Pith. sign in

REVIEW 5 major objections 7 minor 21 references

CUE-M: Contextual Understanding and Enhanced Search with Multimodal Large Language Model

T0 review · 5 major / 7 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read The paper claims CUE-M, a five-stage multimodal search pipeline with intent refinement and relevance filtering, lifts knowledge-based visual question answering on Encyclopedic VQA to 0.762 accuracy with a 7B model, near the 0.870 oracle…

desk verdict Solid multimodal RAG system with a strong public-benchmark result; the real-world ablation numbers hinge on an undisclosed split between prompt tuning and evaluation. read the letter →

arxiv 2411.12287 v3 pith:PHWNVDBJ submitted 2024-11-19 cs.CL

Dongyoung Go , Taesun Whang , Chanhee Lee , Hwa-Yeon Kim , Sunghoon Park , Seunghwan Ji , Jinho Kim , Dongchan Kim
show 1 more author
Young-Bum Kim
This is my paper · ORCID
classification cs.CL
keywords multimodalretrieval-augmentedgenerationlargelanguagemodelknowledge-basedvisualquestionansweringintentrefinementrelevancefilteringsafetyevaluationEncyclopedicVQAMM-SafetyBench
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 is trying to establish that multimodal search should be organized as a deliberate sequence of retrieval and reasoning steps, not left to a single model call. CUE-M turns a user's image-plus-text question into enriched image context (captions, tags, similar-image documents), refines it into a structured intent, generates supplementary search queries, routes to general web, shopping, or map APIs, filters the retrieved documents by relevance, and only then lets a multimodal LLM compose the final answer. On 2.1k real queries from a large Q&A service the full pipeline beats a searchless MLLM baseline with a 0.639 win rate, and each ablation step contributes to that number. On the public Encyclopedic VQA benchmark, CUE-M with Qwen2.5-7B-Instruct reaches 0.762 accuracy, close to the 0.870 oracle that uses ground-truth documents, while LLaVA-1.5-7B with the same pipeline reaches 0.467. The paper also claims the attached safety filter does not sacrifice protection, with an average attack success rate of 0.43 on MM-SafetyBench, on par with prior multimodal models.

What carries the argument

The load-bearing object is the CUE-M pipeline itself. It chains five stages: image context enrichment (an MLLM caption plus similar-image search and image-tag text retrieval), intention refinement (a few-shot prompted LLM that collapses the query and retrieved summaries into one structured sentence of what the user needs), contextual query generation (one to three supplementary searches derived from that intent), application selection (a prompted API Selector that ranks shopping, map, or general text-search APIs by the likelihood of a positive token), and relevance-based filtering (an ELECTRA-based cross-encoder that ranks and keeps the top documents). A multi-stage safety filter runs in parallel, combining lightweight image and text classifiers, a prompted multimodal detector, an instance-level query-answer database, and category-level policy answers. The argument works by showing that removing intent refinement or relevance filtering measurably lowers win rate, so the pipeline's decomposition itself, not any single component, is what carries the result.

What would settle it

A concrete check: split the 2.1k Knowledge-iN queries into disjoint prompt-tuning and held-out test sets, run CUE-M against the no-retrieval baseline on the held-out set, and have human raters judge the responses; if the win rate falls from 0.639 toward 0.5 or disagrees with GPT-4o on the same pairs, the pipeline's claimed effectiveness does not generalize. A second check on Encyclopedic VQA: swap the external similar-image and text-search APIs for an equally capable open retriever; if the 0.762 accuracy collapses, the gain belongs to the search APIs rather than to the pipeline design.

Watch

Extended reading notes

Core claim

CUE-M's central discovery is that a decomposed retrieval pipeline, whose components are each prompt-tuned rather than trained, can convert a small open MLLM into a competitive knowledge-based visual question answerer. The intent refiner is the piece the ablations single out: translating the raw query plus image-derived summaries into a single structured intent is what lets the query generator form useful supplementary queries, and it also feeds the API selector and the safety detector. The relevance classifier, an ELECTRA-based cross-encoder, then removes low-quality documents before answer generation. With these stages, CUE-M beats previous knowledge-augmented methods on Encyclopedic VQA (Wiki-LLaVA 0.218, DPR*V+T 0.291, EchoSight 0.418) and approaches the ground-truth oracle (PaLM 0.870) without access to ground-truth documents. On the curated real-world set the win rate rises from 0.5 for the no-retrieval baseline to 0.639 for the full pipeline, and the safety experiments show comparable attack-success rates while acknowledging vulnerabilities to multimodal-retrieval-specific jailbreaks.

