When Prompts Ignore Structure: Graph-Based Attribute Reasoning for Calibrated VLMs

Aditya Sharma; Pranav Murthy Yeluripaty; Ramya Hebbalaguppe; Tanay Sodha; Vinti Agarwal

REVIEW 5 major objections 6 minor 42 references

When Prompts Ignore Structure: Graph-Based Attribute Reasoning for Calibrated VLMs

T0 review · 5 major / 6 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Graph-structured attribute selection, not the tuning loss, drives calibration gains in vision-language models: choosing which descriptive prompts to use at test time cuts average Expected Calibration Error from roughly 6–8% to 4.45% across

desk verdict Plausible incremental improvement to TCA with a useful graph-based selection idea, but the reported numbers contain internal inconsistencies that make the headline calibration gains hard to trust. read the letter →

arxiv 2607.07395 v2 pith:ZQSM6BXA submitted 2026-07-08 cs.CV cs.AIcs.LG

Tanay Sodha , Aditya Sharma , Ramya Hebbalaguppe , Vinti Agarwal , Pranav Murthy Yeluripaty This is my paper

classification cs.CVcs.AIcs.LG

keywords test-timeadaptationcalibrationexpectederrorvision-languagemodelsgraphattentionnetworkattributeselectionprompttuningsymbolic

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that how you choose which descriptive attributes to put into a vision-language model's test-time prompts—not just the tuning objective—determines how well calibrated the confidence scores are. It builds a Symbolic Attribute Graph whose nodes are (class, attribute) pairs, trains a Graph Attention Network offline with contrastive loss to get relational attribute embeddings, and then selects attributes per class by either intra-class diversity (ARGTCA-DIV) or inter-class discrimination (ARGTCA-DISC). Across nine benchmarks with a frozen vision-language backbone, the diversity variant reduces average Expected Calibration Error to 4.45%, a 26.7–48% reduction in miscalibration over four prior test-time tuning baselines, while the discrimination variant matches the best accuracy with lower ECE. If correct, this means calibration can be improved at zero additional test-time cost by a purely structural selection step that ignores image content.

What carries the argument

The central object is the Symbolic Attribute Graph (SAG): a node for every (class, attribute) pair, with edges defined purely symbolically—intra-class edges between all attributes of a same class, inter-class edges between nodes that share an identical attribute string. A Graph Attention Network (GAT) propagates information over these edges under a supervised contrastive loss, refining the frozen CLIP text embeddings into relational embeddings. The two selection criteria (DIV: minimum cosine similarity pair within a class; DISC: maximum mean angular distance from all other-class nodes) then pick M'=2 attributes per class, which initialize the TCA test-time prompt-tuning loss. The GAT and sel

What would settle it

Run ARGTCA on a dataset where the same visual property is described by different words across classes (e.g., some classes use 'rough', others 'bumpy'), so the exact-string inter-class edges do not connect them; if the reported ECE gains shrink or vanish compared with a variant that adds semantic-similarity edges (e.g., cosine threshold on CLIP embeddings), then the symbolic edge rule, not relational reasoning, is the load-bearing component.

Watch

Extended reading notes

Core claim

The central claim is that modeling symbolic attribute interactions—the relationships among descriptive words attached to classes—improves calibration of vision-language models at test time. Instead of treating attributes as a flat set ranked by similarity to the class name, ARGTCA constructs a graph over (class, attribute) pairs, with intra-class edges connecting all attributes of a class and inter-class edges connecting nodes that share the exact same attribute string. A Graph Attention Network trained with supervised contrastive loss then produces embeddings in which same-class attributes cluster and cross-class attributes separate. Two selection criteria read off these embeddings: DIV pic

Load-bearing premise

The entire gain rests on the assumption that cosine geometry in the frozen text-encoder space, plus exact string equality of attribute tokens, is enough to capture the relational structure relevant for calibration—if that geometry or the exact-token edge rule fails to relate near-synonym attributes, the graph has nothing useful to propagate and the selection inherits the limitation.

