acceptodds
Under review as a conference paper at ICLR 2027

AnTAGAD: Generalist Text-Attributed Graph Anomaly Detection with Textual and Topological Anchors

Abstract

Representing the rich textual semantics of text-attributed graphs (TAGs) with large language models markedly improves cross-domain generalization, enabling effective anomaly detection across domains as varied as social platforms, financial transaction systems, and multi-agent collaboration environments. Concretely, this generalizability allows detectors to recall a large fraction of anomalies across domains, as reflected in high AUROC. However, this apparent success is undermined by persistently low AUPRC, meaning many flagged nodes are false positives and individual predictions carry unreliable confidence, leading to unacceptable misjudgments in practical deployment. We trace this failure to the widespread reliance on generalizable prototype representations, which enhance cross-domain adaptation at the expense of instance-specific detail and thus blur the distinctions among individual anomalies. We term this the Over-Generalization () problem. To address this, we propose a novel multi-anchor representation scheme to replace prototypes, encoding each node against a set of high-confidence anomaly and normal anchors to improve generalization while preserving the instance-specific characteristics of individual anomalies. Extensive experiments on eight real-world cross-domain datasets show that our method consistently outperforms all competitive baselines, with substantial gains in both AUPRC and AUROC. Beyond these empirical gains, we provide a theoretical characterization of sufficient conditions for controlling together with matched-coverage empirical analyses showing reduced anomaly acceptance on most evaluated settings.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Text-attributed Graph Anomaly Detection, Generalist Graph Learning, Large Language Models.

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_muUpOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+2)
Read review
Paper summary

AnTAGAD addresses generalist anomaly detection on text-attributed graphs using LLM-selected textual anchors and neighborhood-consistency-based topological anchors. It scores unseen graphs through relative top-k anchor similarity and neighborhood dispersion, and diagnoses over-generalization through anomaly acceptance at matched normal coverage. Experiments on eight target graphs report leading AUPRC on seven targets and leading mean AUROC on all eight, alongside ablations, controlled comparisons, and a conditional separation guarantee.

Critical strengths
  1. Strong reported transfer performance: AUPRC improves over the strongest baseline by 22.14 percentage points on Pubmed and 15.86 on Texas (Table 1), supporting the practical promise of source-only anchor learning.
  2. The matched-coverage analysis and frozen-encoder controls (Tables 4–6) provide useful evidence beyond qualitative visualization. The paper also clearly limits the scope of its conditional guarantee.
Critical weaknesses

Update after author response: The authors have substantially addressed my main concerns by correcting Eq. (15), clarifying score normalization and its connection to Theorem 2, and providing capacity-matched ablations and reference-budget details. These clarifications and additional experiments strengthen the support for the method and its reported gains. I am therefore raising my rating from weak rejection to weak acceptance.

Original concerns before the author response:

  1. The stated method and released implementation differ materially. Equation (15) is not the contrastive probability implemented in code; loss weighting and separately normalized score components also differ from Equations (16)/(18). This leaves the table-producing detector and its correspondence to Theorem 2 unclear, affecting reproducibility and ablation interpretation.
  2. The anchor contribution remains confounded. Removing TeA/ToA roughly halves bank capacity, while reference counts in Tables 4/6 are unspecified. Appendix J mixes original and injected anomalies without separate results; several injected text corruptions align with the LLM selection criteria. These gaps limit attribution of the gains to instance anchoring.
Questions for authors
  1. Which code revision/configuration produced the tables? Is Equation (15) a typo, and how do the implemented losses and normalized score correspond to Equations (16)/(18) and Theorem 2?
  2. Are Tables 4/6 matched in candidate information and per-class storage? Please specify reference counts and clarify whether single-branch gains persist at matched capacity.
  3. What are the proportions and separate results for original versus injected anomaly types? Please provide the confidence prompts and data/score generation resources needed to verify this benchmark.
Reviewer_FAw3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

AnTAGAD uses LLM-derived textual anchors and node-neighborhood consistency to learn transferable normal and anomalous references. At inference, it combines local anchor similarities with a neighborhood-dispersion term to detect anomalies in unseen graphs.

