Pith. sign in

REVIEW 3 major objections 6 minor 30 references

Multi-Sourced Compositional Generalization in Visual Question Answering

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read VQA models fail on novel compositions whose parts come from different modalities; retrieval-augmented training that aligns cross-modal primitive representations improves this ability without sacrificing standard accuracy.

desk verdict A useful new benchmark and a plausible method, undercut by test-set hyperparameter selection and no error bars; the headline gain needs a cleaner evaluation. read the letter →

arxiv 2505.23045 v1 pith:7IHO7F3G submitted 2025-05-29 cs.CV cs.AI

Chuanhao Li , Wenbo Ye , Zhen Li , Yuwei Wu , Yunde Jia This is my paper
classification cs.CVcs.AI
keywords compositionalgeneralizationvisualquestionansweringretrieval-augmentedtrainingcross-modalalignmentmulti-sourcednovelcompositionsGQA-MSCGvisionandlanguageprimitiverepresentation
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

This paper asks whether vision-and-language models can recombine primitives they have seen separately when those primitives come from different modalities, such as the word 'white' and a pictured dog. It argues that current VQA models struggle with these multi-sourced novel compositions, and that the underlying weakness is poor alignment between linguistic and visual primitive representations. The proposed fix is retrieval-augmented training: during training, each primitive feature is replaced by itself plus weighted averages of semantically similar primitive features retrieved from per-modality databases built only from the training set, nudging the model to represent the same meaning consistently across modalities. To measure this, the authors construct GQA-MSCG, seven test splits from GQA containing unseen pairs of primitives, and show that the framework raises accuracy for a small VQA model (from 72.41 to 74.21 overall) and a large multimodal model (from 68.49 to 71.53 overall) while preserving or slightly improving IID accuracy. If correct, the paper establishes multi-sourced compositional generalization as a distinct, measurable failure mode and offers a training-side remedy that does not change the loss function or require external data.

What carries the argument

The central mechanism is the retrieval-aggregation equation: for a primitive feature $p$, the aggregated feature is $p_a = p + \frac{w_q}{K_q}\sum_{i=1}^{K_q}\cos(p,p_i^{(q)}) + \frac{w_v}{K_v}\sum_{i=1}^{K_v}\cos(p,p_i^{(v)})$, where $p_i^{(q)}$ and $p_i^{(v)}$ are the top-$K$ retrieved features from the linguistic and visual primitive databases $\mathcal{D}_q$ and $\mathcal{D}_v$, and $w_q, w_v$ control each modality's contribution. The aggregated feature replaces the original primitive feature during training, so the feature extractor is continuously refined to produce context-invariant, modality-consistent primitive representations. The benchmark machinery pairs every primitive in a sample, labels a pair as novel if that pair never occurs in the training set, and classifies novel pairs into LL, VV, or LV depending on whether the two primitives come from language, vision, or one of each.

What would settle it

Hold out a random portion of GQA-MSCG as validation, select $w_q$ and $w_v$ using only that validation accuracy, then report accuracy on the remaining test samples; if the retrieval-augmented model no longer beats its baseline across all seven splits under this protocol, or if the gains become small or negative, the central claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that VQA models can be made to generalize to novel compositions of primitives drawn from different modalities by aligning primitive representations at training time through retrieval. Concretely, for each word-level or object-level primitive in a training sample, the framework retrieves the top-K most cosine-similar primitive features from a linguistic primitive database and a visual primitive database, both built solely from the training set, and replaces the original feature with a weighted sum of itself and the two retrieval aggregates. Because each database stores the same primitive in many contexts, the aggregated feature pulls together representations of semantically equivalent primitives across modalities and contexts. On the constructed GQA-MSCG benchmark, the framework improves overall accuracy from 72.41 to 74.21 for the CFR baseline and from 68.49 to 71.53 for Qwen-VL across all seven test splits, and it slightly improves IID accuracy on GQA test-dev and VQA v2 val. The paper also claims that difficulty grows with the number of novel-composition types co-occurring in a sample (Level-1 to Level-3), and that existing models degrade accordingly.

Load-bearing premise

The reported gains depend on the mixture weights $w_q=0.6$ and $w_v=0.4$ used to combine retrieved linguistic and visual features, and those weights were chosen by inspecting GQA-MSCG test accuracy; if that choice carries hidden information about the test set, the improvement would not transfer to new data.

