Pith. sign in

REVIEW 3 major objections 5 minor 244 references

Natural captions are incomplete constraints: align vision to the consensus core, suppress residual over-commitment, and confidence follows how strongly text pins the image.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-12 04:34 UTC pith:D4AXM4J3

load-bearing objection Solid dual-encoder recipe that treats captions as partial constraints; gains look real, but synthetic multi-view construction is the soft underbelly. the 3 major comments →

arxiv 2607.03143 v1 pith:D4AXM4J3 submitted 2026-07-03 cs.CV cs.AI

Text as Partial Constraint: Core-Residual Alignment for Robust Vision-Language Learning

Chengzhen Yu , Canran Xiao , Siyuan Ma , Yang Liu This is my paper
classification cs.CV cs.AI
keywords vision-language alignmentpartial language constraintscore-residual decompositionmulti-view consensusuncertainty-aware contrastive learningzero-shot robustnessLVLM hallucination
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Vision–language models trained on natural captions treat each description as a complete target, yet real captions omit details, paraphrase freely, and leave much unsaid. That mismatch makes similarity brittle under rewording and overconfident when language is weak. This paper argues that multi-view captions of the same image expose a shared semantic core and an orthogonal “unsaid” residual, and that robust alignment should match images only to the core while explicitly discouraging residual commitment. TPC distills a multi-view consensus core as the training target, learns a lightweight single-view core filter so standard one-query inference still works, penalizes image–residual correlation, and softens contrastive sharpness when caption views disagree. The resulting representations improve clean and adversarial zero-shot accuracy and transfer as a drop-in vision encoder that reduces hallucination in large vision–language models. A sympathetic reader cares because underspecified language is the default in open-vocabulary recognition, retrieval, and interactive multimodal systems; calibrating alignment to partial constraints is a practical route to more reliable matching.

Core claim

Treating multi-view natural captions as incomplete supervision—aligning images to a consensus semantic core, predicting that core from a single caption at test time, suppressing correlation with the orthogonal unsaid residual, and scaling contrastive temperature by caption-view disagreement—yields vision–language representations whose matching is more stable under paraphrase and whose confidence better tracks how strongly text constrains the image, with measured gains on zero-shot recognition, adversarial robustness, and LVLM transfer.

What carries the argument

Core–residual alignment (TPC): consensus core ci as the normalized mean of multi-view text embeddings; single-view core filter ψθ; unsaid residual ri,k orthogonal to the predicted core; non-commitment loss that squares image–residual correlation; and uncertainty-aware InfoNCE with temperature τi = τ0(1 + γui).

Load-bearing premise

The normalized average of several captions of the same image is a faithful target for the shared meaning, and the extra captions built for single-caption web data (shortened cores, paraphrases, back-translations) truly expose underspecification rather than train on artifacts.

What would settle it

Train and evaluate with only genuine multi-caption human data versus the paper’s synthesized-view recipe on single-caption corpora; if the clean/robust ImageNet and Avg-14 margins and residual-leakage reduction disappear when synthetic views are removed or replaced by unfiltered generation, the partial-constraint claim does not hold under realistic supervision.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes Text as Partial Constraint (TPC), a core–residual vision–language alignment framework that treats multi-view captions as incomplete supervision. It defines a multi-view consensus core ci (Eq. 2) as the alignment target, trains a linear single-view core filter ψθ (Eq. 4) for one-text inference, penalizes image–residual correlation via Lnc (Eqs. 5–6), and softens InfoNCE with disagreement-based temperature τi=τ0(1+γui) (Eqs. 3, 7). Theory motivates residual non-commitment as a DRO-style worst-case bound (Thm. 5.1) and multi-view consensus as denoising (Thm. 5.2). Empirically, TPC reports 81.42/64.05 ImageNet clean/robust Top-1, 76.19/52.03 on Avg-14, and improved LLaVA-1.5-7B transfer (85.16 POPE F1, 59.57 OKVQA), with ablations, residual-leakage, risk–coverage, caption-sensitivity, controlled-deletion, and backbone analyses.