Critical strengths
  • The reported AUPRC improvements on several target graphs are substantial, suggesting the combined approach is effective for cross-domain anomaly detection.
  • The matched-coverage analysis offers a useful view of missed anomalies, and the theoretical discussion states its assumptions clearly.
Critical weaknesses
  • The constructed anomalies introduce textual and structural inconsistencies that appear closely aligned with the signals the method is designed to detect. Without separate results for pre-existing annotations and constructed anomaly types, it is unclear how well the method generalizes to naturally occurring anomalies. In addition, the LLM-derived confidence scores provide extra information whose contribution does not appear to be isolated from the anchor design itself.
  • The ablations show that removing the neighborhood-dispersion term increases anomaly acceptance, but a dispersion-only baseline is missing. Removing either anchoring branch also changes the size of the reference banks. It is therefore unclear how much of the gain comes from anchor complementarity, reference capacity, or the neighborhood score. Also, the joint anchor loss in Eq. 15 appears to differ in form from the textual and topological losses in Eqs. 11 and 13. Please clarify whether this is intentional or a typesetting issue.
Questions for authors
  • How accurate are the anchor pseudo-labels across anomaly types, particularly structural anomalies whose text appears normal?
3 comments

We are grateful for the reviewers' constructive feedback. Our responses to Q1, Q2 and Q3 are as follows.

Q1: Generalization to Naturally Occurring Anomalies

A1: We thank the reviewer for raising this important concern regarding the generalization of AnTAGAD to naturally occurring anomalies. We agree that evaluating the model on realistic anomalies is essential for assessing its practical applicability. However, publicly available text-attributed graph (TAG) datasets with naturally occurring, ground-truth anomaly annotations remain limited. To provide a more comprehensive evaluation, we further conduct experiments on representative TAG anomaly detection benchmarks introduced by CMUCL. These benchmarks provide a complementary evaluation setting for assessing the generalization ability of AnTAGAD beyond our original experimental datasets. The experimental results are presented below.

Table 1: Performance Comparison (AUROC/AUPRC)

MethodCiteSeerComputersPhotoPubMed
ARC(2024)72.37/52.0459.23/5.0258.59/4.8574.57/15.96
GCTAM(2025)47.16/5.5647.38/3.6748.72/3.8051.54/4.94
UNPrompt(2025)55.28/4.6046.93/3.5945.62/3.5050.49/3.88
IA-GGAD(2025)73.57/50.7059.66/5.1359.03/4.9574.64/16.82
CMUCL(2025)77.76/14.7856.79/4.6156.25/4.5357.75/4.64
OWLEYE(2026)72.64/52.6853.92/4.3353.62/4.3072.18/14.37
ProMoS(2026)60.71/36.4351.80/4.1751.31/4.0866.99/7.69
AnTAGAD75.90/54.9173.58/24.9771.79/17.3977.28/34.54

Q2: Contribution of LLM-Derived Confidence

A2: We sincerely thank the reviewer for raising this insightful question. To further investigate the contribution of instance-level anchor design independently of LLM-derived confidence information, we compare directly using LLM-selected anchors (Anchors) with aggregating the same anchors into K-means prototypes (K-means), while keeping the LLM-derived information consistent. As shown in Table 2, Anchors outperforms K-means in AUROC across all eight datasets and achieves higher AUPRC on seven datasets. These results demonstrate that the performance improvements are not solely attributable to LLM-derived information, but also benefit from preserving instance-level anchor representations.

Table 2: Performance Comparison Between K-means Prototypes and Instance-level Anchors (AUROC/AUPRC).

MethodCiteSeerPubMedGroceryMoviesToysFitnessCornellTexas
K-means47.50±6.59 / 7.42±1.1974.08±1.56 / 13.15±1.2062.09±1.91 / 8.03±0.5262.17±1.41 / 7.75±0.2665.67±1.92 / 9.69±0.6169.97±2.43 / 20.59±6.5850.60±4.10 / 11.30±1.5153.33±4.77 / 14.31±2.50
Anchors52.36±6.11 / 8.86±2.0177.67±0.48 / 15.01±0.3664.53±0.94 / 8.33±0.1966.79±0.20 / 8.79±0.0568.61±0.89 / 10.06±0.1873.34±0.96 / 17.17±2.2156.13±3.21 / 14.50±1.4357.24±3.08 / 21.04±2.09

Q3: Contribution of Neighborhood Dispersion

A3: We thank the reviewer for highlighting the need to clarify the contribution of neighborhood dispersion. To address this concern, we conduct additional experiments comparing anchor-only, dispersion-only, normal-bank-only, anomaly-bank-only, and the full scoring model. As shown in Table 3, the full model achieves the best average AUROC, AUPRC, and OG@95 across eight datasets, demonstrating that neighborhood dispersion complements anchor-based scoring to improve anomaly detection and reduce anomaly acceptance.

