Pith. sign in

REVIEW 4 major objections 6 minor 46 references

FIFA: Unified Faithfulness Evaluation Framework for Text-to-Video and Video-to-Text Generation

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read FIFA claims that one reference-free framework can score faithfulness and correct hallucinations in both text-to-video and video-to-text outputs.

desk verdict FIFA is a credible extension of fact-decomposition faithfulness evaluation to video, but its headline correlation rests on 120 samples and an 80%-accurate verifier, and the post-correction numbers are partly circular. read the letter →

arxiv 2507.06523 v1 pith:OBEFM6XH submitted 2025-07-09 cs.CV cs.CLcs.GR

Liqiang Jing , Viet Lai , Seunghyun Yoon , Trung Bui , Xinya Du This is my paper
classification cs.CVcs.CLcs.GR
keywords faithfulnessevaluationhallucinationdetectionvideo-to-textgenerationtext-to-videovideoquestionansweringsemanticdependencygraphfactverificationpost-correction
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 claims that hallucination in video-to-text and text-to-video generation can be caught by one reference-free framework, FIFA. FIFA turns generated text into descriptive facts, wires those facts into a dependency graph, and asks a video-question-answering model a yes/no question for each one. When a fact depends on a false prerequisite, the graph marks all downstream facts wrong too, which fixes errors that independent fact checks miss, like attributing a red shirt and a blue hat to the wrong person. The paper reports that FIFA correlates more strongly with human faithfulness ratings than existing metrics on both tasks, and that the same verification answers can drive a Post-Correction step that edits hallucinated text and video. If these claims hold, open-ended video-language outputs can be automatically audited and repaired at fine granularity.

What carries the argument

The load-bearing object is the Spatio-Temporal Semantic Dependency Graph (STSDG), a directed acyclic graph whose nodes are yes/no questions derived from extracted facts and whose edges encode that one question is only meaningful if another is answered yes. Facts come in two granularities that together give full semantic coverage: atomic facts, the smallest indivisible units such as entity, attribute, relation, and scene facts, and event-level facts, composite facts that bind several atomic facts into one high-level claim. Verification uses a VideoQA model to answer every question from the video, and score aggregation sets the refined score of a fact to the indicator of a yes answer times the product of its dependency parents' scores, so a false prerequisite propagates a zero to all facts that depend on it.

What would settle it

Take a held-out set of video-text pairs in which every extracted fact has a human gold yes/no answer, run FIFA with the best VideoQA verifier, and compare FIFA's per-fact verdicts with the human majority; if agreement is no better than the verifier's standalone accuracy of about 80%, the claimed alignment with human judgment is falsified.

Watch

Extended reading notes

Core claim

The central claim is that faithfulness of a video-text pair can be estimated reference-free by decomposing the text into atomic facts (entities, attributes, relations, scenes) and event-level facts, translating each fact into a yes/no question, and verifying those questions with a VideoQA model. The new element is the Spatio-Temporal Semantic Dependency Graph: a directed acyclic graph over the questions that encodes prerequisite structure, so the refined score of a fact is zero whenever any fact it depends on is judged false. The paper reports the highest correlation with human judgment among the compared metrics (Pearson r of 58.20 on V2T and 67.92 on T2V), near-perfect human agreement on fact extraction, fact-to-question conversion, and dependency generation, and identifies the VideoQA verifier as the accuracy bottleneck, with the best model reaching about 80% per-fact accuracy.

Load-bearing premise

The VideoQA verifier answers the generated fact-questions correctly; since the strongest verifier in the paper is about 80% accurate per fact, every FIFA score inherits an error floor of roughly one wrong fact-verdict in five.

Editorial extensions

If this is right

  • The same FIFA pipeline applies to both V2T and T2V outputs, so one evaluation suite can replace task-specific metrics.
  • Dependency-aware aggregation guarantees that a fact is never counted as faithful when the fact it presupposes is judged false.
  • Post-Correction improves the faithfulness of generated text and video across several video-language models, measured by FIFA and by COAHA.
  • Temporal relations and event-level facts show the lowest faithfulness scores, pointing to temporal reasoning as the main weakness of current video generation models.
  • Because ablation shows better verifiers yield better correlation, the framework's alignment with human judgment scales with VideoQA accuracy.

