VideoRAG: Retrieval-Augmented Generation over Video Corpus

Jinheon Baek; Kangsan Kim; Soyeong Jeong; Sung Ju Hwang

REVIEW 3 major objections 5 minor 6 cited by

VideoRAG: Retrieval-Augmented Generation over Video Corpus

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Retrieving videos as knowledge improves RAG answers over text-only and image-based retrieval.

desk verdict VideoRAG is a solid integration of video retrieval, LVLM generation, and frame selection, but the experiments don't isolate visual grounding from retrieval quality, so the headline claim is suggestive, not proven. read the letter →

arxiv 2501.05874 v3 pith:ZVYJABDV submitted 2025-01-10 cs.CV cs.AIcs.CLcs.IRcs.LG

Soyeong Jeong , Kangsan Kim , Jinheon Baek , Sung Ju Hwang This is my paper · ORCID

classification cs.CVcs.AIcs.CLcs.IRcs.LG

keywords retrieval-augmentedgenerationvideoretrievallargelanguagemodelsmultimodalknowledgeframeselectionquestionansweringHowTo100Mautomaticspeechrecognition

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

VideoRAG asks whether a video corpus can serve as external knowledge for retrieval-augmented generation, and answers yes. Instead of flattening videos into text or assuming the relevant video is already known, the method retrieves videos by embedding queries and video frames plus transcripts through a large video-language model, then feeds the retrieved video's most informative frames and text into the same model to generate the answer. The paper reports that this joint visual-textual video pipeline outperforms text-only, text-image, and text-video RAG baselines on instructional question answering built from WikiHowQA and HowTo100M. It also introduces a trained frame-selection step that beats uniform sampling, and uses speech-to-text to supply transcripts when subtitles are missing. If correct, the work extends the RAG paradigm from text and images to open-domain video knowledge.

What carries the argument

The load-bearing object is the large video-language model (LVLM), a model that encodes video frames and text into a shared token stream and can both embed a query-video pair for retrieval and generate an answer from the concatenation of retrieved frames, transcripts, and query. Around it sit two supporting mechanisms: a learned frame-scoring function that scores subsets of frames, with k-means++ first reducing the search space, and picks the most informative frames separately for retrieval and generation; and an automatic-speech-recognition fallback that produces auxiliary text for videos without subtitles. These allow the pipeline to treat the video itself as the knowledge unit rather than a pre-chosen video or a text summary.

What would settle it

An experiment that retrieves with a single fixed retriever and then generates from transcript only, frames only, and frames plus transcript for the same retrieved videos would isolate the visual contribution; if frames plus transcript does not beat transcript only, the central claim fails. A second check would swap VideoRAG's video-specific retrieval encoder for a text retriever while keeping the generation input identical, to show how much of the gap comes from retrieval rather than modality fusion.

Watch

Extended reading notes

Core claim

The paper is trying to establish that a video corpus can be a practical external knowledge source for retrieval-augmented generation when both retrieval and generation are handled by a large video-language model. Concretely, it claims that dynamically retrieving videos by embedding similarity and then generating from selected frames plus transcripts yields better answers on instructional question answering than retrieving text documents, text-image pairs, or text-only video descriptions. The experiments compare six baselines and report VideoRAG as the best on ROUGE-L, BLEU-4, BERTScore, and G-Eval across three LVLMs, with an oracle setting showing that better retrieval would help further. The paper also claims its adaptive frame selection improves both retrieval and generation relative to uniform sampling, and that automatic-speech-recognition-derived transcripts stand in for missing subtitles.

Load-bearing premise

The argument assumes the reported gains come from jointly using visual and textual video content, even though the video pipeline also uses a different retrieval encoder than the text baselines and the retrieval step is never held fixed while only the generation modality changes.

Editorial extensions

If this is right

For how-to question answering, a system can answer without knowing in advance which video is relevant, because retrieval happens dynamically over the video corpus.
Answers grounded in video frames plus transcripts surpass answers grounded in transcripts alone, so video corpora should be treated as first-class RAG knowledge sources.
Selecting a small number of informative frames is enough to beat uniform sampling, which lowers the cost of processing long videos.
Videos without subtitles can still contribute textual signal by running automatic speech recognition, making the approach applicable to raw video collections.
The method scales with LVLM size and holds across query categories, so larger video-language models are a direct lever for further gains.