Load-bearing premise

The main load-bearing premise is that the curated 2.1k-query evaluation set, which was filtered toward web, shopping, and map intents and used to tune prompts, is representative enough that GPT-4o's win-rate judgments on it reflect real user preference; if tuning and test queries overlap or the judge diverges from humans, the reported gains are fitted rather than predictive.

Editorial extensions

If this is right

  • Open 7B MLLMs can reach near-oracle performance on single-hop knowledge-based VQA without fine-tuning, provided the retrieval context is enriched and filtered through intent refinement.
  • The pipeline is backbone-agnostic: the same prompts move LLaVA-1.5-7B from 0.169 to 0.467 and Qwen2.5-7B-Instruct from 0.304 to 0.762 on Encyclopedic VQA.
  • Component-level ablations imply that intent refinement and relevance filtering are the main contributors to answer quality; dropping the relevance classifier lowers win rate from 0.639 to 0.547.
  • Safety filtering can be layered on without a major accuracy trade-off, at least on the six MM-SafetyBench categories tested, with average ASR 0.43.
  • Because the API Selector is prompted rather than trained, adding a new external service reduces to adding an API description, making the design extensible to shopping, maps, or other specialized tools.

Reading between the lines

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

  • Editorial inference: the 0.639 win rate over a no-retrieval baseline is measured by GPT-4o on a filtered query distribution; a human-preference panel on held-out queries would be a stronger test that the gain is real rather than judge-aligned.
  • Editorial inference: the intent refiner should matter most when image and text point in different directions; a targeted dataset of mismatched captions and questions would isolate whether the win-rate gain comes from disambiguation or simply from more retrieval budget.
  • Editorial inference: the safety evaluation covers only six scenarios and the paper admits multimodal retrieval jailbreaks; a stress test that composes harmful intents with the API selector, such as indirect or meme-based queries, would reveal whether the category-level filters generalize beyond the benchmark.
  • Editorial inference: the modular stages could be reused independently, e.g., the intent refiner and relevance classifier as a standalone multimodal query-understanding and filtering benchmark, decoupled from the answer generator.
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

5 major / 7 minor

Summary. The paper proposes CUE-M, a multi-stage multimodal retrieval-augmented generation pipeline that combines image enrichment, intention refinement, contextual query generation, external API selection, relevance-based document filtering, and a layered safety filter. The authors evaluate the system on a curated Naver Knowledge-iN dataset, on Encyclopedic-VQA, and on MM-SafetyBench, reporting that CUE-M reaches a 0.639 win rate over a baseline MLLM on the curated set, 0.762 accuracy with Qwen2.5-7B-Instruct on Encyclopedic-VQA (close to the 0.870 oracle), and an average ASR of 0.43 on MM-SafetyBench. The central claim is that each pipeline stage contributes distinct benefits and that the full system outperforms prior knowledge-augmented multimodal methods without additional training.

Significance. If the reported numbers are predictive, CUE-M is a meaningful engineering contribution: it demonstrates that a modular, search-API-based RAG pipeline can substantially improve knowledge-based multimodal question answering, and the public Encyclopedic-VQA result (Table 2) is notably strong for a zero-/few-shot system. The paper is also unusually transparent about sub-module prompts, failure cases, and the limitations of its own metrics (App. C and App. D). However, the central real-world claim rests on a win-rate evaluation whose tuning/evaluation split is not documented, the evaluation set is curated into API-aligned categories by the same model family that judges the responses, and the safety comparison lacks uncertainty quantification. These concerns are fixable but currently limit the strength of the paper's conclusions.