Reading between the lines

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

  • Editorial inference: the roughly 80% verifier accuracy caps FIFA's per-fact reliability, so reported FIFA scores are best read with a per-fact error band of about one in five, and improving the verifier is the most direct route to a better metric.
  • Editorial inference: the STSDG construction could transfer to other multimodal faithfulness settings, such as text-to-image evaluation or audio-visual grounding, which the paper does not test.
  • Editorial inference: Post-Correction is bounded by the verifier's recall; a false 'no' answer could prompt the corrector to edit content that is actually faithful, so correction quality depends on verifier calibration.
  • Editorial inference: event-level facts look like a general remedy for cross-claim attribution errors, which also appear in long-form text summarization when individually true sentences combine into a false composite claim.
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

4 major / 6 minor

Summary. The paper proposes FIFA, a unified, reference-free faithfulness evaluation framework for both Video-to-Text (V2T) and Text-to-Video (T2V) outputs. FIFA extracts atomic and event-level descriptive facts from text, converts them into yes/no questions, constructs a Spatio-Temporal Semantic Dependency Graph (STSDG), verifies the questions with VideoQA models, and aggregates scores with dependency-based propagation. The paper also introduces Post-Correction, which uses FIFA's intermediate QA pairs to revise hallucinated text or video content. The main empirical claims are that FIFA correlates with human faithfulness judgments more strongly than existing metrics (Table 1), that its intermediate stages are reliable (Tables 2 and 3), and that Post-Correction improves factual consistency (Table 4). The code and human-annotated data are released.

Significance. If the headline correlation result is robust, FIFA would be a genuinely useful contribution: it is one of the first unified metrics to address both V2T and T2V hallucination, it targets open-ended free-form responses rather than binary QA, and it introduces a dependency structure that goes beyond independent fact-checking. The paper also provides valuable intermediate human evaluations for fact extraction, question conversion, dependency validity, and VideoQA accuracy, and it releases code. These are concrete strengths: the component-wise human checks give the pipeline a level of transparency that many evaluation papers lack. However, the meta-evaluation is small, the verifier has a substantial per-fact error rate, and the correction results are evaluated with the same oracle that drives the corrections. These issues leave the central superiority and correction claims not fully established.