Editorial extensions

If this is right

If the claim holds, calibration gains can be obtained purely from attribute selection, with zero added per-image compute.
The diversity and discrimination criteria give practitioners a dial: DIV for broad cross-dataset calibration, DISC when inter-class confusion dominates.
The results reinforce the view that calibration and top-1 accuracy are largely decoupled and shaped by the geometry of text feature space rather than by accuracy alone.
Because the graph is built from class names and attribute strings, its structure encodes dataset-specific semantics; extending the method to new datasets requires re-running the offline graph phase.
The ablation isolating random selection suggests that the geometric selection criteria—not graph training per se—are the operative mechanism for calibration improvement.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A reader could test whether the GAT is necessary: applying the same contrastive objective to attributes grouped only by class membership, without graph edges, might yield similar selection quality on non-texture datasets; if so, the symbolic edges mainly matter on datasets where near-synonym attributes cluster.
The exact-string edge rule (a_i = a_j) misses semantically equivalent attributes expressed with different words; replacing it with a cosine-similarity threshold or LLM-synonym edges could plausibly extend the gains to datasets like DTD, which the paper identifies as its own failure case.
Because the graph is built from the test-time class vocabulary, the method as stated cannot transfer to open-set classes or streaming class discovery; a dynamic graph that grows with new classes would be needed to preserve calibration in those settings.
The reported averages rest on hyperparameters tuned on a single validation set (Caltech101); the ordering of DIV and DISC may shift when hyperparameters are selected per dataset.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: worth reading if you work on test-time calibration for CLIP, but don't take the headline numbers at face value. The core idea is to replace TCA's flat attribute selection with a graph attention network over (class, attribute) nodes, then select attributes by intra-class diversity or inter-class discriminability. That is a genuine, if incremental, departure from the TCA/C-TPT/O-TPT/A-TPT line, and the offline training means zero extra test-time cost. The two selection rules are simple, the ablation includes the right controls (edges removed, no GAT training, random selection), and the writing is clear.

The problems are about evidence quality. First, the reported average ECE for ARGTCA-DISC in Table 1 (5.90%) does not match the nine table cells, which average to about 5.46%. That's an arithmetic inconsistency that should have been caught. Second, the ablation text says removing edges degrades DTD ECE for both DIV and DISC, but Table 2 shows DISC improves from 6.36 to 5.31 when edges are removed. That directly contradicts the claim that edge topology drives the gains. Third, there are no repeated runs or error bars anywhere, so we don't know how stable the ECE numbers are. The hyperparameter sweeps are a good start but don't compensate for missing seed variance.

The deeper concern is whether the graph actually does what the motivation says. The motivating failure mode is near-synonymous attributes like rough/bumpy/grainy clustering on the hypersphere. But the SAG edge rule only connects identical attribute strings across classes, and intra-class edges connect everything in the same class. So near-synonyms are not specially treated; they're just connected like any other same-class pair. Meanwhile the contrastive objective pulls same-class embeddings together, which would compress the intra-class diversity that ARGTCA-DIV is supposed to exploit. That's a real tension the paper doesn't address. The examples in Table 3 look reasonable—"crystals, glittering" is a decent diverse pair—but that's anecdotal. The exact-token edge rule is a genuine limitation and should be acknowledged, ideally with a synonym-aware variant.

All that said, I don't think the central idea is wrong. The method is sensible and the DIV variant's average ECE is consistent with its table cells. With better reporting—corrected averages, seed variance, and a discussion of the edge-rule limitation—this could be a solid contribution. As is, I'd send it to peer review, but I'd want the authors to fix the arithmetic and run the ablations multiple times. The paper is for researchers working on VLM calibration and test-time adaptation; it's not a major breakthrough, but it's a useful incremental step.

Referee Report

5 major / 6 minor

Summary. The paper proposes ARGTCA, a two-phase method for test-time prompt-tuning calibration of CLIP. Phase 1 constructs a Symbolic Attribute Graph (SAG) whose nodes are (class, attribute) pairs initialized from CLIP text embeddings; intra-class edges connect all attributes of the same class and inter-class edges connect identical attribute strings across classes. A Graph Attention Network is trained with a supervised contrastive loss to refine these node embeddings. Phase 2 selects M' attributes per class using one of two criteria: ARGTCA-DIV (minimum pairwise cosine similarity within a class) or ARGTCA-DISC (maximum mean angular distance to all other-class nodes), then runs the existing TCA test-time tuning procedure unchanged. The authors report that ARGTCA-DIV achieves the best average ECE of 4.45% across nine benchmarks, with ARGTCA-DISC as second best, and argue from ablations that graph edge structure, contrastive training, and geometric selection each contribute independently.