major comments (5)
  1. [Sec. 4.1, Table 1; Sec. 3.7; App. E] The paper never states whether the 2.1k Knowledge-iN queries used for the win-rate table are disjoint from the validation samples on which automatic prompt tuning selects final prompts. Since App. E says 'the final prompt is selected after iterative tuning, based on its performance on validation samples,' and Sec. 3.7 says the tuning dataset is 'derived from user query data,' the reported win rates (0.639 full, 0.595 without Intention Refiner, 0.547 without Relevance Classifier) would be in-sample fitted values if the splits overlap. This is load-bearing because Table 1 is the primary evidence that each pipeline component contributes. Please specify the exact split between prompt-tuning validation and final evaluation, and if no disjoint split exists, re-run the ablations on a held-out test set.
  2. [Sec. 4.1; App. F] The evaluation set was filtered by GPT-4o into Web, Shopping, and Maps categories that align with the pipeline's available APIs (App. F, Table 14). This makes the win rate a measure on a curated subset of multimodal queries rather than on the general multimodal query distribution. Additionally, GPT-4o is used both to curate the data and to judge the win-rate comparisons, creating a risk of systematic preference for pipeline-style answers. Please report agreement between GPT-4o curation and human raters, provide the category distribution of the unfiltered query pool, and include a human-agreement study for the GPT-4o judgments, ideally with a second judge.
  3. [Sec. 4.3, Table 3] The safety evaluation reports point estimates of attack success rate over six scenarios with no confidence intervals or significance tests. Comparisons such as 0.36 vs. 0.38 for the PO scenario are within likely sampling noise, and the scenarios were selected ('six relevant scenarios') rather than using the full benchmark. Please report the number of trials per scenario, compute confidence intervals, and either justify the scenario selection or evaluate on the full MM-SafetyBench split.
  4. [Sec. 4.2, Table 2 and Table 15] The headline comparison against EchoSight and Wiki-LLaVA mixes backbone models, retrieval APIs, and evaluation protocols. CUE-M uses Google Lens plus Google Search restricted to Wikipedia, while EchoSight uses BLIP-2 + Mistral-7B and Wiki-LLaVA uses LLaVA-1.5 with a different hierarchical retrieval scheme; the 'Oracle' rows use ground-truth entities and are not directly comparable zero-/few-shot numbers. Please state the exact retrieval configuration and prompting protocol for every row, report any variance across retrieval runs, and soften the 'state-of-the-art' claim unless the comparison is made under matched conditions.
  5. [Sec. 4.1, Sec. 4.2] The win-rate metric is the only headline result for the Knowledge-iN study, and it is defined relative to a baseline MLLM response rather than as an absolute quality measure. The paper should report absolute agreement with the reference answers or include a secondary metric on the same queries; App. C provides ROUGE-L, BERTScore, and NER, but these are presented as auxiliary and are not tied to the win-rate evaluation in Table 1.
minor comments (7)
  1. [Abstract] 'real-word datasets' should be 'real-world datasets'.
  2. [Tables 2 and 15] 'Vanlia' should be 'Vanilla' in both tables.
  3. [Tables 4-7] The prompt templates contain typographical errors such as 'docummnt_summaries' and 'usre_query'; please correct them.
  4. [App. C] The BERTScore discussion correctly notes a known limitation, but the paragraph should also state whether the differences in Table 11 are statistically significant or merely descriptive.
  5. [Table 14] The 'Internal' category is defined by an 'out_of_scope_questions' placeholder; please either show the actual examples or explain that they are masked for policy reasons.
  6. [App. D and Sec. 4.3] The main text says safety performance is 'comparable,' while App. D documents systematic jailbreak vulnerabilities in multimodal retrieval; please reconcile these statements in the conclusion or abstract so the claims match the evidence.
  7. [Table 13] The Maps category is notably small (138 samples) relative to Web and Shopping; the paper should report per-category win rates or note the imbalance as a limitation of the curated set.

Circularity Check

1 steps flagged · score 4.0 of 10

Table 1's component win rates may be in-sample if the 2.1k Knowledge-iN eval queries overlap the automatic-prompt-tuning validation set; public benchmarks keep the overall pipeline claim independent.