Significance. If the gains isolate the partial-constraint principle rather than multi-text augmentation, the work is a clear and practical contribution to robust open-vocabulary alignment and LVLM grounding under underspecified language. Strengths include a diagnostic preliminary study (Sec. 3), single-text deployability via ψθ, multi-front empirical stress tests (residual leakage Fig. 5, risk–coverage Fig. 6, controlled deletion Fig. 11, caption-choice sensitivity Fig. 7), seed variability (Table 5), and compute disclosure. The theory is motivational rather than machine-checked, but Thm. 5.1 cleanly links core alignment and residual sensitivity to worst-case similarity. The main significance risk is whether headline numbers credit core–residual alignment or simply richer multi-view text supervision.

major comments (3)
  1. [§6.2–6.3, Tables 1–3] The central claim that gains come from modeling text as a partial constraint is not isolated from multi-text augmentation. Table 3 removes TPC components but does not include a matched multi-view baseline that uses the same K captions (human or synthetic) with standard multi-positive contrastive loss and no core/residual machinery. Without that control, the ImageNet/Avg-14/LVLM improvements (Tables 1–2) may partly reflect extra text views rather than consensus-core alignment and residual non-commitment (Eqs. 6–9).
  2. [Appendix B.2; §5.2, Thm. 5.2; Eqs. 2–6] Appendix B.2 synthesizes views for single-caption web data via core-only deletion, paraphrase, and back-translation with a weak frozen-encoder filter (cosine ≥0.25). Thm. 5.2 models views as zero-mean noise around a latent core μi, but systematic deletion removes structure rather than adding noise, so the denoising justification does not transfer to the large-scale regime that drives dual-encoder training. The paper should either (i) report human-multi-caption-only vs. synthetic-view training splits for residual leakage and main metrics, or (ii) substantially strengthen faithfulness checks and explicitly limit the theory’s scope.
  3. [§6.1, Table 1; Appendix B.4–B.5] Training data composition, step budget, and compute matching for the headline ImageNet/Avg-14 numbers versus robust baselines (TeCoA, FARE, Adv-W2S) remain underspecified beyond “match when feasible” (§6.1, B.4–B.5). Because adversarial methods incur different costs and TPC’s multi-view text path changes supervision density, a compute- and data-matched block (or clear external-checkpoint labeling) is needed before attributing +5–7 point robust gains primarily to core–residual alignment.
minor comments (5)
  1. [Table 1; Appendix B.3] Adversarial radius is inconsistent: Table 1 caption states AutoAttack ϵ=2/255, while Appendix B.3 defaults to ϵ=4/255. Please unify the protocol and restate which radius was used for all methods.
  2. [Abstract; §1; §4; §6] Title/method casing is broken in several places (“TEXT ASPARTIAL CONSTRAINT”, “TEXT ASPARTIALCONSTRAIN”). Normalize to “Text as Partial Constraint (TPC)” throughout.
  3. [Fig. 4] Figure 4 axis labels show encoding artifacts (“/glyph1197umber of views”, “/glyph1197on-com m itm ent”). Clean the PDF text layer.
  4. [§2] Related Work cites three concurrent ICLR 2026 papers (Xiao et al., Zhang et al., Zhou et al.) as “adjacent”; a one-sentence differentiation of partial-constraint alignment vs. continual/compositional adaptation would help readers.
  5. [§4.6; §6.1] In §4.6, state explicitly that τ(y) calibration preserves ranking while only changing softmax sharpness; this is said once but should be repeated near the LVLM transfer protocol so readers do not confuse confidence calibration with re-ranking.

Circularity Check

0 steps flagged

No significant circularity: TPC is a standard method+benchmark paper; theory motivates objectives and reported metrics are external evaluations, not quantities forced by construction.