Editorial extensions

If this is right

  • Across the seven GQA-MSCG splits, the framework raises accuracy for both tested baselines on every split, with Level-1 gains from roughly +1.5 to +3.4 points and Level-3 gains from roughly +1.4 to +3.6 points.
  • Because the method adds no auxiliary loss and builds its databases from the training split, it can be inserted into existing VQA models without changing the training objective or introducing external data.
  • The framework preserves or slightly improves IID generalization, with CFR gaining +1.43 points and Qwen-VL gaining +1.13 points on GQA test-dev, so the compositional improvement is not bought by sacrificing ordinary accuracy.
  • Linguistic-only and visual-only retrieval each help most on their own modality's novel compositions, while joint retrieval gives the best overall accuracy, indicating the two databases are complementary.
  • Existing VQA models show inconsistent abilities across LL, VV, and LV splits, so a model that is strong on one type of novel composition is not necessarily strong on another; the framework pushes all three in the same direction.

Reading between the lines

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

  • The retrieval-aggregation step acts as a form of train-time feature smoothing; if the reported gains replicate, it suggests that representation alignment across modalities, rather than model capacity, is a primary bottleneck for multi-sourced compositional generalization in VQA.
  • Because novelty is defined as an unseen pair of recorded primitives, the benchmark does not control for synonyms or paraphrases; a harder variant that swaps synonyms for seen words would test whether the framework aligns semantics or merely surface forms.
  • The visual primitive database relies on GQA's scene-graph annotations, so an extension worth testing is whether the same retrieval scheme works with detector-free or caption-derived visual primitives, which would determine if the gains survive without oracle object labels.
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.

Referee Report

3 major / 6 minor

Summary. The paper introduces the task of multi-sourced compositional generalization (MSCG) for VQA, constructs a new benchmark GQA-MSCG with seven test splits corresponding to different combinations of novel composition types (LL, VV, LV and their co-occurrences), and proposes a retrieval-augmented training framework (RAG) that aggregates each primitive feature with retrieved features of semantically equivalent primitives from linguistic and visual databases built from the training set. Experiments on CFR and Qwen-VL report accuracy gains on GQA-MSCG as well as on IID splits of GQA and VQA v2.

Significance. If the empirical claims were validated, the paper would make a useful contribution: the GQA-MSCG benchmark is well-motivated and carefully constructed (primitives are required to be seen while compositions are novel, and the multi-level splits are a nice design), and the framework is simple, model-agnostic, and does not add training losses. The release of the dataset is a concrete positive. However, the core empirical claim is currently compromised by the evaluation protocol: hyperparameters are selected on the test set, and no uncertainty estimates are provided. The significance of the contribution therefore depends on a corrected evaluation.

major comments (3)
  1. [Section 5.5 and Section 5.1 (Implementation Details)] The aggregation weights w_q=0.6 and w_v=0.4 are selected by maximizing accuracy on the GQA-MSCG dataset itself, as shown in Figure 5, and the final results in Table 1 are reported on the same set. This constitutes test-set model selection, so the reported improvements (e.g., +1.8 points for CFR+RAG on Overall) are optimistically biased. Please select hyperparameters on a held-out validation split of GQA-MSCG (or via cross-validation) and report final results on a separate test split, or explicitly justify why this selection does not bias the conclusion.
  2. [Tables 1–3] No standard deviations, confidence intervals, or multiple-seed runs are reported. Given that Figure 5 shows accuracy fluctuations of roughly 1–2 points across neighboring hyperparameter values, the reported gains of about 1–2 points cannot be distinguished from selection noise. Please report mean and standard deviation over at least three independent runs with different seeds for all main tables.
  3. [Section 3.3, Eq. (1)] The aggregation formula is dimensionally inconsistent: p_a is a feature vector, while the terms (1/K_q)Σ cos(p, p_q^(i)) and (1/K_v)Σ cos(p, p_v^(i)) are scalars. As written, the equation adds scalars to a vector, which is not a valid operation. The intended aggregation (e.g., a weighted sum of feature vectors with cosine-similarity weights, or a gated combination) must be specified precisely, because this operation is the core of the framework.
minor comments (6)
  1. [Figures 3 and 4] The figures use the labels 'QQ' and 'QV' for what the text calls 'LL' and 'LV' (e.g., Figure 3(a) 'QQ split' vs. Section 4 'LL'). Please unify the notation.
  2. [Section 5.3] The citations for LLaVA-1.5 and LLaVA-1.6 are swapped relative to the reference list; LLaVA-1.5 is [Liu et al., 2023] and LLaVA-1.6 is [Liu et al., 2024a].
  3. [Section 5.5] There are two occurrences of the typo 'CFG+RAG' that should read 'CFR+RAG'.
  4. [Abstract and Section 4] The term 'multi-sourced' is used to describe all three novel composition types, but LL and VV compositions are composed of primitives from a single modality; consider clarifying that the framework addresses compositions whose primitives can originate from either modality, not that each composition must be cross-modal.
  5. [Section 3.3] It is not specified how visual primitives are localized in the image when the visual encoder produces patch-level features (as in Qwen-VL); please clarify the procedure used for matching a primitive to its corresponding feature.
  6. [Section 4] The dataset construction samples 5,000 samples per category, yielding 35,000 test samples; it would be helpful to report the number of candidate samples in D_c before sampling to inform the reader of the sampling ratio.

