A Knowledge Noise Mitigation Framework for Knowledge-based Visual Question Answering

Jinyuan Liu; Sihang Liu; Xinru Zhang; Zhiyue Liu

REVIEW 3 major objections 5 minor 23 references

A Knowledge Noise Mitigation Framework for Knowledge-based Visual Question Answering

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Knowledge noise, not missing knowledge, is the main drag on KB-VQA, and a training-free filter fixes it.

desk verdict A solid training-free KB-VQA pipeline with a clean ablation, but a confidence-gate definition that is internally inconsistent and key hyperparameters tuned on the test set keep the headline numbers from being fully trusted. read the letter →

arxiv 2509.09159 v1 pith:7IHQJG24 submitted 2025-09-11 cs.CV cs.AI

Zhiyue Liu , Sihang Liu , Jinyuan Liu , Xinru Zhang This is my paper

classification cs.CVcs.AI

keywords KB-VQAknowledgenoiseretrieval-augmentedgenerationselectiveintegrationlargelanguagemodelsvisualtraining-freeOK-VQA

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that knowledge-based visual question answering (KB-VQA) fails less because of missing knowledge than because of noisy, redundant retrieved knowledge, and proposes a training-free framework, KF-VQA, to clean knowledge before it reaches the answering LLM. It first condenses image-question pairs into low-noise keyword queries for retrieval, then uses a VLM+LLM collaboration to extract only answer-relevant segments, and finally lets the LLM decide whether it needs external knowledge at all based on its own generation confidence. On OK-VQA and A-OKVQA, KF-VQA with Llama 3 8B reaches 63.2% and 60.9% test accuracy, respectively, surpassing both training-based and training-free state-of-the-art baselines. The main reason to care is that the method obtains these gains without any training or finetuning, making it immediately deployable with frozen models.

What carries the argument

The load-bearing mechanism is the selective knowledge integration gate in Equation (6): an exponential transform of the summed log-probabilities of the LLM's candidate answer, compared against a threshold τ=0.8. Below the threshold, the model receives h=7 filtered knowledge segments; above it, it answers from implicit knowledge alone. This gate is what converts knowledge retrieval from a fixed augmentation into a conditional intervention, and it is the component whose removal drops accuracy in the ablation study.

What would settle it

A direct check: compute accuracy on OK-VQA for questions where the LLM's confidence is high (s > 0.8) and compare answers with versus without injected knowledge; if adding filtered knowledge on this subset does not reduce accuracy, the selective gate is not doing the causal work claimed. A second check: re-tune τ on A-OKVQA validation and see whether the OK-VQA-tuned value still yields the reported test gains.

Watch

Extended reading notes

Core claim

The central discovery is that a three-stage noise-mitigation pipeline—low-noise retrieval queries, fine-grained knowledge-segment filtering, and confidence-gated knowledge integration—lets a frozen LLM outperform models that are trained or that use larger, more verbose retrieval. The confidence gate is the key: when the LLM is already confident, injecting even filtered external knowledge hurts accuracy, so the framework withholds knowledge in those cases. The paper demonstrates this with ablations: using the same retrieval and filtering but no gate gives 62.4% on OK-VQA, while the full gated framework reaches 63.2%.

Load-bearing premise

The framework assumes that the LLM's generation probability reliably indicates when external knowledge would help, and that a single threshold τ=0.8 set on OK-VQA transfers to other datasets; if the probability is miscalibrated, the gate either withholds useful knowledge or lets noisy knowledge through.

Editorial extensions

If this is right

KF-VQA improves over the strongest training-free baseline DKA by 1.1 points on OK-VQA and 1.0 points on A-OKVQA test, and over the best training-based method SKSQA by 0.4 points on OK-VQA.
The framework is training-free, so the same pipeline can be applied to any frozen LLM/VLM pair without finetuning.
The confidence gate implies that external knowledge is not unconditionally beneficial; models should receive knowledge only when their own confidence is low.
Filtering knowledge to a few segments shortens prompts, which the paper notes accelerates inference speed.

