ]-β· D_KL ( E(v|G_p)|| E(u|G_t) ), (10) where E^(v;Gp) E(v;G_p) and E^(u;Gt) E(u;G_t) denote the energy score from the prompt-enhanced and original test graphs, respectively. Following the standard expectation formulation x=∑xp(x)f(x)Ex= _xp(x)f(x), we treat E^(v;Gp) E(v;G_p) as the sampling distribution over nodes, and interpret the KL divergence term as an energy-weighted log-ratio: DKL(E^(v;Gp)||E^(u;Gt))=∑vE^(v;Gp)logE^(v;Gp)E^(u;Gt). D_KL ( E(v;G_p)|| E(u;G_t) )= _v E(v;G_p) log E(v;G_p) E(u;G_t). (11) Substituting this into the original objective yields: max∑vE^(v;Gp)⋅[r(Pm,v)−βlogE^(v;Gp)E^(u;Gt)]. _v E(v;G_p)· [r(P_m,v)-β log E(v;G_p) E(u;G_t) ]. (12) To maximize the objective, the reward for each outcome v should satisfy: r(Pm,v)=β⋅logE^(v;Gp)E^(u;Gt).r(P_m,v)=β· log E(v;G_p) E(u;G_t). (13) The above provides the derivation of Eq.(8), which assigns higher rewards to samples exhibiting increased energy after prompt enhancement, indicating a stronger likelihood of being OOD. References X. Bi, J. Lu, B. Liu, X. Cun, Y. Zhang, W. Li, and B. Xiao (2025) Customttt: motion and appearance customized video generation via test-time training. In AAAI 2025, p. 1871–1879. Cited by: Test-time Training. R. A. Bradley and M. E. Terry (1952) Rank analysis of incomplete block designs: i. the method of paired comparisons. Biometrika 39 (3/4), p. 324–345. Cited by: Step 4: Feedback Optimization. M. M. Breunig, H. Kriegel, R. T. Ng, and J. Sander (2000) LOF: identifying density-based local outliers. In SIGMOD 2000, p. 93–104. Cited by: 1st item. Y. Cao, F. Shi, Q. Yu, X. Lin, C. Zhou, L. Zou, P. Zhang, Z. Li, and D. Yin (2025) IBPL: information bottleneck-based prompt learning for graph out-of-distribution detection. Neural Networks 188, p. 107381. Cited by: Graph OOD Detection. S. Chen, L. Huang, J. R. Schwarz, Y. Du, and Y. Wei (2023) Secure out-of-distribution task generalization with energy-based models. In NeurIPS 2023, p. 67007–67020. Cited by: Overview. K. Dalal, D. Koceja, J. Xu, Y. Zhao, S. Han, K. C. Cheung, J. Kautz, Y. Choi, Y. Sun, and X. Wang (2025) One-minute video generation with test-time training. In CVPR 2025, p. 17702–17711. Cited by: Test-time Training. Z. Ding, J. Shi, S. Shen, X. Shang, J. Cao, Z. Wang, and Z. Gong (2024) Sgood: substructure-enhanced graph-level out-of-distribution detection. In CIKM 2024, p. 467–476. Cited by: Introduction. L. Fu, B. Deng, S. Huang, T. Liao, C. Zhang, and C. Chen (2025) Learn from global rather than local: consistent context-aware representation learning for multi-view graph clustering. In IJCAI 2025, p. 5145–5153. Cited by: Introduction. D. Fuchsgruber, T. Wollschläger, and S. Günnemann (2024) Energy-based epistemic uncertainty for graph neural networks. In NeurIPS 2024, p. 34378–34428. Cited by: Graph OOD Detection. Y. Hou, H. Zhu, R. Liu, Y. Su, J. Xia, J. Wu, and K. Xu (2025) Structural entropy guided unsupervised graph out-of-distribution detection. In AAAI 2025, p. 17258–17266. Cited by: Introduction. Z. Jiang, J. Lu, H. Fan, T. Wang, and J. Yan (2025) Learning structured universe graph with outlier ood detection for partial matching. In ICLR 2025, p. 1–16. Cited by: Step 3: OOD Signal Recognition. D. Jin, L. Wang, Y. Zheng, G. Song, F. Jiang, X. Li, W. Lin, and S. Pan (2023a) Dual