Table 3: Average Performance of Different Components (AUROC/AUPRC/OG@95)

ComponentAverage AUROCAverage AUPRCAverage OG@95
Anchor64.8124.3785.43
Dispersion70.2525.7273.10
Normal-bank56.9215.7784.56
Anomaly-bank56.2311.6788.71
Full77.7641.0659.87

We sincerely thank the reviewer for their valuable comments and suggestions. Our point-by-point responses to Q4, Q5 and Q6 are provided below.

Q4: Impact of Anchor Bank Capacity on Ablation Results

A4: We thank the reviewer for raising this concern. To distinguish the effects of anchor complementarity from anchor bank capacity, we conduct additional capacity-matched ablation experiments. Specifically, TeA and ToA retain only textual or topological anchoring, respectively, while all variants use the same number of normal (1,900) and anomalous (1,854) anchors, with all other settings unchanged. As shown in Table 4, AnTAGAD achieves the best AUROC, AUPRC, and OG@95 across all eight datasets, demonstrating that the improvements primarily arise from the complementary information provided by textual and topological anchors rather than increased anchor capacity.

Table 4: Performance Comparison of Capacity-Matched Ablation Variants (AUROC/AUPRC/OG@95)

MethodCiteSeerPubMedGroceryMoviesToysFitnessCornellTexas
TeA67.56/14.02/79.7975.83/20.07/88.8272.35/12.95/90.7671.79/10.94/96.2674.95/13.60/83.0079.22/25.28/77.4975.85/28.05/69.5277.58/30.93/75.24
ToA65.10/16.78/65.6376.69/13.31/99.9766.32/8.51/96.5066.98/8.45/98.0669.88/9.91/90.2974.57/13.44/86.0752.20/12.03/94.2954.28/14.18/91.43
AnTAGAD73.11/53.07/47.4077.81/51.84/54.5174.21/33.64/63.8574.51/30.37/62.2276.95/32.24/61.2380.90/30.19/60.2081.31/49.82/64.7683.28/47.30/64.76

Q5: Inconsistency between the Joint Anchor Loss and the Textual and Topological Losses.

A5: We sincerely thank the reviewer for pointing out this inconsistency. We clarify that the difference in formulation between Eq. (15) and Eqs. (11) and (13) was not intentional, but resulted from an error in the original manuscript. We sincerely apologize for this oversight. After carefully examining our implementation, we have corrected Eq. (15) in the revised manuscript to accurately reflect the implemented contrastive loss. The corrected equation is presented below:

LA=−∑i∈A+∣N(i)∣>0log∑j∈N(i)exp(sim(hi,hj)/τ)∑j∈N(i)exp(sim(hi,hj)/τ)+∑k∈A−exp(sim(hi,hk)/τ).\mathcal{L}_\mathcal{A}=-\sum_{\substack{i\in{\mathcal A}^{+}\\\\|\mathcal N(i)|>0}} \mathrm{log} \frac{\sum_{j \in \mathcal{N}(i)} \mathrm{exp}(\mathtt{sim}(\mathbf{h}_i,\mathbf{h}_j)/\tau)}{\sum_{j \in \mathcal{N}(i)}\mathrm{exp}(\mathtt{sim}(\mathbf{h}_i,\mathbf{h}_j)/\tau)+\sum_{k \in \mathcal{A}^-}\mathrm{exp}(\mathtt{sim}(\mathbf{h}_i,\mathbf{h}_k)/\tau)}.

Q6: Accuracy of Anchor Pseudo-Labels.

