Pith. sign in

REVIEW 5 major objections 5 minor 54 references

MAG: MAnifold Guided Semi-Supervised Multi-modal In-Context Learning

T0 review · 5 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read Unlabeled image-question pairs, mined by graph propagation and pseudo-labeling, lift few-shot multi-modal in-context learning under label scarcity.

desk verdict Plausible two-stage label-propagation method for few-shot multimodal ICL, but the undisclosed split between the unlabeled pool and test set could change the headline result. read the letter →

arxiv 2608.12724 v1 pith:2QEY2XHY submitted 2026-08-13 cs.LG

Zirui Cheng , Xun Xu , Tiankai Chen , Fady Rezk , Bowen Zheng , Xiaodong Shi , Shijie Li , Kangkang Lu
show 2 more authors
Bharadwaj Veeravalli Nancy F. Chen
This is my paper · ORCID
classification cs.LG
keywords in-contextlearningmulti-modallargelanguagemodelssemi-supervisedlabelpropagationdemonstrationselectionpseudo-labelingvisualquestionansweringgraph-basedretrieval
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

Multi-modal in-context learning adapts a model to a new task by showing it a few image-question-answer examples, with no weight updates, but performance depends heavily on which examples are shown. This paper claims that abundant unlabeled image-question pairs can be exploited without training: first, a relevance score is propagated on a graph built from descriptions produced by a multi-modal large language model (MLLM), isolating a compact set of unlabeled samples whose predicted answers are then used as provisional labels (pseudo-labels); second, the final demonstrations are chosen by propagating query-conditioned relevance on separate image and text graphs and combining the two scores. The reported result is consistent gains over labeled-only retrieval baselines across eight benchmarks under extreme label scarcity (15 labeled examples), with the largest gains on reasoning-heavy tasks. The paper concludes that pseudo-label quality matters more than quantity, that textual structure is the right carrier for the coarse filtering stage, and that late-fused visual and textual structure is needed for the final selection.

What carries the argument

The carrying object is a pair of $k$-nearest-neighbour graphs (each node links to its $k$ most similar nodes) with closed-form relevance-score propagation, and the paper packages the two-stage use of that machinery under the name MAG (manifold-guided semi-supervised demonstration selection). In Stage 1, nodes are text embeddings $h_i^{(t)} = f_t([d_i, q_i])$, where $d_i$ is an MLLM-generated description and $q_i$ is the question; starting from a vector that is 1 on labeled nodes and 0 on unlabeled nodes, the relevance scores converge to $I^{(t)}_{(\infty)} = (1-\alpha)(I-\alpha \hat{W}^{(t)})^{-1} I^{(t)}_{(0)}$, and the top-scoring unlabeled nodes are pseudo-labeled. In Stage 2, the same propagation is run separately on a visual graph and a textual graph over the expanded candidate pool, each initialized from the query's nearest neighbours, and the two score vectors are combined by the late fusion $I = \beta I^{(v)}_{(\infty)} + (1-\beta) I^{(t)}_{(\infty)}$. The closed form is what makes the method training-free: no learning is involved, and computation reduces to solving a sparse linear system, which the paper reports scales near-linearly in the number of samples.

What would settle it

Compute the overlap between Stage 1's top-ranked unlabeled samples and an oracle ranking built from ground-truth labels or human-written captions on the same eight benchmarks, then corrupt or paraphrase a random fraction of the MLLM-generated descriptions and re-run MAG. If overlap is low yet accuracy is unchanged, the paper's dependence on description quality is contradicted; if accuracy collapses under corruption, the dependence is confirmed.