  1. fitted input called prediction [Sec. 3.7 (Automatic prompt tuning), Sec. 4.1 (Real-World Evaluation), App. E, Table 1]
    "A curated dataset is used to define the desired behavior for each sub-module, derived from user query data. ... The final prompt is selected after iterative tuning, based on its performance on validation samples, ensuring optimal effectiveness for downstream tasks. ... we extract multimodal queries combining text and user-provided images ... yielding 2.1k queries covering diverse categories."

    The 2.1k Knowledge-iN queries evaluated in Table 1 are drawn from the same 'user query data' source used for automatic prompt tuning, and the paper never documents a disjoint split between the validation set used to select prompts and the evaluation set. If the validation samples overlap with the 2.1k queries, the reported win rates (0.639 full pipeline; 0.595 without Intention Refiner; 0.547 without Relevance Classifier) are the prompt tuner's validation objective, not out-of-sample predictions, and the ablation increments become fitted rather than predictive. The public Encyclopedic-VQA and MM-SafetyBench evaluations are independent, so this concern affects the component-level claims rather than the overall external benchmark results.

full rationale

CUE-M is an empirical system paper, not a derivation, so the main circularity risk is in the evaluation protocol. The paper's automatic prompt tuning (Sec. 3.7, App. E) selects prompts by performance on validation samples derived from user query data, while the real-world evaluation (Sec. 4.1) uses a 2.1k Knowledge-iN set drawn from the same kind of user query data with no documented train/validation/test split. If the evaluation set overlaps the prompt-tuning validation set, Table 1's win rates are in-sample fitted values rather than predictive estimates, making the component-ablation evidence partially circular. The public Encyclopedic-VQA and MM-SafetyBench results provide independent support for the overall pipeline's effectiveness and safety, so the central claim retains external validity. I found no other load-bearing circular steps: citations to prior work, including the authors' Naver CUE reference, are not used as a substitute for evidence, and the public benchmark comparisons against Wiki-LLaVA, DPR*, and EchoSight are external. The main issue is a protocol gap: the paper should state whether the 2.1k evaluation queries are disjoint from the prompt-tuning validation samples. This is a moderate circularity risk rather than a fully forced reduction, hence a score of 4.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

No new physical or mathematical entities are introduced. The load-bearing inputs are empirical: a filtered real-world dataset, GPT-4o judgments, hand-set thresholds, and prompt tuning on validation samples. The central claim does not rest on a derivation.

free parameters (2)
  • Positive-to-negative sample ratio for relevance classifier = 1:4
    Appendix H states this ratio performed best on the evaluation set; it is chosen by data, not derived.
  • NLI thresholds for relevance training data = positive > 0.9, negative < 0.3
    Appendix H defines these thresholds to build positive and negative examples; they are hand-set.
assumptions (4)
  • domain assumption GPT-4o simulated human judgments on Naver Knowledge-iN responses approximate real user preference
    Sec 4.1 uses GPT-4o to compute Win Rate; no human agreement study is reported.
  • domain assumption Knowledge-iN answers with more than two upvotes are reliable gold references
    Sec 4.1 filters by upvotes, assuming crowd agreement implies correctness.
  • ad hoc to paper Curated dataset categories (Web, Shopping, Maps) are representative of real multimodal user queries
    Appendix F refines selection to categories most likely to benefit from external APIs; this is chosen for the method, not a neutral sample.
  • domain assumption Google Lens and Google Search API restricted to Wikipedia provide a fair retrieval setting for Encyclopedic VQA
    Appendix I imposes a Wikipedia restriction; removing the constraint improves performance, so the comparison is setting-specific.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CUE-M: Contextual Understanding and Enhanced Search with Multimodal Large Language Model." pith.science (2026). https://pith.science/paper/PHWNVDBJ

@misc{pith2026241112287,
  author       = {Pith},
  title        = {Pith review of: CUE-M: Contextual Understanding and Enhanced Search with Multimodal Large Language Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PHWNVDBJ}},
  note         = {Machine review of arXiv:2411.12287}
}
read the original abstract