full rationale

Walking the load-bearing chain: (i) the consensus core ci (Eq. 2), residual ri,k (Eq. 5), Lnc (Eq. 6), Lcore (Eq. 7), and Lagree (Eq. 8) are design choices that define the training objective, not claimed first-principles predictions of external observables; (ii) Theorems 5.1–5.2 are motivational DRO/PAC analyses showing that worst-case similarity under a residual ambiguity set improves by raising core alignment and lowering residual sensitivity, and that multi-view averaging reduces noise—standard justification of the chosen losses, not identities that make the reported accuracies true by definition; (iii) headline numbers (ImageNet clean/robust, Avg-14, POPE, OKVQA, COCO/Flickr) are measured on external held-out benchmarks against independent baselines; (iv) hyperparameter tuning of γ, λnc, λagree, τ0, K on validation is ordinary ML practice, not renaming a fit as a prediction; (v) self-citations (e.g., Xiao/Zhang/Zhou 2026) appear only as adjacent continual-learning work and are not load-bearing uniqueness or ansatz imports for TPC. Weaknesses about synthetic multi-view construction (Appendix B.2) and whether ci is a faithful latent core are assumption/correctness risks, not circular reductions of outputs to inputs. Score 0 with empty steps is therefore the honest finding.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 4 invented entities

TPC’s claim rests on standard dual-encoder contrastive learning plus several modeling choices: multi-view captions share a usable consensus core; orthogonal residual text is “unsaid” and should not drive similarity; disagreement ui is a valid uncertainty proxy; and synthetic views can stand in for human multi-captions. Free parameters control temperature, residual penalty, agreement, and view count. Invented operational entities are the consensus core, residual, core filter, and disagreement statistic—engineering constructs with diagnostic handles inside the paper, not external physical entities.

free parameters (6)
  • base temperature τ0 = 0.07
    Sets contrastive sharpness; default 0.07 in CLIP range, with sensitivity sweep in Fig. 4.
  • uncertainty scale γ = 1.0
    Scales instance temperature τi=τ0(1+γ ui); tuned in {0,0.5,1,2}, default 1.0.
  • non-commitment weight λnc = 0.1
    Weight on residual correlation penalty Lnc; tuned in {0.05,0.1,0.2,0.5}, default 0.1.
  • core agreement weight λagree = 1.0
    Weight on single-view core matching multi-view consensus; tuned around 0.5–2, default 1.0.
  • number of caption views K = 4–5
    Controls multi-view consensus/uncertainty estimation; K=5 on COCO/Flickr, K=4 on synthesized web views; performance saturates near 4–5.
  • synthetic-view cosine filter threshold = ≥0.25
    Frozen text-encoder similarity gate (≥0.25 default) used to accept/reject generated paraphrases and deletions for single-caption corpora.
axioms (6)
  • domain assumption Multiple textual views of an image share a stable semantic core while differing in omitted or ambiguous details.
    Stated as the motivating premise in §1 and formalized via consensus core ci in Eq. (2).
  • ad hoc to paper Components of a caption embedding orthogonal to the (predicted) core are unsaid residuals that should not drive image–text similarity.
    Operationalized by residual definition Eq. (5) and non-commitment loss Eq. (6); central design choice rather than a standard VL axiom.
  • ad hoc to paper Caption-view disagreement ui is a monotone proxy for language underspecification and should soften contrastive updates.
    Defined in Eq. (3) and used for τi in Eq. (7); supported by preliminary diagnostics but still a modeling choice.
  • domain assumption Hemisphere condition t⊤i,k ci ≥ 0 for observed views so that residual radius and ambiguity set A_i are well-defined.
    Assumed in Lemma A.1 for containment of observed views in the spherical-cap ambiguity set.
  • domain assumption Multi-view text embeddings are noisy observations of a latent unit core with bounded zero-mean noise (Eq. 14).
    Noise model for the PAC multi-view denoising theorem (Thm. 5.2).
  • standard math Standard dual-encoder cosine similarity and InfoNCE-style contrastive learning are appropriate for open-vocabulary transfer.
    Inherited from CLIP/ALIGN-style VLP; used throughout setup and Lcore.
invented entities (4)
  • consensus semantic core ci no independent evidence
    purpose: Multi-view normalized mean text direction used as the alignment target instead of any single caption.
    Defined in Eq. (2); central training target. Independent evidence is internal (stability diagnostics, ablations), not an external physical quantity.
  • unsaid residual ri,k no independent evidence
    purpose: Orthogonal component of a caption embedding after removing the predicted core; target of non-commitment penalty.
    Defined in Eq. (5). Diagnostic residual leakage uses a related construction; no external measurement outside this framework.
  • single-view core filter ψθ no independent evidence
    purpose: Linear map recovering the multi-view core from one caption embedding for standard single-query inference.
    Eq. (4); near-identity initialization. Deployable estimator, not an independently observed entity.
  • caption-view disagreement uncertainty ui / residual energy û(y) no independent evidence
    purpose: Scalar uncertainty used to soften training temperature and calibrate inference confidence.
    Eqs. (3) and §4.6; validated via risk–coverage curves but defined within the method.