Watch

Extended reading notes

Core claim

The paper's central discovery is that relevance propagation on manifolds can turn an unlabeled pool into better in-context demonstrations. Specifically, the proposed MAG (manifold-guided semi-supervised demonstration selection) framework shows that closed-form label propagation on a graph of text embeddings—formed by concatenating MLLM-generated image descriptions with the task question—ranks unlabeled samples by relevance to the labeled data well enough that pseudo-labeling only the top 45 out of 585 yields larger ICL gains than random pseudo-labeling or no pseudo-labeling. It further shows that for selecting the demonstrations actually shown to the model, neither modality alone is enough: visual-only graphs mislead on language-heavy tasks such as emotion recognition and scene text reading, while text-only graphs collapse on visually grounded reasoning, and late fusion of two separately propagated relevance scores consistently beats early or mid fusion. In the paper's reported experiments, MAG outperforms the strongest baseline on all eight benchmarks, with the largest margins on CLEVR (+30.5% over the best retrieval baseline), TextOCR (+25.1%), and MMStar (+7.1%).

Load-bearing premise

Everything rests on the generated image descriptions being faithful enough that textual similarity finds unlabeled samples worth labelling with provisional labels; if the descriptions are noisy or biased, the first stage can pick textually similar but visually irrelevant samples and the reported gain disappears.

Editorial extensions

If this is right

  • Under the paper's reported regime (15 labeled examples, 585 unlabeled, 45 pseudo-labeled), a deployment can improve multi-modal few-shot accuracy without finetuning, so the bottleneck shifts from labels to cheap MLLM calls for descriptions and pseudo-labels.
  • The largest reported gains on compositional and fine-grained reasoning benchmarks imply that graph-selected demonstrations matter most exactly where local similarity retrieval fails to cover the task's structure.
  • Random pseudo-labeling can hurt accuracy, so any adoption should keep relevance filtering; the pseudo-label budget is best spent on a compact, manifold-ranked subset rather than a large random sample.
  • Late fusion of separately propagated visual and textual relevance scores is the design that makes both modalities help; this recommends keeping modality-specific structure intact rather than concatenating features before graph construction.

Reading between the lines

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

  • An unstated consequence is that the same triage pattern could generalize beyond ICL: use a cheap single-modality manifold to filter a large unlabeled pool, then reserve expensive cross-modal scoring for a shortlist; the paper's near-linear scaling measurements make this pattern practical as pools grow.
  • The paper does not test whether the MLLM that writes the descriptions and the MLLM that consumes the demonstrations should be the same; using one generator may select images that match that model's perceptual biases, which could be either a benefit or a confirmation-bias risk.
  • A testable extension left open is an adaptive pseudo-label budget: instead of a fixed 45 samples, Stage 1 relevance scores could be thresholded, and the reported accuracy peak at a 25% labeled rate suggests such a rule might transfer across tasks.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes MAG, a training-free two-stage framework for multi-modal in-context learning under label scarcity. Stage 1 builds a text-only k-NN graph over labeled and unlabeled samples, propagates relevance scores from labeled to unlabeled nodes, and pseudo-labels the top-K unlabeled samples using the MLLM. Stage 2 constructs separate visual and textual graphs over the expanded candidate pool, propagates query-conditioned relevance scores in each modality, fuses them late, and selects the top-k demonstrations. Experiments on eight vision-language benchmarks compare MAG against retrieval, pseudo-labeling, and zero-shot baselines, report better mean accuracy across all eight datasets, and include ablations on pseudo-labeling strategy, demonstration selection, modality usage, fusion strategy, and hyperparameter sensitivity.

Significance. If the empirical claims hold, this is a useful contribution: it provides a parameter-free, closed-form way to exploit unlabeled multi-modal data for ICL without fine-tuning, with a sparse-graph formulation that is computationally attractive. The paper ships a clearly specified algorithm, reasonable ablations, and a scalability analysis. However, the current validation has several load-bearing gaps: no explicit statement that the unlabeled pool and test set are disjoint, no significance testing on the headline gains, an internal inconsistency in the pseudo-label accuracy results on Emotion6, and hyperparameter choices informed by the same test sets used for the final numbers. These issues must be resolved before the central claim is credible.