Reading between the lines

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

The paper's evaluation is instructional; the same pipeline may transfer to news, sports, or documentary video, where the visual signal carries information absent from transcripts, but that transfer is not tested here.
Because the baselines and VideoRAG differ in both retrieval encoder and generation input, part of the reported gap could come from retrieval quality rather than joint visual-textual generation; a retrieval-held-fixed ablation would settle this.
The frame-selection scorer is trained on downstream retrieval and generation success labels, so it could be retrained for other video tasks, such as video grounding or long-video QA, with the same label-generation recipe.
A direct prediction of the visual-value claim is that on queries whose answers live only in the image, such as how a dog's expression changes, the gap over transcript-only RAG should be larger than on text-redundant queries.

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

VideoRAG is a solid systems paper and worth a real referee. It does something new: it puts together corpus-level video retrieval, LVLM-based generation over frames and transcripts, and a learned frame selector, and it ships code. The gains over text-only and text-image RAG are consistent across three LVLMs and several metrics, and the small human evaluation agrees. That is real evidence.

The paper also does a few things right. It is clear about the components, uses off-the-shelf encoders and LVLMs, and honestly reports retrieval quality by modality (Table 2). The ASR fallback for missing subtitles is a practical, sensible addition. The frame-selection ablation shows adaptive selection beats uniform sampling.

The soft spot is the one the stress-test flags, and I think it is legitimate. The main comparison changes two things at once: retrieval pipeline and generation input. TextVideoRAG retrieves with text-only indexing and generates from transcripts; VideoRAG retrieves with InternVideo2 plus adaptive frame selection and generates from frames. Table 2 shows retrieval quality varies a lot by feature type (textual R@1 0.088, visual 0.054, ensemble 0.103), so the two conditions are starting from different videos. There is no condition that holds retrieval fixed and varies only whether generation consumes frames or transcripts. That means the reported gains could be driven by better retrieval rather than by visual grounding, which is the paper's specific novelty claim. This does not kill the paper—the system still appears to work better than baselines—but it does mean the headline interpretation is not fully supported.

A related smaller issue: the frame selector for generation is trained to optimize ROUGE-L, and the retrieval selector is trained on retrieval similarity, both of which are used as evaluation metrics. That is a mild circularity for the frame-selection component, not for the central claim, and it is worth noting but not fatal.

The lack of error bars or significance tests is also a weakness, but the consistency across three LVLMs mitigates it somewhat.

Bottom line: this is a useful contribution to multimodal RAG, the code is out, and the questions I raised are answerable with one controlled experiment plus standard errors. I would send it to review and ask for that experiment. A reader who works on RAG or video understanding will get value from it.

Referee Report

3 major / 5 minor

Summary. The paper introduces VideoRAG, a retrieval-augmented generation framework that operates over a video corpus. For a given query, it retrieves videos using LVLM-based embeddings that combine visual frames (selected by a trained adaptive frame selector) and textual transcripts (original subtitles or ASR-generated auxiliary text), then feeds the retrieved videos jointly with the query into an LVLM to generate answers. The authors construct evaluation sets from WikiHowQA with HowTo100M and a synthetic QA dataset, and compare against textual, text-image, and text-video RAG baselines across three LVLMs (LLaVA-Video, InternVL2.5, Qwen2.5-VL). Results in Table 1 show VideoRAG outperforming all baselines on ROUGE-L, BLEU-4, BERTScore, and G-Eval, with a supporting human evaluation. The paper also reports retrieval ablations, frame-selection comparisons, and category-wise analyses.