Significance. The problem addressed is important: entropy-based test-time prompt tuning tends to degrade calibration, and prior work has mostly treated LLM-generated attributes as a flat set. The idea of using relational structure among attributes to choose which attributes initialize prompts is timely and operationally attractive because the graph is trained offline and adds no test-time cost. The manuscript evaluates across nine datasets against several recent baselines. However, the main causal claim that graph-structured reasoning is the key driver is currently undermined by internal inconsistencies in the reported numbers and ablations, and by the absence of repeated-run statistics. If these issues are resolved, the contribution could be a useful step toward structure-aware attribute selection for calibrated test-time adaptation.

major comments (5)

[§5.1, Table 1] The reported average ECE for ARGTCA-DISC is 5.90%, but the nine per-dataset ECE values in Table 1 (1.62, 3.37, 4.29, 3.57, 2.06, 4.25, 6.36, 19.81, 3.81) average to 5.46%. The reductions '2.8%, 9.2%, 23.7%, and 31.0%' in §5.1 and the abstract's '~17%' are therefore not reproducible from the table. Please correct the arithmetic and propagate the corrected reductions, or explain the table's definition.
[§5.4, Table 2] The text states that removing edge structure 'degrades ECE on DTD for both Div (3.17→4.91) and Disc (6.36→4.91).' Table 2 lists the zero-edge Disc DTD ECE as 5.31, not 4.91; this is an improvement over the full-model value of 6.36, not a degradation. Moreover, zero-edge Div improves Caltech101 ECE from 2.52 to 2.36. The ablation evidence in the table contradicts the stated conclusion that edge topology is necessary for both variants. Please correct the text and provide a more nuanced analysis of when edges help versus hurt.
[§3.2.3, Eq. (7)-(8)] The GAT's supervised contrastive objective treats every same-class node as a positive, and the SAG adds dense intra-class edges (Eq. 5). Both mechanisms pull same-class attribute embeddings toward a common centroid, reducing intra-class angular spread. ARGTCA-DIV then selects the minimum-cosine pair in this refined space (Eq. 8). The paper does not quantify how much intra-class diversity remains after Phase 1; if the representation is compressed, DIV may be selecting among near-duplicates, which is exactly the failure the paper motivates. Please report intra-class pairwise cosine-similarity distributions for h(0) versus h(L) and demonstrate that the selected pairs retain meaningful CLIP-space diversity.
[§3.2.2, Eq. (5)] Inter-class edges connect only exactly identical attribute strings (a_i = a_j). Near-synonyms such as 'rough', 'bumpy', and 'grainy'—the motivating example in §1—are never directly linked. Thus the SAG cannot encode cross-class relational structure for semantically related but lexically distinct attributes. The paper should either relax the edge rule (e.g., thresholded cosine similarity in CLIP embedding space) or present an analysis showing that exact-token edges already resolve the motivating failure. As written, the claim that the graph captures inter-attribute dependencies is overstated.
[§5.4, Table 2] All ECE numbers in Table 2 are point estimates from a single run. On Caltech101, differences among variants are small (e.g., full Div 2.52 vs. zero-edge Div 2.36), and ECE is sensitive to binning and the stochasticity of test-time prompt tuning. Without repeated runs, standard deviations, or significance tests, the claimed 'key driver' role of edge structure and selection criteria cannot be distinguished from noise. Please provide variance estimates for the main results and ablations.

minor comments (6)