Circularity Check

1 steps flagged · score 6.0 of 10

GQA-MSCG headline gain is partially a fitted result: aggregation weights w_q/w_v are selected on the target test set (Fig. 5) and then used to report the final Table 1 accuracies on that same test set.

  1. fitted input called prediction [Section 5.5 Parameter Analysis; Implementation Details; Eq. (1); Table 1]
    "CFR+RAG performs best with setting w_q = 0.6 and w_v = 0.4 simultaneously. Based on the above experimental results, we set w_q to 0.6 and w_v to 0.4 for all experiments."

    The aggregation weights w_q and w_v are hyperparameters in Eq. (1) that control how much retrieved linguistic and visual primitive features contribute to training. Section 5.5 chooses these weights by sweeping them on the GQA-MSCG dataset (Figure 5) and taking the configuration that maximizes accuracy on that dataset. The final results in Table 1, which are used to claim that the framework 'significantly improves' multi-sourced compositional generalization, are then reported on the very same GQA-MSCG test splits. This is not an unbiased evaluation: the reported MSCG accuracy is partly an output of test-set model selection.

full rationale

The paper's core retrieval-augmented training idea is not circular in the sense of being defined in terms of its own output, and there is no load-bearing self-citation chain or imported uniqueness theorem. However, the primary empirical claim about MSCG generalization is weakened by a fitted-input problem: the hyperparameters w_q and w_v are explicitly tuned on the GQA-MSCG test set (Figure 5) and then the same test set is used to produce the headline numbers in Table 1. Because the paper reports no error bars or held-out validation for these weights, the reported improvements are partly a consequence of test-set selection rather than a genuine prediction of generalization. The IID results on GQA test-dev and VQA v2 val provide some independent evidence that the framework does not hurt IID performance, which keeps the paper from collapsing entirely into circularity. But the main contribution as evaluated on GQA-MSCG does not support the strong 'significantly improves' claim without a proper validation protocol, so a moderate circularity score is warranted.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The framework introduces no new physical entities. The free parameters are the retrieval and aggregation hyperparameters, several of which are tuned on the target test set. The axioms are standard assumptions about GQA annotations, NLTK reliability, feature similarity semantics, data leakage, and baseline reimplementation.