Significance. If the reported gains are attributable to the joint use of visual and textual video content, this is a valuable step toward extending RAG to video corpora, which are rich but underexplored as knowledge sources. The paper ships code, evaluates across multiple LVLMs and metrics, includes a human study corroborating the automatic results, and provides detailed ablations of retrieval modalities and frame-selection strategies. These are concrete strengths. However, the central claim that visual grounding drives the improvement is not yet isolated from retrieval-quality differences, so the significance of the specific mechanism remains conditional on additional controlled experiments.

major comments (3)

[Section 3.2, Table 1] The main comparison between VideoRAG and TEXTVIDEORAG changes two factors simultaneously: the retriever (InternVideo2 with ensemble features and adaptive frame selection versus text-only indexing/encoding) and the generation input modality (frames versus text-only). Because Table 2 shows that the ensemble retriever has higher recall (R@1 0.103) than textual features alone (R@1 0.088), the improved generation results may be driven by retrieving more relevant videos rather than by the visual grounding that the paper claims as its contribution. Without a condition that holds retrieval fixed while varying the generation modality (e.g., generating from the same top-1 video with text-only versus frames-only versus frames+text), the central claim about visual video information is not isolated.
[Section 3.1 / Table 1] The paper repeatedly uses the word 'significantly' to describe the improvements of VideoRAG over baselines, but no error bars, confidence intervals, or significance tests are reported. Several gaps are small (e.g., for LLaVA-Video, VideoRAG-V and VideoRAG-VT differ by 0.02 ROUGE-L on WikiHowQA), and the magnitudes vary across LVLMs and datasets. The authors should report variances (e.g., bootstrap confidence intervals or multiple random seeds) and perform paired significance tests for the key pairwise comparisons to support the claimed significance.
[Section 2.3 / Appendix A.3, Table 3] The adaptive frame selectors are trained with labels derived from the same metrics used in evaluation: the retrieval selector uses retrieval-similarity-based labels and the generation selector uses ROUGE-L-based labels, while the reported improvements (Table 3) are measured on retrieval recall and ROUGE-L. This creates a risk that the gains over uniform sampling reflect direct optimization of the evaluation objectives rather than generally informative frame selection. The authors should either evaluate the selected frames on a metric not used for training (e.g., human ratings or G-Eval conditioned on frame choice) or discuss this potential inflation and provide evidence that the selection transfers to held-out metrics.

minor comments (5)

[Section 3.1, Datasets] It is unclear how the WikiHowQA queries are paired with HowTo100M videos for the Oracle condition and how the retrieval corpus is constructed; please clarify the association procedure and whether the corpus contains all videos or a filtered subset.
[Section 2.3, Eq. (1)-(2)] The notation Comb(·) is described as randomly sampling a subset of m frames from n, but the method first applies k-means++ clustering to reduce the frame space; please clarify how Comb operates on the reduced candidate set and whether the clustering step is included in the theoretical formulation.
[Appendix A.3, Generation frame selection] The paper states that 40 random subsets are sampled from the 64C32 combinations, but no rationale is given for this number; a short justification or sensitivity analysis would be helpful.
[Table 6] The human evaluation description does not specify how many evaluators assessed each query, how individual Likert scores were aggregated into the reported mean, or how the inter-annotator correlation was computed; please provide this detail.
[Figure 4] The x-axis label appears truncated as 'Combination Ratio ()'; the symbol should be written as α and defined in the caption.

Circularity Check

2 steps flagged · score 4.0 of 10

Frame-selection improvements in Table 3 are trained on the exact metrics used to report them (cosine similarity for R@1, ROUGE-L for generation), making those component-level gains partly by construction; the central VideoRAG-vs-baseline comparison is independent and not circular.

fitted input called prediction [Appendix A.3 'Retrieval'; Sec. 3.2 'Effectiveness of Frame Selection' (Table 3)]
"given the pair of the query and its relevant video, we sample multiple sets of 4 frames, and measure their similarities with the given query, so that we label the top 3 combinations with the highest similarities as True and the bottom 3 combinations as False. Then, the network is optimized via cross-entropy loss based on these labels."