A6: We greatly appreciate the reviewer for raising this valuable and thought-provoking question. The selected normal anchors achieve a purity of approximately 95%–98%, while anomaly-anchor purity is relatively lower (20%–30%) due to the low prevalence of anomalous nodes. However, our method does not directly use individual anchor pseudo-labels for anomaly prediction. Instead, it considers the relative similarities between each node and multiple normal and anomalous anchors. Further representation-space analysis shows that normal-anchor regions contain fewer anomalies than random regions (2.32±0.11% vs. 5.06±0.23%), whereas anomaly-anchor regions exhibit a higher anomaly ratio (8.21±0.25% vs. 2.94±0.16%). These results indicate that, despite imperfect node-level pseudo-labels for some anomaly anchors, the anchor banks still provide informative references with distinct local distributions in the representation space.

Table 5: Within-Region Anomaly Ratio (%)

TypeAnchor_regionRandom_region
Normal2.32±0.115.06±0.23
Anomaly8.21±0.252.94±0.16
Reviewer_FAw3OpenReview ↗

Official response from reviewer

Thank you for additional experiments. I acknowledge that the authors' rebuttal and additional experiments address several of my technical concerns, so I have increased my soundness score. However, my overall assessment remains as before, as my primary concerns regarding the limited methodological novelty and reliance on constructed anomalies remain insufficiently addressed. I’d like to see clearer evidence that the approach generalizes to naturally occurring anomalies. I’m also still unsure what benefit the anchor-based mechanism provides.

Reviewer_GwXAOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies generalist anomaly detection on text-attributed graphs and identifies a failure mode termed Over-Generalization, where detectors built on aggregated prototype representations achieve high AUROC but low AUPRC because coarse aggregation collapses instance-level detail and admits anomalies into the normal acceptance region. The authors formalize this as the false-negative rate at a prescribed normal-node coverage, derive sufficient conditions for controlling it, and define a contamination measure. They then propose AnTAGAD, which replaces prototypes with instance-level textual anchors selected through an Agent-Skill-Confidence LLM reliability score and topological anchors chosen by neighborhood consistency, merged into normal and anomaly banks. Scoring uses the gap between top-k similarities to the two banks plus neighborhood dispersion. Eight cross-domain datasets show consistent gains over strong generalist baselines.

Critical strengths

The paper identifies and formally defines Over-Generalization in TAG anomaly detection, a concrete and practically important failure mode. It supports the definition with a theoretical characterization of sufficient conditions and matched-coverage empirical evidence, giving high novelty and rigor.

Figures and text are tightly staged, with Figure 1 motivating the problem, Figure 2 overviewing the pipeline, and Figure 3 detailing the agent module. Notation and preliminaries are cleanly introduced, making a complex design easy to follow.

Eight real-world cross-domain datasets, strong generalist baselines including multi-prototype controls, and gains in both AUROC and AUPRC make the validation solid. Reporting at matched normal coverage directly targets the claimed failure mode.

Critical weaknesses

Eq. (7) defines selection through TopK and BottomK without specifying tie handling, per-graph budget normalization, or the relation between selection budget K and retrieval size k. Please give a general, dataset-agnostic formulation and discuss the degenerate case where K exceeds available candidates.

AnTAGAD is not the best on Fitness. Please diagnose why by quantifying domain shift, anchor coverage, class separation, and textual signal quality, and report OG(q) and contamination curves. Clarifying whether anomalies there are semantic or structural would test the claim's universality.

Figure 2 lacks an informative caption. Please add a step-by-step caption naming each module, including textual anchoring, topological anchoring, bank merging, top-k scoring, and the dispersion term, plus the data flow, so readers can parse the architecture without searching the text.

Questions for authors
  1. AnTAGAD performs well on most datasets, but its AUPRC is slightly lower on Fitness. Could you briefly explain why this dataset is more challenging?

  2. Could you explain a little more about how the textual and topological anchors complement each other? A simple example would help readers understand their different roles.

  3. Figure 2 provides a useful overview, but it contains many details. Could you expand the caption to briefly explain the main steps and how the two types of anchors are combined?

3 comments

We sincerely appreciate the reviewers' valuable suggestions, which have provided important guidance for improving our work. Our responses to Q1 are provided below.

Q1: Further clarification of the anchor selection and retrieval mechanisms in Eq. (7) is needed.