free parameters (6)
  • w_q = 0.6
    Aggregation weight for linguistic primitive database. Chosen based on GQA-MSCG test set accuracy (Section 5.5, Figure 5).
  • w_v = 0.4
    Aggregation weight for visual primitive database. Chosen based on GQA-MSCG test set accuracy (Section 5.5, Figure 5).
  • T_q = 8 (1 for VQA v2)
    Number of sampled questions per linguistic primitive in the database. Hyperparameter set by the authors.
  • T_v = 32
    Number of sampled images per visual primitive in the database. Hyperparameter set by the authors.
  • K_q = 4
    Number of retrieved linguistic primitive features per primitive. Hyperparameter set by the authors.
  • K_v = 16 (4 for VQA v2)
    Number of retrieved visual primitive features per primitive. Hyperparameter set by the authors.
assumptions (5)
  • domain assumption GQA scene graph annotations accurately identify object categories and attributes for visual primitives.
    Section 3.2 uses scene graphs to construct the visual primitive database; if these annotations are noisy, the alignment of visual primitives would be unreliable.
  • domain assumption NLTK lemmatization and part-of-speech tagging reliably extract linguistic primitives from questions.
    Section 3.2 uses NLTK to define linguistic primitives; errors in lemmatization or POS tagging would affect both the database and the dataset construction.
  • domain assumption Cosine similarity between feature vectors is a valid measure of semantic similarity between primitives.
    Equation (1) uses cosine similarity for retrieval and aggregation; if feature similarity does not align with semantic equivalence, the aggregation may be harmful.
  • domain assumption The retrieval databases built from the GQA train split do not leak test-set information.
    The databases are constructed only from D_t, and test samples come from the val split; the paper states this explicitly to avoid external data, but the assumption is critical for the benchmark's validity.
  • domain assumption The reimplemented baseline models (CFR, Qwen-VL) match the official code behavior.
    Section 5.1 says the baselines are reimplemented based on official code; any discrepancy could affect the reported comparisons.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Sourced Compositional Generalization in Visual Question Answering." pith.science (2026). https://pith.science/paper/7IHO7F3G

@misc{pith2026250523045,
  author       = {Pith},
  title        = {Pith review of: Multi-Sourced Compositional Generalization in Visual Question Answering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7IHO7F3G}},
  note         = {Machine review of arXiv:2505.23045}
}
read the original abstract

Compositional generalization is the ability of generalizing novel compositions from seen primitives, and has received much attention in vision-and-language (V\&L) recently. Due to the multi-modal nature of V\&L tasks, the primitives composing compositions source from different modalities, resulting in multi-sourced novel compositions. However, the generalization ability over multi-sourced novel compositions, \textit{i.e.}, multi-sourced compositional generalization (MSCG) remains unexplored. In this paper, we explore MSCG in the context of visual question answering (VQA), and propose a retrieval-augmented training framework to enhance the MSCG ability of VQA models by learning unified representations for primitives from different modalities. Specifically, semantically equivalent primitives are retrieved for each primitive in the training samples, and the retrieved features are aggregated with the original primitive to refine the model. This process helps the model learn consistent representations for the same semantic primitives across different modalities. To evaluate the MSCG ability of VQA models, we construct a new GQA-MSCG dataset based on the GQA dataset, in which samples include three types of novel compositions composed of primitives from different modalities. Experimental results demonstrate the effectiveness of the proposed framework. We release GQA-MSCG at https://github.com/NeverMoreLCH/MSCG.

Figures

Figures reproduced from arXiv: 2505.23045 by the authors.

Figure 1
Figure 1. Multi-sourced novel compositions in the context of VQA. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The overall framework of the proposed framework. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Level-1 samples in the GQA-MSCG dataset. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Level-2 samples and Level-3 samples in the GQA-MSCG [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Parameter analysis using CFR as the baseline model on the GQA-MSCG dataset. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: The qualitative comparison between CFR+RAG (Ours) [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 26 canonical work pages

  1. [1]

    Robust visual rea- soning via language guided neural module networks

    [Akulaet al., 2021 ] Arjun Akula, Varun Jampani, Soravit Changpinyo, and Song-Chun Zhu. Robust visual rea- soning via language guided neural module networks. InAdvances in Neural Information Processing Systems (NeurIPS), pages 11041–11053,

    work page 2021

  2. [5]

    Qwen-vl: A frontier large vision- language model with versatile abilities.arXiv preprint arXiv:2308.12966,

    [Baiet al., 2023 ] Jinze Bai, Shuai Bai, Shusheng Yang, Shi- jie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision- language model with versatile abilities.arXiv preprint arXiv:2308.12966,

    arXiv 2023

  3. [9]

    The paradox of the compositionality of natural language: A neural machine translation case study

    [Dankerset al., 2022 ] Verna Dankers, Elia Bruni, and Dieuwke Hupkes. The paradox of the compositionality of natural language: A neural machine translation case study. InProceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), pages 4154–4175,

    work page 2022

  4. [13]

    Compositional attention networks for machine reasoning

    [Hudson and Manning, 2018] Drew A Hudson and Christo- pher D Manning. Compositional attention networks for machine reasoning. InProceedings of the International Conference on Learning Representations (ICLR),

    work page 2018

  5. [14]

    Gqa: A new dataset for real-world vi- sual reasoning and compositional question answering