major comments (5)
  1. [4.2 Implementation Details] The experimental setup never states that the 585-sample unlabeled pool is disjoint from the 300 test samples; because Stage 1 (Eq. 4) pseudo-labels the top-K unlabeled samples and Stage 2 (Eq. 8) selects demonstrations from DE, any overlap between DU and the test set would directly leak ground-truth labels into the prompt and could fully explain the reported gains. The authors must explicitly specify the data split construction (e.g., unlabeled samples drawn from the train split and test samples from the test split with no shared images) and, ideally, release the exact split indices.
  2. [Table 1 (main results)] The abstract and Section 4.3 claim that MAG 'significantly' outperforms baselines, but with only 3 runs and no significance tests, many margins are within one standard deviation (e.g., EmoSet 76.3±1.4 vs. CVR-LLM 74.8±1.5; Emotion6 64.3±1.4 vs. Few-shot 63.0±1.9; GQA 59.0±3.9 vs. VICL 55.1±1.0; OKVQA 57.6±2.1 vs. VICL 56.3±1.5). Please report per-example paired tests or bootstrap confidence intervals across the three runs, and temper the 'significant' language accordingly.
  3. [Appendix B.3, Table 7] On Emotion6 the manifold-guided pseudo-labeling gives lower pseudo-label accuracy (0.400) and slightly lower final accuracy (0.650) than random selection (0.556 and 0.657), which directly contradicts the paragraph's claim of a 'strong positive correlation between pseudo-label accuracy and final performance' and the main-text claim that relevance-guided pseudo-labeling 'significantly improves performance across all benchmarks.' The authors should explain why Emotion6 is an exception or qualify the universal claim.
  4. [Section 4.7, Figure 3] The hyperparameters (α, β, knn, pool size, label rate, demo count) are fixed after inspecting performance on EmoSet, Emotion6, and MMStar, which appear to be the same test sets used for the final evaluation in Table 1; this selection-on-test-set circularity can inflate the reported numbers on those three datasets. Please tune on held-out validation splits (or report sensitivity computed on validation folds) and state the split protocol explicitly.
  5. [Section 3.1 (Textual Description Generation)] Stage 1 assumes that MLLM-generated descriptions faithfully represent the visual content needed for label propagation, but the paper provides no measure of description quality (e.g., human agreement or a reference-based metric) and no robustness analysis under corrupted or noisy descriptions. Because Eqs. (1) and (4) build the entire Stage-1 graph and pseudo-label selection on these descriptions, this assumption is load-bearing and should be validated.
minor comments (5)
  1. [Section 4.3] The heading 'Quantitave Evaluations' contains a typo; it should be 'Quantitative Evaluations.'
  2. [Appendix A, Algorithm 1] Line 11 says to compute relevance scores 'via Eq. (4)', but Eq. (4) is the pseudo-labeling equation; the relevance scores are computed by the closed-form solution in Eq. (3), so the reference should be corrected.
  3. [Section 3.1] The text encoder is named 'Contriver' but the reference is Contriever (Izacard et al., 2021); please correct the spelling throughout.
  4. [Tables 3 and 4] Tables 3 and 4 report accuracy without standard deviations, unlike Tables 1 and 2, making it difficult to judge whether the modality and fusion differences are meaningful; please include variance information.
  5. [Figure 3 caption] The caption should state which split is used for the sensitivity analysis; if it is the test split, this reinforces the concern in major comment 4.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the two-stage propagation pipeline derives demonstration selection from label- and query-conditioned diffusion on external-encoder graphs, and the reported gains are empirical rather than forced by construction.