pith-pipeline@v1.1.0-grok45 · 32229 in / 4204 out tokens · 43784 ms · 2026-07-12T04:34:43.417194+00:00 · methodology

0 comments
read the original abstract

Vision-language alignment powers open-vocabulary recognition, retrieval, and LVLM grounding, yet natural captions are often underspecified, making similarity brittle and overly confident under paraphrase and omitted details. We aim to learn representations whose matching is stable across caption views and whose confidence reflects how strongly text constrains an image. We propose Text as Partial Constraint (TPC), a core-residual alignment framework that treats multi-view captions as incomplete supervision. It distills a consensus semantic core as the alignment target, learns a single-view core predictor for standard inference with one query, and explicitly discourages vision-language similarity from depending on the orthogonal unsaid residual. An uncertainty-aware contrastive objective further softens alignment when caption views disagree, reducing overconfident updates under weak language constraints. Across zero-shot recognition and adversarial robustness, TPC achieves 81.42/64.05 Top-1 clean/robust accuracy on ImageNet and 76.19/52.03 on an Avg-14 transfer suite, while improving LVLM transfer with 85.16 POPE F1 and 59.57 OKVQA accuracy under an LLaVA-1.5-7B stack. These results suggest that modeling text as a partial constraint is a practical and principled route to more reliable vision-language representations under underspecified language supervision.

Figures

Figures reproduced from arXiv: 2607.03143 by Canran Xiao, Chengzhen Yu, Siyuan Ma, Yang Liu.