Reading between the lines

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

The confidence-gating principle could generalize to other retrieval-augmented generation tasks, but the threshold τ is tuned on OK-VQA; a calibrated or learned gate may transfer better across datasets.
The method's gains hinge on the VLM's keyword extraction being faithful; a testable extension would vary the VLM or prompt and measure retrieval precision directly, not just end accuracy.
Because generated token probabilities are used as confidence, the framework is vulnerable to LLM overconfidence; using entropy-based measures across multiple sampling passes might make the gate more robust.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Short version: this is a competent training-free KB-VQA paper with a new and sensible three-part recipe. The gains over DKA are small (~1 point) but consistent on two benchmarks, and the ablation shows each component does something. The paper should be reviewed, but only after the authors fix a genuine inconsistency in their confidence-gate definition and supply the missing prompts/code.

The new thing: using VLM-generated keywords to shorten retrieval queries, using an LLM conditioned on visual details to extract answer-beneficial segments, and gating knowledge injection on a model confidence score. None of these is novel in isolation, but the combination is not in the prior work I know, and the experiments are set up cleanly. Credit where due: the ablation study is straightforward and supports the contribution of each module. The qualitative examples are illustrative rather than cherry-picked-looking.

Now the soft spots. The most important is the confidence score. Eq. 6 defines s as the exponentiated sum of per-token log-probabilities (i.e., sequence likelihood). The caption of Eq. 7 defines s as e^{max_j p(aj|...)}. These are not even on the same scale; tau=0.8 cannot be applied consistently to both. The paper never says which one was implemented. That's not a minor typo, because the entire SKI component depends on this threshold, and SKI is the paper's stated conceptual contribution. Also, tau is tuned on OK-VQA itself (Fig. 3c) with no validation split; the same holds for r and h. The paper reports no error bars or significance tests, so the 1-point margins over DKA could be noise. No code, prompts (P1-P4) are not in the text, so I can't reproduce the numbers.

Importantly, though, the SKI component is not the whole story. In the ablation, LNQ and KRF together get to 62.4, already above DKA's 62.1. So even if the confidence gate is miscalibrated or overfit, the core retrieval-and-filtering pipeline still gives a small edge. That tempers the stress-test claim that the result is fragile end-to-end. Still, the paper's headline result uses the full model, and the definition problem is real.

Bottom line: this is a solid empirical contribution that deserves a serious referee, but not acceptance in its current form. The authors need to clarify the confidence metric, report variance, and release prompts/code with validation set details. If those are addressed, the paper is a useful addition to the KB-VQA literature. If not, the SKI claim is unverifiable and the +0.8 gain should be treated as exploratory.

For my own work: I'd probably not cite the numbers without the code, but it's a relevant related-work item.

Referee Report

3 major / 5 minor

Summary. The paper proposes KF-VQA, a training-free framework for knowledge-based visual question answering. It has three components: (1) low-noise retrieval queries generated by prompting a frozen VLM to summarize image-question pairs; (2) knowledge redundancy filtering in which an LLM selects fine-grained knowledge segments using question-specific visual details; and (3) selective knowledge integration (SKI), which gates external knowledge on an LLM confidence score. Experiments on OK-VQA and A-OKVQA report state-of-the-art results among compared methods, e.g., 63.2% on OK-VQA with Llama 3 8B versus 62.1% for DKA, with ablations attributing gains to each component.

Significance. If the results are reproducible, the framework is a useful training-free alternative for KB-VQA: it improves knowledge relevance without fine-tuning and is model-agnostic. The component-wise ablation in Table III supports the overall design. However, the main conceptual component, selective knowledge integration, depends on a confidence score whose definition is inconsistent, and the headline numbers come from hyperparameters selected on the target test set without error bars or significance tests. These issues must be resolved before the claim of outperforming state-of-the-art baselines is fully supported.

major comments (3)