full rationale

None of MAG's load-bearing derivations reduces to its inputs by construction. Stage 1 (Eqs. 1-4) is standard relevance propagation on a textual k-NN graph using external encoders (Contriever, CLIP ViT-L/14), with the closed-form solution cited to Zhu et al. and Zhou et al.; the pseudo-labeled set DP is selected by propagated relevance score, not by the target accuracy, and Eq. 4 invokes the MLLM only to produce labels after selection. Stage 2 (Eqs. 5-8) builds separate visual and textual graphs over DE, propagates query-conditioned scores, and late-fuses them; Eq. 9 then predicts by calling the MLLM on the selected demonstrations. The final accuracy numbers are empirical outcomes, not algebraic consequences of the selection equations, and the central comparison (MAG vs. baselines sharing the same backbone and pool) does not depend on any equation whose output is identical to its input. The graph-propagation machinery is cited to external prior work (Zhu et al. 2002, Zhou et al. 2004), and the method's relation to MAPLE is a contrast, not an imported conclusion; MAPLE's authors do not overlap with the present paper, so there is no self-citation chain that carries the argument. Two protocol concerns should be noted but they are not circularity: the hyperparameters (alpha, knn, beta, pool size, label rate, demo count) are selected from sensitivity analysis on EmoSet, Emotion6, and MMStar (Fig. 3) and then used in the main results, and Section 4.2 says only 'we randomly sample 15 labeled examples, 585 unlabeled examples as the candidate pool, and 300 test samples' without an explicit disjointness guarantee between the unlabeled pool and the test set. Both are potential evaluation-contamination or leakage risks, not instances where a 'prediction' is equivalent to a fitted parameter or to the input by construction. Likewise, using Gemini-2.0-Flash to generate descriptions and pseudo-labels and then evaluating with Gemini-2.0-Flash creates a same-model coupling, but that is a methodological confound, not a derivation that assumes its own conclusion. Overall, the claimed chain—relevance propagation selects candidates, pseudo-labeling expands the pool, and query-conditioned multimodal propagation selects demonstrations—is self-contained and externally benchmarked, so the circularity score is 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method is not a derivation and introduces no new physical or mathematical entities. It relies on standard label propagation (closed-form solution), the classic manifold assumption, and two domain-specific assumptions: that MLLM-generated descriptions faithfully represent images for graph construction, and that the unlabeled pool is representative of and disjoint from the test queries. Five hyperparameters are fixed across experiments, chosen via sensitivity analysis rather than fit to per-dataset labels; they are listed as free parameters.

free parameters (5)
  • alpha (propagation strength) = 0.9
    Fixed across experiments; chosen from sensitivity analysis in Figure 3.
  • k (nearest neighbors) = 5
    Fixed across experiments; chosen from sensitivity analysis in Figure 3.
  • beta (late fusion weight) = 0.5
    Fixed across experiments; chosen from sensitivity analysis in Figure 3.
  • K (number of pseudo-labeled samples) = 45
    Forms candidate pool of 60 with 15 labeled samples, matching label rate of 25 percent.
  • Number of demonstrations |DS(q)| = 10
    Fixed; sensitivity analysis shows performance plateaus at 10 demonstrations.
assumptions (4)
  • standard math Label propagation converges to the closed-form solution (1-alpha)(I-alpha W_hat)^(-1) I_0
    Standard result from Zhu et al. 2003 and Zhou et al. 2004; used in Eq. 3 and Eq. 6.
  • domain assumption k-NN graphs with cosine similarity and label propagation capture task-relevant semantic structure, i.e., nearby nodes share labels
    Classic label propagation assumption; invoked in Sections 3.1 and 3.2.
  • ad hoc to paper MLLM-generated textual descriptions faithfully represent visual content for graph construction
    Stage 1 and Stage 2 textual graphs use descriptions generated by Gemini-2.0-Flash; no error analysis of descriptions is provided.
  • domain assumption Unlabeled pool and test set are independent samples from the same distribution and are disjoint
    Section 4.2 samples 15 labeled, 585 unlabeled, and 300 test samples without stating disjointness.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MAG: MAnifold Guided Semi-Supervised Multi-modal In-Context Learning." pith.science (2026). https://pith.science/paper/2QEY2XHY

@misc{pith2026260812724,
  author       = {Pith},
  title        = {Pith review of: MAG: MAnifold Guided Semi-Supervised Multi-modal In-Context Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2QEY2XHY}},
  note         = {Machine review of arXiv:2608.12724}
}
read the original abstract