The retrieval scorer f is trained to label as True the frame subsets that maximize the same cosine similarity that the retriever uses for ranking and that defines R@1. Table 3 then reports that adaptive selection raises R@1 over uniform sampling (visual 0.079 vs 0.054; ensemble 0.118 vs 0.097). The reported retrieval gain is the training objective itself, so the improvement is forced by construction rather than an independent test of frame informativeness.
fitted input called prediction [Appendix A.3 'Generation'; Sec. 3.2 'Effectiveness of Frame Selection' (Table 3)]
"we automatically collect the training dataset by labeling the top 3 combinations with the highest ROUGE-L scores as True and the bottom 3 with the lowest scores as False, according to their ROUGE-L score and with the LLaV A-Video (7B) as the LVLM for generation."

The generation selector is optimized to pick the 32-frame subsets whose generated responses have the highest ROUGE-L. Table 3 then evaluates the selector with ROUGE-L (adaptive 23.24 vs uniform 21.04) as evidence that selected frames are 'more useful.' Since the selector was fit to ROUGE-L, the ROUGE-L advantage is the training signal reported back as a result; this is a fitted input presented as a predicted improvement.

full rationale

The paper's main claim—that VideoRAG retrieves videos and uses frames plus transcripts to outperform text, text-image, and text-video RAG—rests on a pipeline built from off-the-shelf encoders (InternVideo2, CLIP) and LVLMs (LLaVA-Video, InternVL2.5, Qwen2.5-VL). No load-bearing result is imported from the authors' own prior work; citations to Adaptive-RAG and multimodal interleaved retrieval are contextual, not computational. The comparison against TextVideoRAG is confounded because the systems differ in both retriever and consumed modality, but confounding is an experimental-design threat, not a circular reduction, so it is not scored here. The concrete circularity is confined to the frame-selection analysis: Appendix A.3 trains the retrieval selector on cosine-similarity labels and the generation selector on ROUGE-L labels, and Section 3.2 (Table 3) then reports improvements in R@1 and ROUGE-L as evidence of effectiveness. Those gains are partly the training objective, so the frame-selection claim reduces to a fit. This does not infect the central RAG comparison, which is a separate empirical claim requiring the retrieval-quality confound to be addressed as a correctness issue rather than as circularity. Hence a moderate score of 4.

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

The paper introduces no new physical entities, forces, or dimensions. Its contributions are software components, a frame-selection scoring function and an assembled RAG pipeline, which are not the kind of postulated entities tracked in this ledger. The central claim rests mainly on data assumptions about the evaluation datasets and on the reliability of the off-the-shelf encoders, LVLMs, and ASR.

free parameters (5)

Frame counts for retrieval and generation = 4 for retrieval, 32 for generation
Chosen by hand for LVLM context limits and efficiency (Section 3.1, Appendix A.3); these values affect both retrieval quality and generation quality.
k-means++ cluster counts = 8 for retrieval, 64 for generation
Hand-chosen to reduce the frame selection search space (Section 2.3, Appendix A.3).
Ensemble interpolation ratio alpha = about 0.5 to 0.7
Tuned on retrieval data (Figure 4) to balance text and visual features; the paper reports the optimal range but no held-out tuning procedure.
Number of random frame subsets sampled for generation selection = 40
Hand-chosen approximation to the 64C32 combination space (Appendix A.3).
Number of retrieved videos used for generation = 1
Chosen because the authors observed little difference with more videos, but no supporting evidence is shown (Appendix A.1).

assumptions (5)

domain assumption InternVideo2 video and text encoders produce embeddings whose similarity reflects query-video relevance.
Used as the retriever in Section 2.2 and Appendix A.1; retrieval quality is a core component of the pipeline.
domain assumption HowTo100M query-video pairs are valid ground truth for relevance, and the WikiHowQA overlap yields valid query-answer-video triples.
Section 3.1; used to train frame selection, evaluate retrieval, and define the Oracle setting.
domain assumption ROUGE-L, BLEU-4, BERTScore, and G-Eval are adequate measures of answer quality for instructional QA.
Section 3.1 metrics; used for main comparisons and for labeling frame-selection training data.
domain assumption LVLMs can generate correct answers from 32 frames plus the query.
Section 2.2 and 3.1; the generation pipeline assumes the selected frames contain sufficient information.
domain assumption Whisper ASR produces transcripts that capture the textual content needed for retrieval and generation.
Section 2.4; used to supply text for videos without subtitles.