    [Hudson and Manning, 2019] Drew A Hudson and Christo- pher D Manning. Gqa: A new dataset for real-world vi- sual reasoning and compositional question answering. In Proceedings of IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pages 6700–6709,

    work page 2019

  6. [15]

    Overcoming language pri- ors in vqa via decomposed linguistic representations

    [Jinget al., 2020 ] Chenchen Jing, Yuwei Wu, Xiaoxun Zhang, Yunde Jia, and Qi Wu. Overcoming language pri- ors in vqa via decomposed linguistic representations. In Proceedings of the AAAI Conference on Artificial Intelli- gence (AAAI), volume 34, pages 11181–11188,

    work page 2020

  7. [17]

    On compositional generalization of neural ma- chine translation

    [Liet al., 2021 ] Yafu Li, Yongjing Yin, Yulong Chen, and Yue Zhang. On compositional generalization of neural ma- chine translation. InProceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), pages 4767–4780,

    work page 2021

  8. [18]

    Compositional temporal ground- ing with structured variational cross-graph correspondence learning

    [Liet al., 2022 ] Juncheng Li, Junlin Xie, Long Qian, Lin- chao Zhu, Siliang Tang, Fei Wu, Yi Yang, Yueting Zhuang, and Xin Eric Wang. Compositional temporal ground- ing with structured variational cross-graph correspondence learning. InProceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3032–3041,

    work page 2022

Show all 30 references
  1. [19]

    Blip-2: Bootstrapping language-image pre- training with frozen image encoders and large language models.arXiv preprint arXiv:2301.12597,

    [Liet al., 2023c ] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre- training with frozen image encoders and large language models.arXiv preprint arXiv:2301.12597,

    arXiv

  2. [20]

    Improved baselines with visual instruction tuning,

    [Liuet al., 2023 ] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning,

    2023

  3. [21]

    Overcoming language priors in visual question answering with cumulative learning strategy.Neurocom- puting, 608:128419,

    [Maoet al., 2024 ] Aihua Mao, Feng Chen, Ziying Ma, and Ken Lin. Overcoming language priors in visual question answering with cumulative learning strategy.Neurocom- puting, 608:128419,

    2024

  4. [22]

    Learning graph embeddings for compositional zero-shot learning

    [Naeemet al., 2021 ] Muhammad Ferjad Naeem, Yongqin Xian, Federico Tombari, and Zeynep Akata. Learning graph embeddings for compositional zero-shot learning. InProceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 953–962,

    2021

  5. [23]

    Coarse- to-fine reasoning for visual question answering

    [Nguyenet al., 2022 ] Binh X Nguyen, Tuong Do, Huy Tran, Erman Tjiputra, Quang D Tran, and Anh Nguyen. Coarse- to-fine reasoning for visual question answering. pages 4557–4565,

    2022

  6. [24]

    Counter- factual vqa: A cause-effect look at language bias

    [Niuet al., 2021 ] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xian-Sheng Hua, and Ji-Rong Wen. Counter- factual vqa: A cause-effect look at language bias. InPro- ceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12700–12710,

    2021

  7. [25]

    Combine to describe: Evaluating compositional generalization in image captioning

    [Pantazopouloset al., 2022 ] George Pantazopoulos, Alessandro Suglia, and Arash Eshghi. Combine to describe: Evaluating compositional generalization in image captioning. InProceedings of the Association for Computational Linguistics: Student Research Workshop (ACLSRW), pages 115–131,

    2022

  8. [26]

    Overcom- ing language priors for visual question answering based on knowledge distillation

    [Peng and Wei, 2024] Daowan Peng and Wei Wei. Overcom- ing language priors for visual question answering based on knowledge distillation. InProceedings of the IEEE In- ternational Conference on Multimedia and Expo (ICME), pages 1–6. IEEE,

    2024

  9. [27]

    Faster r-cnn: Towards real-time ob- ject detection with region proposal networks.IEEE Trans- actions on Pattern Analysis and Machine Intelligence (T- PAMI), 39(6):1137–1149,

    [Renet al., 2016 ] Shaoqing Ren, Kaiming He, Ross Gir- shick, and Jian Sun. Faster r-cnn: Towards real-time ob- ject detection with region proposal networks.IEEE Trans- actions on Pattern Analysis and Machine Intelligence (T- PAMI), 39(6):1137–1149,