Figure 1
Figure 1. Figure 1: Unlike prior methods that over-align to each caption as a complete target, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Frozen VLM failure under partial captions. Left: larger caption-view dispersion leads to higher rank volatility and stronger hard-negative confusion. Right: residual leakage concentrates confident retrieval errors. All statistics are computed with frozen embeddings; no proposed module or training loss is used. (Truong et al., 2025). Yet these fixes are constrained by the shared embedding space they build u… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of TPC. TPC treats captions as partial constraints: it distills a multi-view consensus core, learns a single-view core predictor, decomposes each caption into core and unsaid residual compo￾nents, and performs uncertainty-aware alignment that suppresses residual over-commitment. ambiguous details. A robust aligner should match vision to the shared semantics and avoid committing to view-specific de… view at source ↗
Figure 4
Figure 4. Figure 4: Hyperparameter sensitivity of TPC. Mean±std over 3 seeds. CLIP (Pretrain) CLIP (Finetune) TPC (Pretrain) TPC (Finetune) 0.00 0.05 0.10 0.15 0.20 0.25 R e sid u al le a k a g e = (v r)2 0.049 0.071 0.049 0.040 COCO Flickr30K Web-SingleCap [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Residual leakage. Distribution of ℓ = (v ⊤r) 2 before/after fine-tuning; lower indicates less residual commitment. Protocol in App. B.6. Uncertainty Calibration. We evaluate whether the inferred uncertainty uˆ ranks queries by reliability under weak text constraints. We sort queries from low to high uˆ and compute risk–coverage curves by retaining increasingly uncertain queries [PITH_FULL_IMAGE:figures/fu… view at source ↗
Figure 6
Figure 6. Figure 6: Risk–coverage from uˆ. Protocol in App. B.6. 7 Conclusion We address brittleness and overconfidence in vision–language alignment under underspecified captions, and propose TEXT AS PARTIAL CONSTRAINT (TPC) to align images to view-invariant semantics while suppressing view-specific residuals with single-text deployability. Our results show that modeling language as an incomplete constraint and calibrating by… view at source ↗
Figure 7
Figure 7. Figure 7: Caption-choice sensitivity. Violin plots show per-image distributions of RankStd and WorstGap on COCO and Flickr30K. Dots and error bars indicate mean and 95% CI; faint points show dataset-wise means. Lower is better: less variation in retrieval rank across captions of the same image. The total compute for experiments reported in the paper is approximately 2.8K–3.2K A100 GPU-hours, including main training,… view at source ↗
Figure 8
Figure 8. Figure 8: Does ψθ predict the consensus core? Empirical CDFs of ρ = cos(cˆ, c) on COCO, Flickr30K, and a single-caption web-style set. Colors denote methods; line styles denote datasets. A right-shifted curve indicates that single-caption cores cˆ better match the multi-view consensus c. C.2 Single-Caption Core Filter: Does ψθ Recover the Multi-View Consensus? We verify that the single-view core filter ψθ is not a s… view at source ↗
Figure 9
Figure 9. Figure 9: Gain vs. text-only solvability. Each point is an equal-count bin (hundreds of bins per dataset); colors indicate datasets. The trend line is a smoothed running mean over bins (same color per dataset) and an overall trend (black). Higher gain at lower solvability indicates TPC is most beneficial when text is a weak/underspecified constraint. min mid max IN Clean 78.1 80.2 82.3 IN Robust 61.0 63.1 65.2 Avg-1… view at source ↗
Figure 10
Figure 10. Figure 10: Cross-backbone robustness and retrieval. Parallel coordinates over multiple metrics (each axis is min–max normalized for readability). Each backbone is shown with three seed runs (faint) and the mean (bold). Consistent trends across architectures indicate TPC does not rely on a particular backbone family [PITH_FULL_IMAGE:figures/full_fig_p023_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Controlled caption deletion stress test. Heatmaps show Recall@1 (%) under increasing deletion ratios (x-axis) and deletion types (y-axis). Each cell is annotated with Recall@1 / ECE (ECE shown as a small number). TPC maintains higher accuracy and lower ECE as deletions grow, indicating robustness to systematically weakened language constraints. text-only span extractor driven by an off-the-shelf instructi… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

244 extracted references · 25 linked inside Pith

  1. [1]

    The Fourteenth International Conference on Learning Representations , year=

    Reversible primitive--composition alignment for continual vision--language learning , author=. The Fourteenth International Conference on Learning Representations , year=

  2. [2]

    The Fourteenth International Conference on Learning Representations , year=

    Pi-cca: Prompt-invariant cca certificates for replay-free continual multimodal learning , author=. The Fourteenth International Conference on Learning Representations , year=

  3. [3]

    The Fourteenth International Conference on Learning Representations , year=

    Comem: Compositional concept-graph memory for vision--language adaptation , author=. The Fourteenth International Conference on Learning Representations , year=

  4. [4]

    Langley , title =

    P. Langley , title =. Proceedings of the 17th International Conference on Machine Learning (ICML 2000) , address =. 2000 , pages =

    2000

  5. [5]

    Proceedings of the 41st International Conference on Machine Learning , pages=

    Robust CLIP: unsupervised adversarial fine-tuning of vision embeddings for robust large vision-language models , author=. Proceedings of the 41st International Conference on Machine Learning , pages=

  6. [6]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Improved baselines with visual instruction tuning , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  7. [7]

    International conference on machine learning , pages=

    Scaling up visual and vision-language representation learning with noisy text supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  8. [8]

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

    Mitigating object hallucinations in large vision-language models through visual contrastive decoding , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  9. [9]

    Advances in Neural Information Processing Systems , volume=

    Text-guided attention is all you need for zero-shot robustness in vision-language models , author=. Advances in Neural Information Processing Systems , volume=

  10. [10]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Pre-trained model guided fine-tuning for zero-shot adversarial robustness , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  11. [11]

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

    Winoground: Probing vision and language models for visio-linguistic compositionality , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  12. [12]

    Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    VALSE: A task-independent benchmark for vision and language models centered on linguistic phenomena , author=. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  13. [13]

    arXiv preprint arXiv:2601.06165 , year=

    What Users Leave Unsaid: Under-Specified Queries Limit Vision-Language Models , author=. arXiv preprint arXiv:2601.06165 , year=

    Pith/arXiv arXiv

  14. [14]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Vision-language models do not understand negation , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  15. [15]

    Advances in Neural Information Processing Systems , volume=

    Sugarcrepe++ dataset: Vision-language model sensitivity to semantic and lexical alterations , author=. Advances in Neural Information Processing Systems , volume=

  16. [16]

    Advances in neural information processing systems , volume=

    Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality , author=. Advances in neural information processing systems , volume=

  17. [17]

    arXiv preprint arXiv:2508.14264 , year=

    Directed-Tokens: A Robust Multi-Modality Alignment Approach to Large Language-Vision Models , author=. arXiv preprint arXiv:2508.14264 , year=

    arXiv

  18. [18]

    arXiv preprint arXiv:2505.17316 , year=

    Analyzing Fine-Grained Alignment and Enhancing Vision Understanding in Multimodal Language Models , author=. arXiv preprint arXiv:2505.17316 , year=

    Pith/arXiv arXiv

  19. [19]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    Evaluating Object Hallucination in Large Vision-Language Models , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    2023

  20. [20]

    arXiv preprint arXiv:2308.01390 , year=

    Openflamingo: An open-source framework for training large autoregressive vision-language models , author=. arXiv preprint arXiv:2308.01390 , year=

    Pith/arXiv arXiv

  21. [21]

    Advances in neural information processing systems , volume=

    Visual instruction tuning , author=. Advances in neural information processing systems , volume=

  22. [22]

    arXiv preprint arXiv:2510.21323 , year=

    VL-SAE: Interpreting and Enhancing Vision-Language Alignment with a Unified Concept Set , author=. arXiv preprint arXiv:2510.21323 , year=

    arXiv

  23. [23]

    The Eleventh International Conference on Learning Representations , year =

    Understanding Zero-shot Adversarial Robustness for Large-Scale Models , author=. The Eleventh International Conference on Learning Representations , year =

  24. [24]

    arXiv preprint arXiv:2303.15389 , year=

    Eva-clip: Improved training techniques for clip at scale , author=. arXiv preprint arXiv:2303.15389 , year=

    Pith/arXiv arXiv

  25. [25]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Sigmoid loss for language image pre-training , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  26. [26]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Reproducible scaling laws for contrastive language-image learning , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  27. [27]

    International conference on machine learning , pages=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  28. [28]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Are transformers effective for time series forecasting? , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  29. [29]

    International Conference on Machine Learning , pages=

    MOMENT: A Family of Open Time-series Foundation Models , author=. International Conference on Machine Learning , pages=. 2024 , organization=

    2024

  30. [30]

    Timer-XL: Long-Context Transformers for Unified Time Series Forecasting , author=

  31. [31]

    ICLR 2025: The Thirteenth International Conference on Learning Representations , year=

    Time-MoE: Billion-Scale Time Series Foundation Models with Mixture of Experts , author=. ICLR 2025: The Thirteenth International Conference on Learning Representations , year=

    2025

  32. [32]

    arXiv preprint arXiv:2507.14507 , year=

    Diffusion models for time series forecasting: A survey , author=. arXiv preprint arXiv:2507.14507 , year=

    Pith/arXiv arXiv

  33. [33]

    International conference on machine learning , pages=

    Autoregressive denoising diffusion models for multivariate probabilistic time series forecasting , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  34. [34]

    Advances in neural information processing systems , volume=

    Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting , author=. Advances in neural information processing systems , volume=

  35. [35]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Informer: Beyond efficient transformer for long sequence time-series forecasting , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  36. [36]

    The eleventh international conference on learning representations , year=

    Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting , author=. The eleventh international conference on learning representations , year=

  37. [37]

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

    Construct-vl: Data-free continual structured vl concepts learning , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  38. [38]

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

    Continual learning for visual search with backward consistent feature embedding , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  39. [39]