[Abstract] The reductions '~37%' and '~17%' are described as 'over baselines' without specifying which baselines are averaged. Please define the baseline set explicitly.
[§5.2] The text cites O-TPT ECE as 3.61% on Caltech101, but Table 1 reports 3.62%. Please ensure consistency.
[Appendix B, Eq. (14)] The appendix states that the node attention score is used for attribute selection in ARGTCA-DIV and ARGTCA-DISC, but §3.3 defines selection via cosine-based Eq. (8) and Eq. (9). Please clarify which quantity is actually used.
[Table 1] Several cells have missing spacing/separators (e.g., TCA row '11.355.25', O-TPT row '1.807.15'). Please reformat the table. Also, the TCA average ECE computed from its cells is approximately 6.45%, not the listed 6.50%; please check all averages.
[Limitations] The limitations section says 'ARGTCA-DIV tends to improve accuracy while ARGTCA-DISC tends to reduce ECE,' but Table 1 shows the opposite: ARGTCA-DIV has the lower average ECE and ARGTCA-DISC has the higher average accuracy. This sentence appears reversed.
[References] Some reference entries contain truncated author lists (e.g., Khosla et al. with '1 others') and inconsistent formatting. Please clean up the bibliography.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ECE is not fitted and graph selection is a fixed offline preprocessing step.

full rationale

The claimed ECE gains are produced by a pipeline that does not optimize ECE. Phase 1 trains a GAT with supervised contrastive loss (Eq. 7) over (class, attribute) nodes initialized from frozen CLIP EOS embeddings (Eq. 4) and symbolic edges (Eq. 5). The two selection criteria (Eqs. 8 and 9) are deterministic functions of the learned h(L). Phase 2 is the unchanged TCA test-time objective (Eq. 3) with the selected attributes as prompt initializers. No equation defines the reported ECE in terms of the selection scores, and no selection or graph parameter is fitted to ECE, so the central claim is not equivalent to its inputs by construction. The paper inherits GPT-4 attributes and the TCA protocol from Hebbalaguppe et al. (2025), a citation with overlapping authorship, but this is baseline reuse rather than load-bearing circularity: the comparison against TCA is an executed experiment, and the method's internal ablations are separate evidence. The consistency problem in Table 2 (zero-edge Disc ECE for DTD is 5.31 vs. full 6.36, while the text claims edge removal degrades Disc) and the exact-token Einter limitation are correctness/design concerns, not circular reductions.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The central claim rests on: a fixed externally generated attribute pool, the assumption that CLIP's embedding geometry encodes attribute semantics, an exact-token edge rule that does not cover synonymy, offline class-label supervision for the GAT, and a single calibration metric. The graph itself is an internal construction with no independent evidence. These are the main assumptions the reader pays for beyond the method's own training.

free parameters (5)

M' (number of selected attributes per class) = 2
Set to 2 in Phase 2; no sweep or sensitivity analysis for this choice is reported, though it directly controls prompt diversity.
M (number of candidate attributes per class) = 10
Inherited from the TCA attribute set; graph size K×M and all selection operate on this fixed candidate pool.
alpha (inter-class regularization weight) = 10
Default tuned on Caltech101 validation; swept 5–40 on three datasets with reported ECE variation up to ~2%.
beta (intra-class regularization weight) = 35
Default tuned on Caltech101 validation; swept 15–50 on three datasets with reported ECE variation up to ~2%.
GAT hyperparameters (layers, heads, hidden dim, epochs, lr, SupCon temperature) = 2 layers, 4 heads, 512 dims, 100 epochs, lr 1e-3, tau=0.07
Chosen for the offline Phase 1; no ablation or sensitivity analysis is provided for these values, yet the refined embeddings drive the selection.

assumptions (5)

domain assumption LLM-generated attributes adopted from Hebbalaguppe et al. (2025) are informative and sufficient for calibration improvement.
The paper neither regenerates nor validates the attribute pool; every selection criterion operates only over these externally supplied attributes.
domain assumption Cosine similarity in frozen CLIP text/EOS embedding space is a faithful proxy for semantic relatedness of attributes and classes.
Used for node initialization (Eq. 4) and in both selection criteria (Eqs. 8–9); the paper itself notes TCA's similarity-based selection fails on texture datasets, so this proxy is load-bearing.
ad hoc to paper Exact equality of attribute strings (a_i = a_j) is sufficient to define inter-class relational structure.
Eq. (5) creates E_inter only for identical tokens. Semantically overlapping but differently worded attributes (rough, bumpy, grainy) are never connected, which is exactly the paper's motivating failure mode.
domain assumption Class labels may be used offline to define supervised contrastive positives/negatives while still calling the setting 'task-free' or 'no task supervision'.
The SupCon objective (Eq. 7) uses class identity to define positives and negatives; no image-label pairs are used at test time, but class supervision is present in Phase 1.
standard math ECE computed with 20 equal-width bins is a sufficient calibration metric for comparing methods.
Borrowed from TCA; no alternative calibration metrics, confidence histograms beyond one dataset, or statistical tests are reported.