A1: We thank the reviewer for highlighting the need to clarify the anchor selection procedure in Eq. (7). We apologize for the insufficient description and the inaccurate statement regarding tie handling. We clarify the selection rules, the relationship between K and k, and the degenerate cases as follows.

(1) Anchor Selection Budget and General Formulation: Anchor selection is performed independently for each source graph, each branch (textual or topological), and each pseudo-class (normal or anomalous). Rather than using a graph-size-dependent proportion, we adopt a fixed selection budget (K=160). For each source graph, let C\mathcal{C} denote the set of available candidate nodes, the actual number of selected anchors is defined as Keff=min⁡(K,∣C∣).K_{\mathrm{eff}}=\min(K,|\mathcal{C}|). The KeffK_{\mathrm{eff}} candidates with the highest and lowest scores are selected as normal and anomalous anchors, respectively. This formulation applies to different source graphs without requiring dataset-specific adjustments.

(2) Tie Handling: Anchor selection is implemented using PyTorch's torch. topk. When candidates have identical scores at the selection boundary, ties are handled by the underlying operator without an explicitly defined secondary ordering. We acknowledge that the original description of fixed-order tie handling was inaccurate and have corrected it in the revised manuscript.

(3) Relationship Between Selection Budget K and Retrieval Size k: The parameter K controls the number of anchors selected from each source graph, branch, and pseudo-class, whereas k determines the number of most similar anchors used for anomaly scoring during inference. We use K=160 and k=20 by default. After merging textual and topological selections and removing duplicate node identities, the resulting anchor banks are used for retrieval. If a bank contains fewer than k anchors, all available anchors are retrieved, and their similarities are averaged over the actual retrieval count.

(4) Degenerate Cases: When the number of available candidates is smaller than K, all candidates are selected without padding or replacement. Since normal and anomalous anchors are selected independently, their selections may overlap when K is large relative to the candidate set.

We have supplemented the explanations of Eq. (7) in the revised manuscript.

We are grateful for the reviewers' constructive feedback. Our responses to Q2 are as follows.

Q2: Why does AnTAGAD show relatively lower performance on the Fitness dataset?

A2: We sincerely thank the reviewer for raising this insightful and important question. We agree that a more in-depth analysis of the Fitness dataset is essential for understanding the factors contributing to the relatively lower performance of AnTAGAD on this dataset. To better investigate the underlying reasons and provide a more comprehensive explanation, we conduct further analyses from the following perspectives:

(1) Domain Shift: To investigate whether domain shift is one of the factors contributing to the relatively lower performance of AnTAGAD on Fitness, we employ Maximum Mean Discrepancy (MMD²) to quantify the distribution discrepancy between the source and target domains. As shown in Table 1, Fitness exhibits a more pronounced distribution discrepancy compared with the other three e-commerce datasets. This discrepancy may be one of the factors contributing to the relatively lower performance of AnTAGAD on Fitness.

Table 1: Domain Shift Analysis Based on MMD².

DatasetMMD²
Grocery0.007639
Movies0.005055
Toys0.005177
Fitness0.013974

(2) Anchor Coverage: As shown in Table 2, Toys exhibits the lowest normal and anomaly anchor coverage rates among the four e-commerce datasets, while Fitness achieves slightly higher coverage rates of 81.25% and 91.07%, respectively. However, Fitness still shows lower anchor coverage than Grocery and Movies in both categories, suggesting that relatively limited anchor coverage may be one of the factors contributing to the weaker performance of AnTAGAD on Fitness.

Table 2: Anchor Coverage Comparison Across E-commerce Datasets (%).

DatasetNormal Coverage (%)Anomaly Coverage (%)
Grocery84.6892.60
Movies82.5891.80
Toys77.2588.39
Fitness81.2591.07

(3) Class Separation: To examine AnTAGAD's weaker performance on Fitness, we evaluate its ability to detect textual and topological anomalies separately and jointly, while examining anchor-based scoring and neighborhood dispersion. As shown in Table 3, the anomaly types differ in detection difficulty. Using anchor-based scoring alone, textual anomalies achieve higher AUROC and AUPRC than topological anomalies (84.75%/16.45% vs. 62.35%/4.47%). However, incorporating neighborhood dispersion improves topological anomaly detection. Under the full model, joint detection achieves the highest AUPRC (30.43%), while topological anomaly detection achieves the highest AUROC (87.06%). These findings suggest that differences in anomaly discriminability may contribute to AnTAGAD's weaker performance on Fitness.