    Advances in neural information processing systems , volume=

    Dark experience for general continual learning: a strong, simple baseline , author=. Advances in neural information processing systems , volume=

  40. [40]

    Workshop on Multi-Task and Lifelong Reinforcement Learning , year=

    Continual learning with tiny episodic memories , author=. Workshop on Multi-Task and Lifelong Reinforcement Learning , year=

  41. [41]

    IEEE Transactions on Instrumentation and Measurement , year=

    GALMOR: Memory-Constrained Continual Learning With Efficient Replay for Fault Diagnosis of Rotating Machinery , author=. IEEE Transactions on Instrumentation and Measurement , year=

  42. [42]

    Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , year =

    Guilin Zhu and Dongyue Wu and Changxin Gao and Runmin Wang and Weidong Yang and Nong Sang , title =. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , year =

  43. [43]

    arXiv preprint arXiv:2503.06683 , year =

    Dynamic Dictionary Learning for Remote Sensing Image Segmentation , author =. arXiv preprint arXiv:2503.06683 , year =

    Pith/arXiv arXiv

  44. [44]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

    SCORE: Scene Context Matters in Open-Vocabulary Remote Sensing Instance Segmentation , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

  45. [45]

    arXiv preprint arXiv:2501.13925 , year =

    GeoPixel: Pixel Grounding Large Multimodal Model in Remote Sensing , author =. arXiv preprint arXiv:2501.13925 , year =

    Pith/arXiv arXiv

  46. [46]

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

    SegEarth-OV: Towards Training-Free Open-Vocabulary Segmentation for Remote Sensing Images , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  47. [47]

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

    SkySense-O: Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  48. [48]

    Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , year =

    Towards Open-Vocabulary Remote Sensing Image Semantic Segmentation , author =. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , year =

  49. [49]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Learning at a glance: Towards interpretable data-limited continual semantic segmentation via semantic-invariance modelling , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 2024 , publisher=

    2024

  50. [50]

    IGARSS 2024-2024 IEEE International Geoscience and Remote Sensing Symposium , pages=

    Self-training and curriculum learning guided dynamic refined network for remote sensing class-incremental semantic segmentation , author=. IGARSS 2024-2024 IEEE International Geoscience and Remote Sensing Symposium , pages=. 2024 , month=

    2024

  51. [51]

    IEEE Transactions on Geoscience and Remote Sensing , volume=

    Domain-Incremental Learning for Remote Sensing Semantic Segmentation With Multifeature Constraints in Graph Space , author=. IEEE Transactions on Geoscience and Remote Sensing , volume=. 2024 , publisher=

    2024

  52. [52]

    Science China Information Sciences , volume=

    Mitigating representation bias for class-incremental semantic segmentation of remote sensing images , author=. Science China Information Sciences , volume=. 2025 , doi=

    2025

  53. [53]

    IEEE Transactions on Geoscience and Remote Sensing , volume=

    MiSSNet: Memory-inspired semantic segmentation augmentation network for class-incremental learning in remote sensing images , author=. IEEE Transactions on Geoscience and Remote Sensing , volume=. 2024 , publisher=

    2024

  54. [54]

    Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , year =

    Yirui Wu and Yuhang Xia and Hao Li and Lixin Yuan and Junyang Chen and Jun Liu and Tong Lu and Shaohua Wan , title =. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , year =

  55. [55]

    Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , year =

    Zhidong Yu and Xiaoman Liu and Jiajun Hu and Zhenbo Shi and Wei Yang , title =. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , year =

  56. [56]

    Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , year =

    Cheng Xu and Weiwen Zhang and Hongrui Zhang and Xuemiao Xu and Huaidong Zhang and Jing Zou and Jing Qin , title =. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , year =

  57. [57]

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

    Instruction-Grounded Visual Projectors for Continual Learning of Generative Vision-Language Models , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  58. [58]

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

    Hongmei Yin and Tingliang Feng and Fan Lyu and Fanhua Shang and Hongying Liu and Wei Feng and Liang Wan , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  59. [59]

    IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews) , volume=

    Understanding video events: A survey of methods for automatic interpretation of semantic occurrences in video , author=. IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews) , volume=. 2009 , publisher=