intent enhanced graph neural network for session-based new item recommendation. In W 2023, p. 684–693. Cited by: Introduction. W. Jin, T. Zhao, J. Ding, Y. Liu, J. Tang, and N. Shah (2023b) Empowering graph representation learning with test-time graph transformation. In ICLR 2023, p. 1–27. Cited by: Introduction, 3rd item. S. Kullback and R. A. Leibler (1951) On information and sufficiency. The Annals of Mathematical Statistics 22 (1), p. 79–86. Cited by: Step 4: Feedback Optimization. J. Liang, R. He, and T. Tan (2025) A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision 133 (1), p. 31–64. Cited by: Test-time Training. X. Lin, Y. Cao, N. Sun, L. Zou, C. Zhou, P. Zhang, S. Zhang, G. Zhang, and J. Wu (2025) Conformal graph-level out-of-distribution detection with adaptive data augmentation. In W 2025, p. 4755–4765. Cited by: Introduction. X. Lin, W. Zhang, F. Shi, C. Zhou, L. Zou, X. Zhao, D. Yin, S. Pan, and Y. Cao (2024) Graph neural stochastic diffusion for estimating uncertainty in node classification. In ICML 2024, p. 30457–30478. Cited by: Graph OOD Detection. F. T. Liu, K. M. Ting, and Z. Zhou (2008) Isolation forest. In ICDM 2008, p. 413–422. Cited by: 1st item. W. Liu, X. Wang, J. Owens, and Y. Li (2020) Energy-based out-of-distribution detection. In NeurIPS 2020, p. 21464–21475. Cited by: Overview. L. M. Manevitz and M. Yousef (2001) One-class svms for document classification. Journal of Machine Learning Research 2, p. 139–154. Cited by: 1st item. M. Neumann, R. Garnett, C. Bauckhage, and K. Kersting (2016) Propagation kernels: efficient graph kernels from propagated information. Machine Learning 102 (2), p. 209–245. Cited by: 1st item. J. Pan, Y. Liu, X. Zheng, Y. Zheng, A. W. Liew, F. Li, and S. Pan (2025) A label-free heterophily-guided approach for unsupervised graph fraud detection. In AAAI 2025, p. 12443–12451. Cited by: Introduction. R. Rafailov, A. Sharma, E. Mitchell, C. D. Manning, S. Ermon, and C. Finn (2023) Direct preference optimization: your language model is secretly a reward model. In NeurIPS 2023, Vol. 36, p. 53728–53741. Cited by: Step 4: Feedback Optimization. D. E. Rumelhart, G. E. Hinton, and R. J. Williams (1986) Learning representations by back-propagating errors. Nature 323 (6088), p. 533–536. Cited by: Step 1: Graph Prompt Generator. V. Sehwag, M. Chiang, and P. Mittal (2021) SSD: A unified framework for self-supervised outlier detection. In ICLR 2021, p. 1–17. Cited by: 1st item. X. Shen, Y. Wang, K. Zhou, S. Pan, and X. Wang (2024) Optimizing ood detection in molecular graphs: a novel approach with diffusion models. In SIGKDD 2024, p. 2640–2650. Cited by: Introduction. N. Shervashidze, P. Schweitzer, E. J. van Leeuwen, K. Mehlhorn, and K. M. Borgwardt (2011) Weisfeiler-lehman graph kernels. Journal of Machine Learning Research 12, p. 2539–2561. Cited by: 1st item. F. Sun, J. Hoffmann, V. Verma, and J. Tang (2020) InfoGraph: unsupervised and semi-supervised graph-level representation learning via mutual information maximization. In ICLR 2020, p. 1–16. Cited by: 2nd item. L. Wang, D. He, H. Zhang, Y. Liu, W. Wang, S. Pan, D. Jin, and T. Chua (2024) Goodat: towards test-time graph out-of-distribution detection. In AAAI 2024, p. 15537–15545. Cited by: Figure 1, Introduction, 3rd item, Implementation