[Section III-C, Eqs. (6)-(7)] The confidence score s is defined in two incompatible ways. Eq. (6) defines s = exp(Σ_t log p(y_t | x, y_<t)), i.e., the joint sequence likelihood, which for any nontrivial output length is typically far below 0.8 and hence cannot be meaningfully compared with τ=0.8. The sentence after Eq. (7) defines s = exp(max_j p_{LLM}(a_j | P4, E_j, I, q)), which is not in [0,1] and differs from Eq. (6) by orders of magnitude. Since SKI and the threshold τ=0.8 depend entirely on s, the claimed gain from SKI (+0.8 in Table III) is not reproducible from the text. Please state the implemented definition, provide the normalization, and include the missing prompts/code.
[Section IV-D, Fig. 3(c), Tables I-III] Hyperparameters r, h, and τ are selected on OK-VQA, the same benchmark used for the headline result, and no validation split or selection procedure is described. Fig. 3(c) sweeps τ on OK-VQA and picks 0.8; the reported 63.2% is therefore a post-selection maximum. This risks overfitting to the test set and makes the comparison against DKA (62.1%) less reliable. Please select hyperparameters on a held-out validation split (e.g., A-OKVQA val) or report sensitivity over all datasets.
[Tables I-III] All accuracy numbers are point estimates without error bars, multiple seeds, or significance tests. The total margin over DKA on OK-VQA is 1.1 points and the SKI ablation is 0.8 points; these differences may lie within run-to-run variability, especially since Eq. (7) describes sampling-based ensembles. Please provide variance estimates, repeated runs with different seeds/example orders, or bootstrap confidence intervals to support the 'outperforms' claim.

minor comments (5)

[Section IV-F] The text says 'In Fig. 3, we present a case study' but the case study is Figure 4; fix the cross-reference.
[References] Reference [20] is cited for Contriever, but the title given ('Webly Supervised Concept Expansion for General Purpose Vision Models') is not the Contriever paper. Please cite the correct source (e.g., Izacard et al., 2022).
[Section III] Prompts P1-P4 are used in Eqs. (1), (3), (5), and (7) but are never included in the text or appendix. Without them the framework is not fully specified; please provide them.
[Section III-C, Eq. (7)] The notation p_{LLM}(a_j | P4, E_j, I, q) is used both for selecting the ensemble answer and for computing confidence. Clarify how the initial knowledge-free pass is performed and how the ensemble selection interacts with the confidence gate.
[Figure 3] The three subpanels are referenced as (a), (b), (c) in the text, but the figure caption does not explicitly state the panel order; add explicit labels to the figure.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the framework is an empirical pipeline evaluated on external benchmarks, with no load-bearing reduction of predictions to fitted inputs or self-citations.

full rationale

KF-VQA is a training-free pipeline whose components (low-noise query extraction, knowledge redundancy filtering, and selective knowledge integration) are defined operationally in terms of frozen VLMs/LLMs and external knowledge bases, not in terms of the target VQA accuracy. The final answer is selected by an LLM argmax over candidate answers; no component's output is defined to be the benchmark accuracy. Hyperparameters r, h, and tau are selected via OK-VQA hyper-parameter analysis, which creates a test-set overfitting risk and could make the reported +0.8 SKI gain optimistic, but this is a validity/reproducibility concern rather than a circularity: the answer is not a re-statement of the chosen threshold. The confidence-score definition is inconsistent between Eq. 6 and the caption of Eq. 7, and prompts/code are not provided, which impedes reproduction, but again this is not an input-output equivalence by construction. The paper contains no load-bearing self-citations: prior works [10], [11], [13] are cited as external baselines or as sources of an empirical observation, not as the sole justification for the central claim. The SOTA comparison is against independently published numbers and the ablations compare the paper's own pipeline variants. Therefore no circular step can be identified under the stated standards.