    2009

  60. [60]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Revisiting the" video" in video-language understanding , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  61. [61]

    Advances in Neural Information Processing Systems , volume=

    A Practitioner's Guide to Real-World Continual Multimodal Pretraining , author=. Advances in Neural Information Processing Systems , volume=

  62. [62]

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

    Yuchen Zhu and Cheng Shi and Dingyou Wang and Jiajin Tang and Zhengxuan Wei and Yu Wu and Guanbin Li and Sibei Yang , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  63. [63]

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

    Kai Fang and Anqi Zhang and Guangyu Gao and Jianbo Jiao and Chi Harold Liu and Yunchao Wei , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  64. [64]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =

    Maoxian Wan and Kaige Li and Qichuan Geng and Weimin Shi and Zhong Zhou , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =. 2025 , pages =

    2025

  65. [65]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =

    Ruitao Wu and Yifan Zhao and Jia Li , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =. 2025 , pages =

    2025

  66. [66]

    Advances in Neural Information Processing Systems , year=

    OVS Meets Continual Learning: Towards Sustainable Open-Vocabulary Segmentation , author=. Advances in Neural Information Processing Systems , year=

  67. [67]

    Advances in Neural Information Processing Systems , year=

    Leveraging Depth and Language for Open-Vocabulary Domain-Generalized Semantic Segmentation , author=. Advances in Neural Information Processing Systems , year=

  68. [68]

    Advances in Neural Information Processing Systems , year=

    Open-Vocabulary Part Segmentation via Progressive and Boundary-Aware Strategy , author=. Advances in Neural Information Processing Systems , year=

  69. [69]

    Advances in Neural Information Processing Systems , year=

    Seg4Diff: Unveiling Open-Vocabulary Segmentation in Text-to-Image Diffusion Transformers , author=. Advances in Neural Information Processing Systems , year=

  70. [70]

    Advances in Neural Information Processing Systems , year=

    OPMapper: Enhancing Open-Vocabulary Semantic Segmentation with Multi-Guidance Information , author=. Advances in Neural Information Processing Systems , year=

  71. [71]

    Advances in Neural Information Processing Systems , year=

    LangHOPS: Language Grounded Hierarchical Open-Vocabulary Part Segmentation , author=. Advances in Neural Information Processing Systems , year=

  72. [72]

    Advances in Neural Information Processing Systems , year=

    Continual Gaussian Mixture Distribution Modeling for Class Incremental Semantic Segmentation , author=. Advances in Neural Information Processing Systems , year=

  73. [73]

    Advances in Neural Information Processing Systems , year=

    Test-Time Adaptation of Vision-Language Models for Open-Vocabulary Semantic Segmentation , author=. Advances in Neural Information Processing Systems , year=

  74. [74]

    Forty-second International Conference on Machine Learning , year=

    Divide and Conquer: Exploring Language-centric Tree Reasoning for Video Question-Answering , author=. Forty-second International Conference on Machine Learning , year=

  75. [75]

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

    VideoLLaMB: Long Streaming Video Understanding with Recurrent Memory Bridges , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  76. [76]

    arXiv preprint arXiv:2507.00469 , year=

    Bisecle: Binding and Separation in Continual Learning for Video Language Understanding , author=. arXiv preprint arXiv:2507.00469 , year=

    Pith/arXiv arXiv

  77. [77]

    Proceedings of the 32nd ACM International Conference on Multimedia , pages=

    Gpt4video: A unified multimodal large language model for lnstruction-followed understanding and safety-aware generation , author=. Proceedings of the 32nd ACM International Conference on Multimedia , pages=

  78. [78]

    2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages=

    Dam: Dynamic adapter merging for continual video qa learning , author=. 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages=. 2025 , organization=

    2025

  79. [79]

    arXiv preprint arXiv:2502.00843 , year=

    VLM-assisted continual learning for visual question answering in self-driving , author=. arXiv preprint arXiv:2502.00843 , year=

    arXiv

  80. [80]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    VQAGuider: Guiding Multimodal Large Language Models to Answer Complex Video Questions , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

Showing first 80 references.