how reviews work

0 comments

Cite this review

Pith. "Pith review of VideoRAG: Retrieval-Augmented Generation over Video Corpus." pith.science (2026). https://pith.science/paper/ZVYJABDV

@misc{pith2026250105874,
  author       = {Pith},
  title        = {Pith review of: VideoRAG: Retrieval-Augmented Generation over Video Corpus},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZVYJABDV}},
  note         = {Machine review of arXiv:2501.05874}
}

read the original abstract

Retrieval-Augmented Generation (RAG) is a powerful strategy for improving the factual accuracy of models by retrieving external knowledge relevant to queries and incorporating it into the generation process. However, existing approaches primarily focus on text, with some recent advancements considering images, and they largely overlook videos, a rich source of multimodal knowledge capable of representing contextual details more effectively than any other modality. While very recent studies explore the use of videos in response generation, they either predefine query-associated videos without retrieval or convert videos into textual descriptions losing multimodal richness. To tackle these, we introduce VideoRAG, a framework that not only dynamically retrieves videos based on their relevance with queries but also utilizes both visual and textual information. The operation of VideoRAG is powered by recent Large Video Language Models (LVLMs), which enable the direct processing of video content to represent it for retrieval and the seamless integration of retrieved videos jointly with queries for response generation. Also, inspired by that the context size of LVLMs may not be sufficient to process all frames in extremely long videos and not all frames are equally important, we introduce a video frame selection mechanism to extract the most informative subset of frames, along with a strategy to extract textual information from videos (as it can aid the understanding of video content) when their subtitles are not available. We experimentally validate the effectiveness of VideoRAG, showcasing that it is superior to relevant baselines. Code is available at https://github.com/starsuzi/VideoRAG.

Figures

Figures reproduced from arXiv: 2501.05874 by the authors.

**Figure 2.** Illustration of the overall pipeline of our VideoRAG, which selects informative frames for retrieval and generation. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Visualization of latent space of features across modalities with Principal Component Analysis (PCA). 1.0 0.8 0.6 0.4 0.2 0.0 Combination Ratio ( ) 0.00 0.05 0.10 0.15 Recall Improvement Impact of Combination Ratio R@1 R@5 R@10 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 6.** Figure 6: Breakdown performance of different models across 10 categories. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 6 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

RAVID: Retrieval-Augmented Visual Detection: A Knowledge-Driven Approach for AI-Generated Image Identification
cs.CV 2025-08 conditional novelty 6.0 of 10

RAVID detects AI-generated images by retrieving similar images from a database and feeding them to a vision-language model, reporting 93.85% average accuracy on UniversalFakeDetect.
ReFoCUS: Reinforcement-guided Frame Optimization for Contextual Understanding
cs.CV 2025-06 conditional novelty 6.0 of 10

A reinforcement-learned frame selection policy, trained with reward margins from a reference video-LLM, improves video QA accuracy of LLaVA-OV and InternVL3 across several benchmarks.
Magic-MM-Embedding: Towards Visual-Token-Efficient Universal Multimodal Embedding with MLLMs
cs.CV 2026-02 conditional novelty 5.0 of 10

Visual token compression (4x fewer tokens) plus a three-stage generative/contrastive/judge-curated training pipeline yields state-of-the-art MLLM-based retrieval accuracy at lower inference cost.
Track and Caption Any Motion: Open-Vocabulary Spatiotemporal Captioning via Trajectory-Conditioned Generation
cs.CV 2025-12 conditional novelty 5.0 of 10

A trajectory-conditioned retrieval system discovers multiple motion descriptions in videos without user queries and grounds them to point tracks, evaluated mainly on MeViS.
SceneRAG: Scene-level Retrieval-Augmented Generation for Video Understanding
cs.CV 2025-06 conditional novelty 5.0 of 10

SceneRAG uses LLM-driven scene segmentation and a scene-level knowledge graph to retrieve and answer questions about long videos, reporting higher LLM-judged win-rates than chunk-based RAG baselines on the LongerVideo...
ReAgent-V: A Reward-Driven Multi-Agent Framework for Video Understanding
cs.CV 2025-06 conditional novelty 5.0 of 10

ReAgent-V is an agentic video understanding framework whose critic agent generates real-time rewards to refine answers and filter training data, yielding gains of up to 6.9%, 2.1%, and 9.8% across three applications.