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

The framework introduces five hyperparameters, three of which (tau, r, h) are tuned on the target benchmark. Its core assumptions are that the Google corpus, Contriever, and frozen VLM/LLM all behave as expected; no new entities are invented.

free parameters (5)

tau (confidence threshold) = 0.8
Chosen from Fig. 3(c) on OK-VQA; controls when external knowledge is injected.
r (number of retrieved documents) = 20
Set in implementation details; Fig. 3(a) shows performance stabilizes around r=15-20.
h (number of knowledge segments) = 7
Chosen from Fig. 3(b); accuracy peaks at h=7.
m (number of sampled predictions) = 5
Following prior work [7],[10] for ensemble robustness.
n (number of in-context examples) = 10
Following prior work [10],[11].

assumptions (5)

domain assumption Google Search Corpus [4] contains relevant knowledge for the test questions
Section III-A uses this corpus as the sole knowledge source; its coverage is taken for granted.
domain assumption Contriever dot-product similarity ranks relevant documents higher when the query is the question plus VLM keywords
Section III-A relies on this embedding quality; no retrieval quality metrics are reported.
domain assumption LLM/VLM prompts produce accurate keywords, visual details, and filtered segments
Section III-B assumes the frozen models respond correctly to P1-P3; examples in Fig. 4 indicate errors can occur.
ad hoc to paper The exponential of summed log probabilities is a calibrated confidence score in (0,1)
Section III-C introduces this score; the paper provides no calibration analysis and the definition is inconsistent with the text after Eq. (7).
standard math Standard probability chain rule, argmax decoding, and dot-product similarity
Used throughout Section III; no need for external evidence.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Knowledge Noise Mitigation Framework for Knowledge-based Visual Question Answering." pith.science (2026). https://pith.science/paper/7IHQJG24

@misc{pith2026250909159,
  author       = {Pith},
  title        = {Pith review of: A Knowledge Noise Mitigation Framework for Knowledge-based Visual Question Answering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7IHQJG24}},
  note         = {Machine review of arXiv:2509.09159}
}

read the original abstract

Knowledge-based visual question answering (KB-VQA) requires a model to understand images and utilize external knowledge to provide accurate answers. Existing approaches often directly augment models with retrieved information from knowledge sources while ignoring substantial knowledge redundancy, which introduces noise into the answering process. To address this, we propose a training-free framework with knowledge focusing for KB-VQA, that mitigates the impact of noise by enhancing knowledge relevance and reducing redundancy. First, for knowledge retrieval, our framework concludes essential parts from the image-question pairs, creating low-noise queries that enhance the retrieval of highly relevant knowledge. Considering that redundancy still persists in the retrieved knowledge, we then prompt large models to identify and extract answer-beneficial segments from knowledge. In addition, we introduce a selective knowledge integration strategy, allowing the model to incorporate knowledge only when it lacks confidence in answering the question, thereby mitigating the influence of redundant information. Our framework enables the acquisition of accurate and critical knowledge, and extensive experiments demonstrate that it outperforms state-of-the-art methods.

Figures

Figures reproduced from arXiv: 2509.09159 by the authors.

**Figure 2.** Overview of the proposed framework. KF-VQA constructs low-noise queries by extracting essential content from the image and question to retrieve [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Hyper-parameter analysis on OK-VQA. [13] for reasoning. Baseline is inferior to the LLM without any external knowledge, as redundant and inaccurate knowledge fails to benefit training-free reasoning. We introduce lownoise queries for knowledge retrieval (i.e., w/ LNQ) and improve the performance of LLMs. That is, the queries with key information are beneficial for locating accurate knowledge. We further integrate … view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Case study on OK-VQA. Retrieved knowledge from different methods is listed along with the corresponding number of knowledge tokens. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

23 extracted references · 3 linked inside Pith

[1]

Ok-vqa: A visual question answering benchmark requiring external knowledge,

Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi, “Ok-vqa: A visual question answering benchmark requiring external knowledge,” in CVPR, 2019, pp. 3195–3204

2019
[2]

Vqa: Visual question answering,

Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh, “Vqa: Visual question answering,” in ICCV, 2015, pp. 2425–2433

2015
[3]

Krisp: Integrating implicit and symbolic knowledge for open-domain knowledge-based vqa,