invented entities (1)

Symbolic Attribute Graph (SAG)
purpose: Represents (class, attribute) pairs as nodes with intra-class and inter-class edges, feeding a GAT that produces refined attribute embeddings for selection.
The graph is trained per dataset and evaluated only inside the ARGTCA+TCA pipeline; the paper gives no standalone falsifiable prediction from the graph itself, so it has no independent handle outside this method.

how reviews work

0 comments

Cite this review

Pith. "Pith review of When Prompts Ignore Structure: Graph-Based Attribute Reasoning for Calibrated VLMs." pith.science (2026). https://pith.science/paper/ZQSM6BXA

@misc{pith2026260707395,
  author       = {Pith},
  title        = {Pith review of: When Prompts Ignore Structure: Graph-Based Attribute Reasoning for Calibrated VLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZQSM6BXA}},
  note         = {Machine review of arXiv:2607.07395}
}

read the original abstract

Reliable confidence estimation remains a key limitation of test-time adaptation in vision-language models (VLMs), where prompt tuning improves zero-shot accuracy but often degrades calibration due to entropy-driven overconfidence. Prior approaches mitigate this using LLM-derived class attributes and contrastive regularization, yet treat attributes independently, ignoring their relational structure. We propose ARGTCA, which represents (class, attribute) pairs as nodes in a Symbolic Attribute Graph and trains a Graph Attention Network (GAT) using contrastive objectives to produce structurally informed embeddings that capture inter-attribute dependencies. We introduce two attribute selection strategies: ARGTCA-DIV for intra-class diversity and ARGTCA-DISC for inter-class discrimination. Experiments across nine benchmarks show that ARGTCA-DIV reduces average Expected Calibration Error (ECE) by approximately ~37% over baselines, while ARGTCA-DISC consistently performs as the second-best variant, reducing average ECE by approximately ~17% over baselines. These results suggest that modeling symbolic attribute interactions provides a principled approach for reliable test-time adaptation in VLMs.

Figures

Figures reproduced from arXiv: 2607.07395 by the authors.

**Figure 2.** Overview of ARGTCA attribute selection pipeline. ❶ An LLM generates M candidate attributes per class. ❷ These are used to construct the Symbolic Attribute Graph (SAG) with intra-class (Eintra) and inter-class (Einter) edges. ❸ A Graph Attention Network (GAT) is trained via supervised contrastive loss (LSupCon) to propagate structural relationships across the SAG, producing attribute embeddings. ❹ Two selection crite… view at source ↗

**Figure 3.** Radar plot comparing ECE (lower is better, smaller area is better) of TCA, O-TPT, A-TPT, and ARGTCA-DIV (Ours) across nine datasets (ViT-B/16). ARGTCA-DIV consistently occupies the innermost region, with the largest gains on Aircraft, Flower, and EuroSAT. (red) predictions at high confidence levels, indicating systematic overconfidence. ARGTCA-DISC achieves the lowest ECE of 1.62% at 93.2% accuracy, with incorrec… view at source ↗

Figures from the paper (6 more)

**Figure 5.** Figure 5: ECE vs. α (inter-class regularization weight) on Caltech101, OxfordPets, and DTD (β=35 fixed). 15 25 35 50 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0 ECE (%) ARGTCA-DIV ECE ( ) Caltech101 OxfordPets DTD= 35 (default) 15 25 35 50 2 3 4 5 6 7 8 9 ECE (%) ARGTCA-DISC ECE ( ) ECE Sen…