Details.. X. Wang, Y. Dong, D. Jin, Y. Li, L. Wang, and J. Dang (2023) Augmenting affective dependency graph via iterative incongruity graph learning for sarcasm detection. In AAAI 2023, p. 4702–4710. Cited by: Introduction. X. Wang, Y. Wang, D. He, Z. Yu, Y. Li, L. Wang, J. Dang, and D. Jin (2025a) Elevating knowledge-enhanced entity and relationship understanding for sarcasm detection. IEEE Transactions on Knowledge and Data Engineering 37 (6), p. 3356–3371. Cited by: Introduction. Y. Wang, Y. Liu, N. Liu, R. Miao, Y. Wang, and X. Wang (2025b) AdaGCL+: an adaptive subgraph contrastive learning towards tackling topological bias. IEEE Transactions on Pattern Analysis and Machine Intelligence, p. 8073–8087. Cited by: Introduction. Y. Wang, Y. Liu, X. Shen, C. Li, R. Miao, K. Ding, Y. Wang, S. Pan, and X. Wang (2025c) Unifying unsupervised graph-level anomaly detection and out-of-distribution detection: A benchmark. In ICLR 2025, p. 1–27. Cited by: Introduction, Datasets.. Q. Wu, Y. Chen, C. Yang, and J. Yan (2023) Energy-based out-of-distribution detection for graph neural networks. In ICLR 2023, Cited by: Introduction. Q. Wu, F. Nie, C. Yang, T. Bao, and J. Yan (2024) Graph out-of-distribution generalization via causal intervention. In W 2024, p. 850–860. Cited by: Introduction. Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang, and P. S. Yu (2020) A comprehensive survey on graph neural networks. IEEE Transactions on Neural Networks and Learning Systems 32 (1), p. 4–24. Cited by: Overview, Step 3: OOD Signal Recognition. Y. You, T. Chen, Y. Sui, T. Chen, Z. Wang, and Y. Shen (2020) Graph contrastive learning with augmentations. In NeurIPS 2020, p. 5812–5823. Cited by: 2nd item. J. Yu, J. Liang, and R. He (2023) Mind the label shift of augmentation-based graph ood generalization. In CVPR 2023, p. 11620–11630. Cited by: Step 2: Prompt Injection. H. Zhang, B. Wu, X. Yang, X. Yuan, X. Liu, and X. Yi (2025a) Dynamic graph unlearning: a general and efficient post-processing method via gradient transformation. In W 2025, p. 931–944. Cited by: Introduction. H. Zhang, B. Wu, X. Yuan, S. Pan, H. Tong, and J. Pei (2024a) Trustworthy graph neural networks: aspects, methods, and trends. Proceedings of the IEEE 112 (2), p. 97–139. Cited by: Introduction. J. Zhang, Y. Wang, X. Yang, S. Wang, Y. Feng, Y. Shi, R. Ren, E. Zhu, and X. Liu (2024b) Test-time training on graphs with large language models (llms). In ACM M 2024, p. 2089–2098. Cited by: Test-time Training. J. Zhang, Y. Wang, X. Yang, and E. Zhu (2024c) A fully test-time training framework for semi-supervised node classification on out-of-distribution graphs. ACM Transactions on Knowledge Discovery from Data 18 (7), p. 1–19. Cited by: Graph OOD Detection. Q. Zhang, Z. Shi, S. Pan, J. Chen, H. Wu, and X. Chen (2024d) EGonc: energy-based open-set node classification with substitute unknowns. In NeurIPS 2024, Vol. 37, p. 66147–66177. Cited by: Introduction. S. Zhang, C. Zhou, Y. Liu, P. Zhang, X. Lin, and S. Pan (2025b) Conformal anomaly detection in event sequences. In ICML 2025, p. 1–17. Cited by: Graph OOD Detection. X. Zheng, D. Song, Q. Wen, B. Du, and S. Pan (2024) Online GNN evaluation under test-time graph distribution shifts. In ICLR 2024, p. 1–22. Cited by: Test-time Training.