Table 3: Class Separation Analysis on Fitness (%).

ComponentTextual AnomalyTopological AnomalyAUROCAUPRC
Anchor Only✓✓73.5515.54
Anchor Only✓84.7516.45
Anchor Only✓62.354.47
Dispersion_V✓✓69.5013.67
Dispersion_V✓53.623.12
Dispersion_V✓85.3714.89
Full✓✓80.8130.43
Full✓74.5720.65
Full✓87.0618.90

(4) Textual Signal Quality: We analyzed the textual signal quality of Fitness and found substantial template redundancy. As shown in Table 3, 98.14% of the node texts share the same fixed prefix, which accounts for approximately 50.48% of the average text length. After removing this boilerplate, the effective item title contains only 10.79 tokens on average. These results indicate that Fitness provides relatively short and highly templated textual signals,

Table 4: Textual Signal Quality of Fitness

MetricFitness
Prefix Coverage98.14%
Boilerplate Token50.48%
Effective Length10.79

(5) OG and contamination curves: We thank the reviewer for suggesting an analysis of the OG(q) and contamination curves. Originally, we only reported the OG curve for Fitness under high-coverage settings. To address this concern, we have extended the analysis to include both OG(q) and contamination curves for Fitness across different coverage levels. Since OpenReview does not support embedding figures in responses directly, we have added these curves and the corresponding analysis to Appendix D.6 of the revised manuscript for the reviewer's reference.

Above all, our findings suggest that Fitness exhibits a larger domain shift, relatively lower anchor coverage, varying detection difficulties across anomaly types, and highly redundant textual information. These factors may jointly contribute to the relatively lower performance of AnTAGAD on Fitness. We have added a detailed analysis of these factors, together with the OG(q) and contamination curves, to Appendix D.6 of the revised manuscript.

We thank the reviewer for the constructive and insightful feedback. Below, we provide our detailed responses to each comment.

Q3: How do textual and topological anchors complement each other in anomaly detection?

A3: We thank the reviewer for raising this important question. Textual and topological anchors provide complementary information for anomaly detection. Specifically, textual anchors are selected using LLM-derived confidence scores that assess the semantic coherence and domain relevance of node texts, whereas topological anchors are selected based on the cosine similarity between a node's attribute representation and the mean representation of its neighbors, capturing node-neighborhood consistency. For example, a node may contain semantically coherent text while being connected to semantically unrelated neighbors, which can be identified through topological anchors. Conversely, a node may exhibit semantic inconsistencies even when its neighborhood provides weak anomaly signals. Therefore, by integrating textual and topological anchors, AnTAGAD captures complementary semantic and structural information, enabling more effective detection of anomalies that may be overlooked by either source alone.

Q4: Further details on the main modules and workflow of Figure 2 are needed.

A4: We thank the reviewer for the valuable suggestion to improve the clarity of Figure 2. We agree that the original caption was insufficiently informative. Accordingly, we have revised the caption in the revised manuscript to provide a step-by-step description of the key modules and the overall data flow. Our description is summarized as follows:

(1) Textual Anchoring: An Agents-Skills-Confidence mechanism evaluates textual reliability to select high-confidence normal and low-confidence anomaly anchors.

(2) Topological Anchoring: Node-neighborhood consistency is used to identify structurally normal and anomalous anchors.

(3) Joint Anchor Learning: Textual and topological anchors are merged into unified normal and anomaly anchor banks, while three contrastive objectives jointly optimize representations.

(4) Anomaly Scoring and Prediction: For an unseen target graph, the trained encoders generate node representations, which are compared with the source-domain anchor banks using Top-k similarity retrieval. The final anomaly score combines the difference between anomaly- and normal-anchor similarities with a neighborhood-dispersion term to identify anomalous nodes.

We have incorporated these detailed descriptions into the revised caption of Figure 2 to improve its clarity and help readers better understand the overall architecture and data flow of AnTAGAD.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.