**Figure 6.** Figure 6: ECE vs. β (intra-class regularization weight) on Caltech101, OxfordPets, and DTD (α=10 fixed). 5.4 Ablation Studies All ablations use ViT-B/16 on Caltech101, OxfordPets, and DTD. We ablate three components of ARGTCA: (i) edge structure — SAG trained with E=∅, removing…

**Figure 4.** Figure 4: Confidence bin plots on Caltech101 (ViT [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 7.** Figure 7: Test time prompt calibration DTD ECE degrades substantially (Disc: 8.95 vs. 6.36 for the full model), showing that contrastive relational training is necessary for reliable attribute discrimination on texture-heavy datasets. C.4 Random attribute selection. We run Phase…

**Figure 8.** Figure 8: Reliability diagrams across all datasets using ViT-B/16. Each row is a dataset; each column is a method. [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]

**Figure 9.** Figure 9: Radar plot comparing ECE (lower is better, smaller area is better) of TCA, O-TPT, A-TPT, and ARGTCA-DISC (Ours) across nine datasets (ViT-B/16). 16 [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 2 linked inside Pith

[1]

Proceedings of the 38th International Conference on Machine Learning , series =

Radford, Alec and Kim, Jong Wook and Hallacy, Chris and Ramesh, Aditya and Goh, Gabriel and Agarwal, Sandhini and Sastry, Girish and Askell, Amanda and Mishkin, Pamela and Clark, Jack and Krueger, Gretchen and Sutskever, Ilya , title =. Proceedings of the 38th International Conference on Machine Learning , series =. 2021 , publisher =

2021
[2]

and Sung, Yunhsuan and Li, Zhen and Duerig, Tom , title =

Jia, Chao and Yang, Yinfei and Xia, Ye and Chen, Yi-Ting and Parekh, Zarana and Pham, Hieu and Le, Quoc V. and Sung, Yunhsuan and Li, Zhen and Duerig, Tom , title =. Proceedings of the 38th International Conference on Machine Learning , series =
[3]

International Journal of Computer Vision , volume =

Zhou, Kaiyang and Yang, Jingkang and Loy, Chen Change and Liu, Ziwei , title =. International Journal of Computer Vision , volume =
[4]

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , series =

Zhou, Kaiyang and Yang, Jingkang and Loy, Chen Change and Liu, Ziwei , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , series =
[5]

Advances in Neural Information Processing Systems , series =

Shu, Manli and Nie, Weili and Huang, De-An and Yu, Zhiding and Goldstein, Tom and Anandkumar, Anima and Xiao, Chaowei , title =. Advances in Neural Information Processing Systems , series =
[6]

Proceedings of the IEEE/CVF International Conference on Computer Vision , series =

Feng, Chun-Mei and Yu, Kai and Liu, Yong and Khan, Salman and Zhong, Wangmeng , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision , series =
[7]

Advances in Neural Information Processing Systems , series =

Mirza, Muhammad Uzair and Khalid, Fahad Shahbaz and Siddiqui, Fawad and Khan, Salman , title =. Advances in Neural Information Processing Systems , series =
[8]

, title =

Guo, Chuan and Pleiss, Geoff and Sun, Yu and Weinberger, Kilian Q. , title =. Proceedings of the 34th International Conference on Machine Learning , series =

Show all 42 references

[9]

International Conference on Learning Representations , series =

Yoon, Jihoon and Oh, Jae-hun and Kim, Byungho and Kweon, In So and Choi, Jongwoo , title =. International Conference on Learning Representations , series =
[10]

arXiv preprint arXiv:2506.22819 , year =

Hebbalaguppe, Ramya and Kandar, Tamoghno and Nagpal, Abhinav and Arora, Chetan , title =. arXiv preprint arXiv:2506.22819 , year =

arXiv
[11]

and others , title =

Sharifdeen, A. and others , title =. International Conference on Learning Representations , series =
[12]

Proceedings of the 22nd International Conference on Machine Learning , series =

Niculescu-Mizil, Alexandru and Caruana, Rich , title =. Proceedings of the 22nd International Conference on Machine Learning , series =
[13]

, title =

Platt, John C. , title =. Advances in Large Margin Classifiers , pages =. 1999 , publisher =

1999
[14]