Few-shot in-context learning (ICL) with multi-modal large language models (MLLMs) enables task adaptation without parameter updates, but its performance is highly sensitive to the quality and coverage of the selected demonstrations. While unlabeled multi-modal data is abundant, it remains elusive how to exploit them for ICL. We propose MAG (MAnifold-Guided semi-supervised in-context demonstra- tion selection), an efficient framework that leverages unlabeled data to improve multi-modal ICL. MAG formulates demonstration selection as a semi-supervised propagation problem on a multi-modal graph and adopts a two-stage strategy: (i) relevance score propagation identifies a compact set of high-impact unlabeled samples for pseudo-labeling, reducing MLLM inference cost; (ii) multi-modal relevance is used to select the final demonstrations. We show that textual represen- tations are more effective for relevance propagation, while both visual and textual modalities are crucial for high-quality demonstration selection. Experiments on eight multi-modal benchmarks demonstrate that MAG consistently outperforms strong baselines in label-scarce regimes, achieving significant gains with a limited pseudo-labeling budget.

Figures

Figures reproduced from arXiv: 2608.12724 by the authors.

Figure 1
Figure 1. Comparison of our proposed semi-supervised [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of MAG. Stage 1 (top): We generate textual descriptions of all images using [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Sensitivity analysis across three datasets (EmoSet, Emotion6, MMStar). Performance is [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison between MAG and baselines (VICL, MMICES, CVR-LLM) across [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]
Figure 5
Figure 5. Figure 5: Visualization of the full MAG method across EmoSet, GQA, MMStar, and OCR tasks. [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: MAG with random Stage 1 pseudo-label selection (vs. relevance-guided). Random selection [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: MAG with random Stage 2 demonstration selection (vs. graph-based). Random selection [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 37 canonical work pages

  1. [1]

    Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks , author=

    work page

  2. [2]

    How do in-context examples affect compositional generalization? , author=

    work page

  3. [3]

    Self-adaptive in-context learning: An information compression perspective for in-context example selection and ordering , author=

    work page

  4. [4]

    Diverse demonstrations improve in-context compositional generalization , author=

    work page

  5. [5]

    MAPLE: Many-Shot Adaptive Pseudo-Labeling for In-Context Learning , author=

    work page

  6. [6]

    Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning , author=

    work page

  7. [7]

    Compositional exemplars for in-context learning , author=

    work page

  8. [8]

    How to configure good in-context sequence for visual question answering , author=

    work page

Show all 54 references
  1. [9]

    Can Multimodal Large Language Models Truly Perform Multimodal In-Context Learning? , author=

  2. [10]

    arXiv preprint arXiv:2402.11574 , year=

    Visual in-context learning for large vision-language models , author=. arXiv preprint arXiv:2402.11574 , year=

    arXiv

  3. [11]

    In-context compositional generalization for large vision-language models , author=

  4. [12]

    Large language models are visual reasoning coordinators , author=

  5. [13]

    arXiv preprint arXiv:2508.12400 , year=

    MPCAR: Multi-Perspective Contextual Augmentation for Enhanced Visual Reasoning in Large Vision-Language Models , author=. arXiv preprint arXiv:2508.12400 , year=

    arXiv

  6. [14]

    Enhancing large vision language models with self-training on image comprehension , author=

  7. [15]

    Taco: Enhancing multimodal in-context learning via task mapping-guided sequence configuration , author=

  8. [16]

    Enhancing advanced visual reasoning ability of large language models , author=

  9. [17]

    IEEE conference on computer vision and pattern recognition , year=

    Efficient diffusion on region manifolds: Recovering small objects with compact cnn representations , author=. IEEE conference on computer vision and pattern recognition , year=

  10. [18]

    Hypergraph label propagation network , author=

  11. [19]

    ACM Web Conference , year=

    VilLain: Self-supervised learning on homogeneous hypergraphs without features via virtual label propagation , author=. ACM Web Conference , year=