The integration of Retrieval-Augmented Generation (RAG) with Multimodal Large Language Models (MLLMs) has revolutionized information retrieval and expanded the practical applications of AI. However, current systems struggle in accurately interpreting user intent, employing diverse retrieval strategies, and effectively filtering unintended or inappropriate responses, limiting their effectiveness. This paper introduces Contextual Understanding and Enhanced Search with MLLM (CUE-M), a novel multimodal search framework that addresses these challenges through a multi-stage pipeline comprising image context enrichment, intent refinement, contextual query generation, external API integration, and relevance-based filtering. CUE-M incorporates a robust filtering pipeline combining image-based, text-based, and multimodal classifiers, dynamically adapting to instance- and category-specific concern defined by organizational policies. Extensive experiments on real-word datasets and public benchmarks on knowledge-based VQA and safety demonstrated that CUE-M outperforms baselines and establishes new state-of-the-art results, advancing the capabilities of multimodal retrieval systems.

Figures

Figures reproduced from arXiv: 2411.12287 by the authors.

Figure 1
Figure 1. CUE-M is a multimodal RAG pipeline de￾signed to interpret multimodal user intent and generate accurate answers through advanced retrieval, seamlessly integrating with diverse applications. Key challenges in multimodal RAG systems lie in three areas: accurately interpreting user intent, employing diverse information retrieval strategies, and effectively filtering unintended or inappropri￾ate responses. Previous syste… view at source ↗
Figure 2
Figure 2. A high-level architecture of CUE-M pipeline, illustrating stages from image-based information retrieval [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. A high-level architecture of safety pipeline. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 16 canonical work pages

  1. [1]

    Determine what information is required to answer the question

    work page

  2. [2]

    Do not generate a query if the provided information is sufficient to answer the question

    work page

  3. [3]

    Accessed: Mar 20, 2024

    An Easy Introduction to Multimodal Retrieval- Augmented Generation. Accessed: Mar 20, 2024. Jason Tang, Garrin McGoldrick, Marie Al-Ghossein, and Ching-Wei Chen. 2024. Captions are worth a thousand words: Enhancing product retrieval with pretrained image-to-text models. arXiv preprint arXiv:2402.08532. Xinyuan Wang, Chenxi Li, Zhen Wang, Fan Bai, Haotian ...

    arXiv 2024

  4. [4]

    arXiv preprint arXiv:2311.17136

    Uniir: Training and benchmarking universal multimodal information retrievers. arXiv preprint arXiv:2311.17136. Zilin Xiao, Ming Gong, Paola Cascante-Bonilla, Xingyao Zhang, Jie Wu, and Vicente Ordonez. 2024. Grounding language models for visual entity recog- nition. arXiv preprint arXiv:2402.18695. Kelvin Xu. 2015. Show, attend and tell: Neural im- age ca...

    arXiv 2024

  5. [5]

    arXiv preprint arXiv:2404.01954

    Hyperclova x technical report. arXiv preprint arXiv:2404.01954. Hongyeon Yu, Seung Hak Yu, and Yong Beom Kim

    arXiv

  6. [6]

    In Communications of the Korean In- stitute of Information Scientists and Engineers, pages 34–41

    Naver cue: Search service based on large lan- guage models. In Communications of the Korean In- stitute of Information Scientists and Engineers, pages 34–41. Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2019. Bertscore: Eval- uating text generation with bert. arXiv preprint arXiv:1904.09675. Yongchao Zhou, Andrei Ioan Mures...

    arXiv 2019

  7. [9]

    image tag + user query

    If additional searching is needed, generate a minimum of 1 and a maximum of 3 queries. {previous_chat} <|user|>{refined_information}<|endofturn|> <|search_result|>{docummnt_summaries}<|endofturn|> <|assistant|> Table 5: The translated prompt template for Query Generator (partially masked due to internal policy). B Example queries B.1 Example of the limita...

    work page

  8. [10]

    Please describe the situation of the painting in more detail

    ROUGE-L (Lin, 2004), which measures summarization quality by evaluating the longest common subsequence (LCS) overlap between the reference text and the generated text. This metric emphasizes sequence-level coherence. 11 {system_prompt} {api_description} Below is a conversation between an AI language model and a user. The AI language model understands exac...

    work page 2004

Show all 21 references
  1. [11]

    This metric evaluates semantic similarity at a fine-grained level

    BertScore (Zhang et al., 2019), which calculates the cosine similarity of contextual embeddings from pretrained BERT models between the reference answer and the generated answer. This metric evaluates semantic similarity at a fine-grained level

    2019

  2. [12]

    What’s a stronger alternative?