Graph Attention Networks , booktitle =

Veli. Graph Attention Networks , booktitle =
[15]

and Welling, Max , title =

Kipf, Thomas N. and Welling, Max , title =. International Conference on Learning Representations , series =
[16]

Advances in Neural Information Processing Systems , series =

Hamilton, Will and Ying, Zhitao and Leskovec, Jure , title =. Advances in Neural Information Processing Systems , series =
[17]

Advances in Neural Information Processing Systems , series =

Wang, Xiyuan and Liu, Muhan and Ji, Meng , title =. Advances in Neural Information Processing Systems , series =
[18]

Advances in Neural Information Processing Systems , series =

Li, Xin and Lian, Wenxuan and Lu, Zhixin and Bao, Jingwen and Liu, Jiaming and Li, Junliang and Lan, Jiaxin and Zhu, Lingling , title =. Advances in Neural Information Processing Systems , series =
[19]

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , series =

Zheng, Xiaobao and Ji, Liqiang and Hong, Changhao and Ying, Nong and Liu, Mengqi and Zhang, Zheng , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , series =
[20]

Pattern Recognition , keywords =

VCGPrompt: Visual Concept Graph-Aware Prompt Learning for Vision-Language Models. Pattern Recognition , keywords =
[21]

Findings of the Association for Computational Linguistics: ACL , year =

Chen, Yukun and Li, Yanggang and others , title =. Findings of the Association for Computational Linguistics: ACL , year =
[22]

Advances in Neural Information Processing Systems , series =

Khosla, Prannay and Tian, Yonglong and Wang, Huiwen and Liu, Chen and Valmadre, Jack and Tian, Chen and Norouzi, Mohammad and others , title =. Advances in Neural Information Processing Systems , series =
[23]

Representation Learning with Contrastive Predictive Coding , journal =

Oord, A. Representation Learning with Contrastive Predictive Coding , journal =
[24]

Proceedings of the 37th International Conference on Machine Learning , series =

Chen, Ting and Kornblith, Simon and Norouzi, Mohammad and Hinton, Geoffrey , title =. Proceedings of the 37th International Conference on Machine Learning , series =
[25]

Brown, Tom B. and Mann, Benjamin and Ryder, Nick and Subbiah, Melanie and Kaplan, Jared and Dhariwal, Prafulla and Neelakantan, Arvind and Shyam, Pranav and Sastry, Girish and Askell, Amanda and others , title =. Advances in Neural Information Processing Systems , series =
[26]

Proceedings of the CVPR 2004 Workshop on Generative-Model Based Vision , year =

Fei-Fei, Li and Fergus, Rob and Perona, Pietro , title =. Proceedings of the CVPR 2004 Workshop on Generative-Model Based Vision , year =

2004
[27]

Proceedings of the 4th IEEE Workshop on 3D Representation and Recognition , pages =

Krause, Jonathan and Stark, Michael and Deng, Jia and Fei-Fei, Li , title =. Proceedings of the 4th IEEE Workshop on 3D Representation and Recognition , pages =
[28]

Proceedings of the 6th Indian Conference on Computer Vision, Graphics and Image Processing , pages =

Nilsback, Maria-Elena and Zisserman, Andrew , title =. Proceedings of the 6th Indian Conference on Computer Vision, Graphics and Image Processing , pages =
[29]

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

Cimpoi, Mircea and Maji, Subhransu and Kokkinos, Iasonas and Mohamed, Sammy and Vedaldi, Andrea , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =
[30]

Maji, Subhransu and Rahtu, Esa and Kannala, Juho and Blaschko, Matthew and Vedaldi, Andrea , title =
[31]

Proceedings of the 13th European Conference on Computer Vision , series =

Bossard, Lukas and Guillaumin, Matthieu and Van Gool, Luc , title =. Proceedings of the 13th European Conference on Computer Vision , series =
[32]

and Vedaldi, Andrea and Zisserman, Andrew and Jawahar, C

Parkhi, Omkar M. and Vedaldi, Andrea and Zisserman, Andrew and Jawahar, C. V. , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =
[33]

arXiv preprint arXiv:1212.0402 , year =

Soomro, Khurram and Zamir, Amir Roshan and Shah, Mubarak , title =. arXiv preprint arXiv:1212.0402 , year =

arXiv
[34]

IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , volume =

Helber, Patrick and Bischke, Benjamin and Dengel, Andreas and Borth, Damian , title =. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , volume =
[35]

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

He, Kaiming and Zhang, Xiangyu and Ren, Shaoqing and Sun, Jian , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =
[36]

International Conference on Learning Representations , series =

Dosovitskiy, Alexey and Beyer, Lucas and Kolesnikov, Alexander and Weissenborn, Dirk and Zhai, Xiaohua and Unterthiner, Thomas and Dehghani, Mostafa and Minderer, Matthias and Heigold, Georg and Gelly, Sylvain and Uszkoreit, Jakob and Houlsby, Neil , title =. International Con...
[37]

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

Zhou, Ziqin and Lei, Yinjie and Zhang, Bowen and Liu, Lingqiao and Liu, Yifan , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =
[38]

International Conference on Learning Representations , series =

A-TPT: Angular Diversity Calibration Properties for Test-Time Prompt Tuning of Vision-Language Models , author=. International Conference on Learning Representations , series =
[39]

Krause, Jonathan and Stark, Michael and Deng, Jia and Fei-Fei, Li , booktitle =. 3
[40]

Blaschko and Andrea Vedaldi , title =

Subhransu Maji and Esa Rahtu and Juho Kannala and Matthew B. Blaschko and Andrea Vedaldi , title =. CoRR , volume =
[41]

Indian Conference on Computer Vision, Graphics and Image Processing , year =

Automated Flower Classification over a Large Number of Classes , author =. Indian Conference on Computer Vision, Graphics and Image Processing , year =
[42]

2023 , eprint=

Learning Hierarchical Prompt with Structured Linguistic Knowledge for Vision-Language Models , author=. 2023 , eprint=

2023

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [1]

Proceedings of the 38th International Conference on Machine Learning , series =

Radford, Alec and Kim, Jong Wook and Hallacy, Chris and Ramesh, Aditya and Goh, Gabriel and Agarwal, Sandhini and Sastry, Girish and Askell, Amanda and Mishkin, Pamela and Clark, Jack and Krueger, Gretchen and Sutskever, Ilya , title =. Proceedings of the 38th International Conference on Machine Learning , series =. 2021 , publisher =

2021

[2] [2]

and Sung, Yunhsuan and Li, Zhen and Duerig, Tom , title =

Jia, Chao and Yang, Yinfei and Xia, Ye and Chen, Yi-Ting and Parekh, Zarana and Pham, Hieu and Le, Quoc V. and Sung, Yunhsuan and Li, Zhen and Duerig, Tom , title =. Proceedings of the 38th International Conference on Machine Learning , series =

[3] [3]

International Journal of Computer Vision , volume =

Zhou, Kaiyang and Yang, Jingkang and Loy, Chen Change and Liu, Ziwei , title =. International Journal of Computer Vision , volume =

[4] [4]

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , series =

Zhou, Kaiyang and Yang, Jingkang and Loy, Chen Change and Liu, Ziwei , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , series =

[5] [5]

Advances in Neural Information Processing Systems , series =

Shu, Manli and Nie, Weili and Huang, De-An and Yu, Zhiding and Goldstein, Tom and Anandkumar, Anima and Xiao, Chaowei , title =. Advances in Neural Information Processing Systems , series =

[6] [6]

Proceedings of the IEEE/CVF International Conference on Computer Vision , series =

Feng, Chun-Mei and Yu, Kai and Liu, Yong and Khan, Salman and Zhong, Wangmeng , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision , series =

[7] [7]

Advances in Neural Information Processing Systems , series =

Mirza, Muhammad Uzair and Khalid, Fahad Shahbaz and Siddiqui, Fawad and Khan, Salman , title =. Advances in Neural Information Processing Systems , series =

[8] [8]

, title =

Guo, Chuan and Pleiss, Geoff and Sun, Yu and Weinberger, Kilian Q. , title =. Proceedings of the 34th International Conference on Machine Learning , series =