    2016

  10. [29]

    Transformer module networks for systematic generaliza- tion in visual question answering.IEEE Transactions on Pattern Analysis and Machine Intelligence,

    [Yamadaet al., 2024 ] Moyuru Yamada, Vanessa D’Amario, Kentaro Takemoto, Xavier Boix, and Tomotake Sasaki. Transformer module networks for systematic generaliza- tion in visual question answering.IEEE Transactions on Pattern Analysis and Machine Intelligence,

    2024

  11. [30]

    Minigpt-4: Enhancing vision-language understanding with advanced large lan- guage models.arXiv preprint arXiv:2304.10592, 2023

    [Zhuet al., 2023 ] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xi- ang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large lan- guage models.arXiv preprint arXiv:2304.10592, 2023

    2023 arXiv

  12. [2009]

    Composi- tional generalization for multi-label text classification: A data-augmentation approach

    [Chaiet al., 2024 ] Yuyang Chai, Zhuang Li, Jiahui Liu, Lei Chen, Fei Li, Donghong Ji, and Chong Teng. Composi- tional generalization for multi-label text classification: A data-augmentation approach. InProceedings of the AAAI Conference on Artificial Intelligence (AAAI), volu...

    2024

  13. [2015]

    Systematic generalization: what is required and can it be learned?arXiv preprint arXiv:1811.12889,

    [Bahdanauet al., 2018 ] Dzmitry Bahdanau, Shikhar Murty, Michael Noukhovitch, Thien Huu Nguyen, Harm de Vries, and Aaron Courville. Systematic generalization: what is required and can it be learned?arXiv preprint arXiv:1811.12889,

    2018 arXiv

  14. [2016]

    Plug-and- play vqa: Zero-shot vqa by conjoining large pretrained models with zero training

    [Tionget al., 2022 ] Anthony Meng Huat Tiong, Junnan Li, Boyang Li, Silvio Savarese, and Steven CH Hoi. Plug-and- play vqa: Zero-shot vqa by conjoining large pretrained models with zero training. InFindings of the Association for Computational Linguistics: EMNLP 2022, pages 951– 967,

    2022

  15. [2017]

    Language-conditioned graph networks for relational reasoning

    [Huet al., 2019 ] Ronghang Hu, Anna Rohrbach, Trevor Darrell, and Kate Saenko. Language-conditioned graph networks for relational reasoning. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10294–10303,

    2019

  16. [2018]

    Vqa: Visual question answering

    [Antolet al., 2015 ] Stanislaw Antol, Aishwarya Agrawal, Ji- asen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zit- nick, and Devi Parikh. Vqa: Visual question answering. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2425–2433,

    2015

  17. [2019]

    LoRA: Low-rank adaptation of large language models

    [Huet al., 2022 ] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. InProceedings of the International Con- ference on Learning Representations (ICLR),

    2022

  18. [2020]

    Retrieval-augmented primitive representa- tions for compositional zero-shot learning

    [Jinget al., 2024 ] Chenchen Jing, Yukun Li, Hao Chen, and Chunhua Shen. Retrieval-augmented primitive representa- tions for compositional zero-shot learning. InProceedings of the AAAI Conference on Artificial Intelligence (AAAI), volume 38, pages 2652–2660,

    2024

  19. [2021]

    Bottom-up and top-down attention for im- age captioning and visual question answering

    [Andersonet al., 2018 ] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for im- age captioning and visual question answering. InProceed- ings of IEEE/CVF Conference on Computer Vision and Pa...

    2018

  20. [2022]

    Making the v in vqa matter: Elevating the role of image understanding in visual question answering

    [Goyalet al., 2017 ] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. InProceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition ...

    2017

  21. [2023]

    ” O’Reilly Media, Inc.”,

    [Birdet al., 2009 ] Steven Bird, Ewan Klein, and Edward Loper.Natural language processing with Python: ana- lyzing text with the natural language toolkit. ” O’Reilly Media, Inc.”,

    2009

  22. [2024]

    Meta module network for compositional visual reasoning

    [Chenet al., 2021 ] Wenhu Chen, Zhe Gan, Linjie Li, Yu Cheng, William Wang, and Jingjing Liu. Meta module network for compositional visual reasoning. InProceed- ings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 655–664,

    2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.