    Named Entity Recall-target (NER) (Nan et al., 2021), defined as the number of overlapping named entities found in both the reference and generated answers, divided by the total number of named entities in the reference answer. This metric assesses the preservation of critical ...

    2023

  3. [13]

    It has a capacity of 23A and a voltage of 12V , which may offer compatibility with the MEHV10 camcorder

    **Toshiba Genuine 23A 12V 10-pack Car Remote Battery** - Used for car remotes, produced by Toshiba. It has a capacity of 23A and a voltage of 12V , which may offer compatibility with the MEHV10 camcorder

  4. [14]

    **Toshiba Lithium Battery CR2032 3V , 10-pack Coin Battery for Smart Key/Remote** - Used for smart keys and remotes, with a capacity of 225mAh and a voltage of 3V , potentially compatible with the MEHV10 camcorder

  5. [15]

    These products are all Toshiba-made and might be compatible with the MEHV10 camcorder battery

    **Toshiba Camileo BW10 HD Waterproof Camcorder Battery** - This camcorder battery, with a capacity of 1.2V and a voltage of 3.6V , is also produced by Toshiba and may be compatible with the MEHV10 camcorder. These products are all Toshiba-made and might be compatible with the ...

  6. [16]

    - Made with mesh and synthetic leather (faux leather), with a 3-4 cm sole height

    **Nike V2K Run Tech Summit White Metallic Silver (Domestic Version)** - Lightweight with shock absorption and height-boosting features. - Made with mesh and synthetic leather (faux leather), with a 3-4 cm sole height. - Features a low-top ankle height and rubber sole for comfort

  7. [17]

    - Made with mesh and synthetic leather (faux leather), with a 3-4 cm sole height

    **Nike V2K Run Tech Summit White Metallic Silver Kang Min-kyung FD0736-100 (Domestic Version)** - Lightweight with shock absorption and height-boosting features. - Made with mesh and synthetic leather (faux leather), with a 3-4 cm sole height. - Features a low-top ankle height...

  8. [18]

    - Made with other materials, mesh, and synthetic leather (faux leather), with a 3-4 cm sole height

    **Nike V2K Run Tech Summit White Metallic Silver Sneakers** - Lightweight with a height-boosting feature. - Made with other materials, mesh, and synthetic leather (faux leather), with a 3-4 cm sole height. - Features a low-top ankle height and rubber sole for comfort. These Ni...

  9. [19]

    - Stylish design and colors with lightweight, height-boosting, and ventilation features

    **Nike V2K Run HJ4497-100 (Domestic Version)** - Released as a women’s product. - Stylish design and colors with lightweight, height-boosting, and ventilation features. - Consider sizing carefully as it can be tricky

  10. [20]

    - Combines black and gray for a sleek look

    **Nike V2K Run Black Anthracite (HJ4497-001)** - Released as a men’s product. - Combines black and gray for a sleek look. - Reviews mention it’s lightweight and comfortable

  11. [21]

    - Affordable with an attractive design

    **Nike V2K Run HJ4497-002** - Released as a women’s product. - Affordable with an attractive design. All these Nike V2K Run models come in various designs and colors but are separated into men’s and women’s versions, so please confirm before purchasing. Table 22: Generation sa...

  12. [2023]

    In International Con- ference on Machine Learning, pages 17283–17300

    Grounding language models to images for multimodal inputs and outputs. In International Con- ference on Machine Learning, pages 17283–17300. PMLR. Paul Lerner, Olivier Ferret, and Camille Guinaudeau

  13. [2024]

    In European Conference on Information Retrieval, pages 421–438

    Cross-modal retrieval for knowledge-based visual question answering. In European Conference on Information Retrieval, pages 421–438. Springer. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Hein- rich Küttler, Mike Lewis, Wen-tau...

    2020 arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.