Kenneth Marino, Xinlei Chen, Devi Parikh, Abhinav Gupta, and Marcus Rohrbach, “Krisp: Integrating implicit and symbolic knowledge for open-domain knowledge-based vqa,” in CVPR, 2021, pp. 14111–14121

2021
[4]

Weakly- supervised visual-retriever-reader for knowledge-based question answer- ing,

Man Luo, Yankai Zeng, Pratyay Banerjee, and Chitta Baral, “Weakly- supervised visual-retriever-reader for knowledge-based question answer- ing,” in EMNLP, 2021

2021
[5]

Conceptnet—a practical commonsense reasoning tool-kit,

Hugo Liu and Push Singh, “Conceptnet—a practical commonsense reasoning tool-kit,” BT technology journal , vol. 22, no. 4, pp. 211–226, 2004

2004
[6]

Wikidata: a free collaborative knowledgebase,

Denny Vrande ˇci´c and Markus Kr ¨otzsch, “Wikidata: a free collaborative knowledgebase,” Communications of the ACM , vol. 57, no. 10, pp. 78– 85, 2014

2014
[7]

An empirical study of gpt-3 for few- shot knowledge-based vqa,

Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Yumao Lu, Zicheng Liu, and Lijuan Wang, “An empirical study of gpt-3 for few- shot knowledge-based vqa,” in AAAI, 2022, vol. 36, pp. 3081–3089

2022
[8]

Promptcap: Prompt-guided image captioning for vqa with gpt-3,

Yushi Hu, Hang Hua, Zhengyuan Yang, Weijia Shi, Noah A Smith, and Jiebo Luo, “Promptcap: Prompt-guided image captioning for vqa with gpt-3,” in ICCV, 2023, pp. 2963–2975

2023

Show all 23 references

[9]

Prompting large language models with answer heuristics for knowledge-based visual question answering,

Zhenwei Shao, Zhou Yu, Meng Wang, and Jun Yu, “Prompting large language models with answer heuristics for knowledge-based visual question answering,” in CVPR, 2023, pp. 14974–14983

2023
[10]

A simple baseline for knowledge-based visual question answering,

Alexandros Xenos, Themos Stafylakis, Ioannis Patras, and Georgios Tzimiropoulos, “A simple baseline for knowledge-based visual question answering,” in EMNLP, 2023, pp. 14871–14877

2023
[11]

Knowledge acquisition disentanglement for knowledge-based visual question an- swering with large language models,

Wenbin An, Feng Tian, Jiahao Nie, Wenkai Shi, Haonan Lin, Yan Chen, QianYing Wang, Yaqiang Wu, Guang Dai, and Ping Chen, “Knowledge acquisition disentanglement for knowledge-based visual question an- swering with large language models,” arXiv preprint arXiv:2407.15346 , 2024

2024 arXiv
[12]

Retrieval augmented visual question answering with outside knowledge,

Weizhe Lin and Bill Byrne, “Retrieval augmented visual question answering with outside knowledge,” in EMNLP, 2022, pp. 11238–11254

2022
[13]

Self-bootstrapped visual-language model for knowledge selection and question answering,

Dongze Hao, Qunbo Wang, Longteng Guo, Jie Jiang, and Jing Liu, “Self-bootstrapped visual-language model for knowledge selection and question answering,” in EMNLP, 2024, pp. 1857–1868

2024
[14]

A-okvqa: A benchmark for visual question answering using world knowledge,

Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi, “A-okvqa: A benchmark for visual question answering using world knowledge,” in ECCV, 2022, pp. 146– 162

2022
[15]

Kat: A knowledge augmented transformer for vision-and-language,

Liangke Gui, Borui Wang, Qiuyuan Huang, Alexander G Hauptmann, Yonatan Bisk, and Jianfeng Gao, “Kat: A knowledge augmented transformer for vision-and-language,” in NAACL, 2022, pp. 956–968

2022
[16]

Revive: regional visual representation matters in knowledge-based visual question answering,