  12. [20]

    From Cross-Task Examples to In-Task Prompts: A Graph-Based Pseudo-Labeling Framework for In-context Learning , author=

  13. [21]

    Language models are few-shot learners , author=

  14. [22]

    arXiv preprint arXiv:2209.12892 , year=

    Learning to learn with generative models of neural network checkpoints , author=. arXiv preprint arXiv:2209.12892 , year=

    arXiv

  15. [23]

    What can transformers learn in-context? a case study of simple function classes , author=

  16. [24]

    arXiv preprint arXiv:2111.02080 , year=

    An explanation of in-context learning as implicit bayesian inference , author=. arXiv preprint arXiv:2111.02080 , year=

    arXiv

  17. [25]

    Fantastically Ordered Prompts and Where to Find Them: Overcoming Few-Shot Prompt Order Sensitivity , author=

  18. [26]

    Learning To Retrieve Prompts for In-context Learning , author=

  19. [27]

    Large language models are human-level prompt engineers , author=

  20. [28]

    gradient descent

    Automatic prompt optimization with" gradient descent" and beam search , author=

  21. [29]

    Learning Transferable Visual Models From Natural Language Supervision , author=

  22. [30]

    Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision , author=

  23. [31]

    BLIP: Bootstrapped Language-Image Pre-training for Unified Vision-Language Understanding and Generation , author=

  24. [32]

    Flamingo: a Visual Language Model for Few-Shot Learning , author=

  25. [33]

    arXiv preprint arXiv:2303.08774 , year=

    GPT-4 Technical Report , author=. arXiv preprint arXiv:2303.08774 , year=

    arXiv

  26. [34]

    Visual Instruction Tuning , author=

  27. [35]

    Learning from Labeled and Unlabeled Data with Label Propagation , author=

  28. [36]

    Learning with Local and Global Consistency , author=

  29. [37]

    Label Propagation for Deep Semi-supervised Learning , author=

  30. [38]

    Pseudo-label guided contrastive learning for semi-supervised medical image segmentation , author=

  31. [39]

    Simmatchv2: Semi-supervised learning with graph consistency , author=

  32. [40]

    Provoking Multi-modal Few-Shot LVLM via Exploration-Exploitation In-Context Learning , author=

  33. [41]

    Semi-supervised learning using gaussian fields and harmonic functions , author=

  34. [42]

    Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results , author=

  35. [43]

    Fixmatch: Simplifying semi-supervised learning with consistency and confidence , author=

  36. [44]

    A Large-scale Visual Emotion Dataset with Rich Attributes , author=

  37. [45]

    Contemplating Visual Emotions: Understanding and Overcoming Dataset Bias , author=

  38. [46]

    arXiv preprint arXiv:2403.13164 , year=

    VL-ICL Bench: The Devil in the Details of Multimodal In-Context Learning , author=. arXiv preprint arXiv:2403.13164 , year=

    arXiv

  39. [47]

    Are We on the Right Way for Evaluating Large Vision-Language Models? , author=

  40. [48]

    Gqa: A new dataset for real-world visual reasoning and compositional question answering , author=

  41. [49]

    Kenneth Marino and Mohammad Rastegari and Ali Farhadi and Roozbeh Mottaghi , title =

  42. [50]

    Making the

    Yash Goyal and Tejas Khot and Douglas Summers. Making the

  43. [51]

    arXiv preprint arXiv:2507.06261 , year=

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities , author=. arXiv preprint arXiv:2507.06261 , year=

    arXiv

  44. [52]

    arXiv preprint arXiv:2112.09118 , year=

    Unsupervised dense information retrieval with contrastive learning , author=. arXiv preprint arXiv:2112.09118 , year=

    arXiv

  45. [53]

    Learning transferable visual models from natural language supervision , author=

  46. [54]

    arXiv preprint arXiv:2409.12191 , year=

    Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution , author=. arXiv preprint arXiv:2409.12191 , year=

    arXiv

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.