Reference graph

Works this paper leans on

17 extracted references · 13 canonical work pages · cited by 6 Pith papers

[1]

Gather materials: You will need clay, a potter’s wheel, a wire cutter, a sponge, and any additional materials you want to use for decoration (such as glaze, paint, or glitter)

work page
[2]

arXiv Preprint arXiv:2406.11931, abs/2406.11931

Deepseek-coder-v2: Breaking the barrier of closed-source models in code intelligence. arXiv Preprint arXiv:2406.11931, abs/2406.11931. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language under- standing. In Proceedings of the 2019 Conference of the North American Chapte...

arXiv 2019
[3]

arXiv preprint arXiv:2406.12846 , abs/2406.12846

Drvideo: Document retrieval based long video understanding. arXiv preprint arXiv:2406.12846 , abs/2406.12846. Muhammad Maaz, Hanoona Abdul Rasheed, Salman Khan, and Fahad Khan. 2024. Video-chatgpt: To- wards detailed video understanding via large vision and language models. In Proceedings of the 62nd Annual Meeting of the Association for Computa- tional L...

arXiv 2024
[4]

You can also use additional materials like glaze, paint, or glitter to decorate the rose

Add details: Use a tool or your fingers to add details to the petals, such as veins or shading. You can also use additional materials like glaze, paint, or glitter to decorate the rose

work page
[5]

Ragtruth: A hallucination corpus for develop- ing trustworthy retrieval-augmented language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024 , pages 10862–10878. Association for Computational Linguistics. OpenAI. 2023. GPT-4 technical...

arXiv 2024
[6]

This will harden the clay and create a durable piece of pottery

Fire the clay: Place the dried clay rose in a kiln and fire it according to the manufacturer’s instructions. This will harden the clay and create a durable piece of pottery

work page
[7]

question

Glaze or finish: Once the clay is fired, you can apply a glaze or finish to give the rose a glossy or matte appearance. Allow the glaze to dry and cure before handling the finished piece. It’s important to note that this is a general guide and the specific steps may vary depending on the type of clay you are using and the techniques you are employing. It’...

work page
[9]

Use a wire cutter to trim any excess clay and smooth out any rough edges

Prepare the clay: Cut a piece of clay to the desired size and shape for your rose. Use a wire cutter to trim any excess clay and smooth out any rough edges

work page

Show all 17 references

[10]

You can start by creating a small ball of clay for the center of the rose, then add petals around it

Shape the clay: Use your hands or tools to shape the clay into a rose. You can start by creating a small ball of clay for the center of the rose, then add petals around it. Use a sponge to smooth out any rough areas and refine the shape of the rose
[12]

This will help prevent any cracking or breaking during the firing process

Allow the clay to dry: Let the clay dry completely before firing it in a kiln. This will help prevent any cracking or breaking during the firing process
[15]

Carefully read the Ground Truth and the Generated Response
[16]

Please penalize the Generated Response that has the far different content and style and is largely longer than the Ground Truth

Assess how well the Generated Response matches the Ground Truth. Please penalize the Generated Response that has the far different content and style and is largely longer than the Ground Truth
[17]

Provide an overall score (1-5) based on your evaluation. Question: {{Question}} Ground Truth Answer: {{Ground_Truth_Answer}} Generated Response: {{Generated_Response}} Please provide only a single numerical rating (1, 2, 3, 4, or 5), without any additional commentary, formatti...
[126]

what shapes your bias?