Yuanze Lin, Yujia Xie, Dongdong Chen, Yichong Xu, Chenguang Zhu, and Lu Yuan, “Revive: regional visual representation matters in knowledge-based visual question answering,” in NeurIPS, 2022, pp. 10560–10571

2022
[17]

Fine-grained late-interaction multi-modal retrieval for retrieval augmented visual question answering,

Weizhe Lin, Jinghong Chen, Jingbiao Mei, Alexandru Coca, and Bill Byrne, “Fine-grained late-interaction multi-modal retrieval for retrieval augmented visual question answering,” in NeurIPS, 2023, pp. 22820– 22840

2023
[18]

Deep modular co-attention networks for visual question answering,

Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian, “Deep modular co-attention networks for visual question answering,” in CVPR, 2019, pp. 6281–6290

2019
[19]

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

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,” in Findings of EMNLP , 2022, pp. 951–967

2022
[20]

Webly supervised concept expansion for general purpose vision models,

Amita Kamath, Christopher Clark, Tanmay Gupta, Eric Kolve, Derek Hoiem, and Aniruddha Kembhavi, “Webly supervised concept expansion for general purpose vision models,” in ECCV, 2022, pp. 662–681

2022
[21]

The llama 3 herd of models,

Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al., “The llama 3 herd of models,” arXiv preprint arXiv:2407.21783, 2024

2024 arXiv
[22]

Minigpt-4: Enhancing vision-language understanding with advanced large language models,

Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed El- hoseiny, “Minigpt-4: Enhancing vision-language understanding with advanced large language models,” arXiv preprint arXiv:2304.10592 , 2023

2023 arXiv
[23]

Instructblip: towards general-purpose vision-language models with instruction tuning,

Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi, “Instructblip: towards general-purpose vision-language models with instruction tuning,” in NeurIPS, 2023, pp. 49250–49267

2023

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

Ok-vqa: A visual question answering benchmark requiring external knowledge,

Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi, “Ok-vqa: A visual question answering benchmark requiring external knowledge,” in CVPR, 2019, pp. 3195–3204

2019

[2] [2]

Vqa: Visual question answering,

Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh, “Vqa: Visual question answering,” in ICCV, 2015, pp. 2425–2433

2015

[3] [3]

Krisp: Integrating implicit and symbolic knowledge for open-domain knowledge-based vqa,

Kenneth Marino, Xinlei Chen, Devi Parikh, Abhinav Gupta, and Marcus Rohrbach, “Krisp: Integrating implicit and symbolic knowledge for open-domain knowledge-based vqa,” in CVPR, 2021, pp. 14111–14121

2021

[4] [4]

Weakly- supervised visual-retriever-reader for knowledge-based question answer- ing,

Man Luo, Yankai Zeng, Pratyay Banerjee, and Chitta Baral, “Weakly- supervised visual-retriever-reader for knowledge-based question answer- ing,” in EMNLP, 2021

2021

[5] [5]

Conceptnet—a practical commonsense reasoning tool-kit,

Hugo Liu and Push Singh, “Conceptnet—a practical commonsense reasoning tool-kit,” BT technology journal , vol. 22, no. 4, pp. 211–226, 2004

2004

[6] [6]

Wikidata: a free collaborative knowledgebase,

Denny Vrande ˇci´c and Markus Kr ¨otzsch, “Wikidata: a free collaborative knowledgebase,” Communications of the ACM , vol. 57, no. 10, pp. 78– 85, 2014

2014

[7] [7]

An empirical study of gpt-3 for few- shot knowledge-based vqa,

Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Yumao Lu, Zicheng Liu, and Lijuan Wang, “An empirical study of gpt-3 for few- shot knowledge-based vqa,” in AAAI, 2022, vol. 36, pp. 3081–3089

2022

[8] [8]

Promptcap: Prompt-guided image captioning for vqa with gpt-3,

Yushi Hu, Hang Hua, Zhengyuan Yang, Weijia Shi, Noah A Smith, and Jiebo Luo, “Promptcap: Prompt-guided image captioning for vqa with gpt-3,” in ICCV, 2023, pp. 2963–2975

2023