National Institute of Standards and Technology (NIST). Jisu Shin, Hoyun Song, Huije Lee, Soyeong Jeong, and Jong Park. 2024. Ask llms directly, "what shapes your bias?": Measuring social bias in large language models. In Findings of the Association for Computa- tional Linguist...

2024
[1994]

In Proceedings of The Third Text REtrieval Conference, TREC 1994, Gaithers- burg, Maryland, USA, November 2-4, 1994 , volume 500-225 of NIST Special Publication , pages 109–

Okapi at TREC-3. In Proceedings of The Third Text REtrieval Conference, TREC 1994, Gaithers- burg, Maryland, USA, November 2-4, 1994 , volume 500-225 of NIST Special Publication , pages 109–

1994
[2019]

In 2019 IEEE/CVF International Confer- ence on Computer Vision, ICCV 2019, Seoul, Ko- rea (South), October 27 - November 2, 2019 , pages 2630–2640

Howto100m: Learning a text-video embed- ding by watching hundred million narrated video clips. In 2019 IEEE/CVF International Confer- ence on Computer Vision, ICCV 2019, Seoul, Ko- rea (South), October 27 - November 2, 2019 , pages 2630–2640. IEEE. Cheng Niu, Yuanhao Wu, Juno ...

2019
[2024]

arXiv Preprint arXiv:2409.11402, abs/2409.11402

NVLM: open frontier-class multimodal llms. arXiv Preprint arXiv:2409.11402, abs/2409.11402. DeepSeek-AI, Qihao Zhu, Daya Guo, Zhihong Shao, Dejian Yang, Peiyi Wang, Runxin Xu, Y . Wu, Yukun Li, Huazuo Gao, Shirong Ma, Wangding Zeng, Xiao Bi, Zihui Gu, Hanwei Xu, Damai Dai, Kai...

arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

Gather materials: You will need clay, a potter’s wheel, a wire cutter, a sponge, and any additional materials you want to use for decoration (such as glaze, paint, or glitter)

work page

[2] [2]

arXiv Preprint arXiv:2406.11931, abs/2406.11931

Deepseek-coder-v2: Breaking the barrier of closed-source models in code intelligence. arXiv Preprint arXiv:2406.11931, abs/2406.11931. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language under- standing. In Proceedings of the 2019 Conference of the North American Chapte...

arXiv 2019

[3] [3]

arXiv preprint arXiv:2406.12846 , abs/2406.12846

Drvideo: Document retrieval based long video understanding. arXiv preprint arXiv:2406.12846 , abs/2406.12846. Muhammad Maaz, Hanoona Abdul Rasheed, Salman Khan, and Fahad Khan. 2024. Video-chatgpt: To- wards detailed video understanding via large vision and language models. In Proceedings of the 62nd Annual Meeting of the Association for Computa- tional L...

arXiv 2024

[4] [4]

You can also use additional materials like glaze, paint, or glitter to decorate the rose

Add details: Use a tool or your fingers to add details to the petals, such as veins or shading. You can also use additional materials like glaze, paint, or glitter to decorate the rose

work page

[5] [5]

Ragtruth: A hallucination corpus for develop- ing trustworthy retrieval-augmented language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024 , pages 10862–10878. Association for Computational Linguistics. OpenAI. 2023. GPT-4 technical...

arXiv 2024

[6] [6]

This will harden the clay and create a durable piece of pottery

Fire the clay: Place the dried clay rose in a kiln and fire it according to the manufacturer’s instructions. This will harden the clay and create a durable piece of pottery

work page

[7] [7]

question

Glaze or finish: Once the clay is fired, you can apply a glaze or finish to give the rose a glossy or matte appearance. Allow the glaze to dry and cure before handling the finished piece. It’s important to note that this is a general guide and the specific steps may vary depending on the type of clay you are using and the techniques you are employing. It’...

work page

[8] [9]

Use a wire cutter to trim any excess clay and smooth out any rough edges

Prepare the clay: Cut a piece of clay to the desired size and shape for your rose. Use a wire cutter to trim any excess clay and smooth out any rough edges

work page