major comments (4)
  1. [4.1, Table 1] The meta-evaluation is based on only 120 samples (60 per task), yet Table 1 reports correlation coefficients without confidence intervals or a description of the significance test. For example, on V2T the difference between FIFA (Pearson r=58.20) and METEOR (45.74) may not be significant at n=60 once sampling variability is accounted for. I ask the authors to report bootstrapped confidence intervals for r, rho, and tau, and to document the exact significance test used for the claimed p<0.01. This is load-bearing because the central claim is that FIFA aligns more closely with human judgment than all baselines.
  2. [3.3, Table 3] The best VideoQA verifier, Qwen2.5-VL-72b, achieves only 80.0-80.1% accuracy on the fact-check questions, meaning roughly one in five binary fact labels is wrong. Since each per-fact score s_i is a binary indicator derived from the verifier, the final FIFA score inherits this error floor. Equation (4) can only zero out dependents of false prerequisites; it cannot correct false positives or false negatives. Moreover, a verifier error on a prerequisite can propagate through the product to all downstream facts. The paper's defense in Section 3.3 that discriminative QA 'tends to generate low hallucination' is contradicted by the paper's own measurement. I ask for a sensitivity analysis: for example, recompute Table 1 correlations with an oracle verifier (human labels on the same questions), or with simulated verifier noise, to show that the observed human-alignment advantage is not an artifact of verifier error patterns.
  3. [5, Table 4] Post-Correction improvements are measured with the same FIFA framework that generates the corrections: the QA pairs produced by the VideoQA model are used both to instruct the corrector and to compute the post-correction FIFA scores. For T2V, Table 4 reports only FIFA as the evaluation metric, so the reported improvement from 54.53 to 60.70 is partly by construction. For V2T, COAHA provides some independent signal, but the primary metric is again FIFA. To establish that Post-Correction genuinely reduces hallucinations, the corrected outputs should be evaluated by human annotators or by an independently developed faithfulness metric.
  4. [3.4, Eq. (4)] Equation (4) defines the refined score of fact i as the raw verifier output I(a_i=yes) multiplied by the raw scores s_j of its direct prerequisites. If dependencies are chained (e.g., q_k depends on q_j depends on q_i), a false prerequisite that is not a direct prerequisite is not propagated unless the dependency generator happens to include transitive edges. The paper's stated guarantee in Section 3.2 and the STSDG description in Section 1 say that 'if the answer to a prerequisite question is negative, all downstream questions that depend on it are skipped.' That guarantee requires either transitive closure of dependencies or computing Eq. (4) in topological order over previously refined scores. As written, a verifier false positive on an intermediate prerequisite can leave downstream facts scored as faithful even when their ultimate prerequisite is false. Please clarify whether the dependency annotations include transitive edges, or modify Eq. (4) to use refined scores.
minor comments (6)
  1. [4.2, Table 2] The precision and recall definitions for question generation are written as 'P mt,q/|Q|' and 'P mt,q/|T|', which is mathematically unclear. Please define them with explicit sums, e.g., precision = (1/|Q|) * sum_q max_t m_{t,q}, and state how partial matches are handled.
  2. [Appendix A, Table 5] The ablation table reports a single FIFA row (63.06 Pearson r), but Table 1 reports separate V2T and T2V correlations. Please clarify whether Table 5 averages across tasks or pools samples, and report the task-wise ablation as well.
  3. [3.4] The notation for s_j in Eq. (4) should be made explicit: s_j is the raw score before refinement, not the refined score. Please also state the intended topological processing order, since the graph is a DAG.
  4. [Appendix I.4] The dependency examples appear to list only direct parent edges; for instance, in the 'a sad man with green hair' example, fact 5 depends on facts 3 and 4, but not on fact 1 or fact 2 which are the ultimate prerequisites. If the intended semantics are transitive dependencies, the examples should be updated; if direct dependencies are intended, the aggregation must propagate through refined scores as noted in major comment 4.
  5. [Throughout] There are numerous typos and formatting artifacts, especially in the appendix and figures: 'decsript' in Figure 1, 'V olcano' in the references, 'ambigity', 'stadning', 'crossrodas', 'expersso', 'Des', 'traffice', 'picece', 'in front off', and 'roadsid' in the in-context examples. A careful proofreading pass is needed.
  6. [4.2] The subsection heading 'Can the tuple be transferred into independent questions correctly?' uses 'tuple' where 'fact' is meant, and the word 'independent' is misleading because the whole framework is built on interdependencies between questions. Please rephrase.

Circularity Check

1 steps flagged · score 6.0 of 10

Post-Correction effectiveness is scored by the same VideoQA model that produces its corrective signals, so the T2V improvement in Table 4 is partly by construction; the FIFA-human correlation itself is externally grounded.

  1. self definitional [Section 5 (Post-Correction), Table 4; reuses Eq. 3 of Section 3.3]
    "2) Claim Verification. Then, we ask a series of questions around them to make the hallucination diagnosis following operations in Eq. 3. For all questions, we apply a VideoQA model to answer the questions conditioned on the video. The first two stages are the intermediate process in our FIFA. ... We use Qwen2.5-VL-72b as the VideoQA model and TokenFlow as the video editing model in our Post-Correction method. Table 4 shows the performance of all the baselines without and with our correction method."

    Post-Correction generates its edits from the answers of Qwen2.5-VL-72b (via Eq. 3), and then Table 4 evaluates those corrected outputs with FIFA, whose score is the average of Eq. 3 answers from the same Qwen2.5-VL-72b. The edit and the outcome measure share the same verifier, so an edit that aligns the output with the verifier's own yes/no answers mechanically raises the FIFA score. For the T2V rows only FIFA is reported (54.53 to 60.70), so the claim that correction improves video factual consistency has no independent measurement; it is the verifier confirming its own corrections. The V2T COAHA column is an external check, but the unified 'both text and video generation' claim rests on the self-scored T2V result.

full rationale

The core FIFA metric is not circular: its headline claim is correlation with human faithfulness judgments (Table 1), and those judgments come from external annotators with a reported Fleiss kappa; the metric's components (fact extraction, question generation, dependency generation) are also checked against human annotations in Table 2. Neither the score definition nor the correlation depends on the paper's own outputs. The verifier accuracy floor in Table 3 (Qwen2.5-VL-72b at about 80%) is a real correctness risk for FIFA as an absolute measure, but it is not circularity because the verifier is an external model and its error rate is measured against human labels. The one load-bearing circular step is Post-Correction: the same VideoQA model both supplies the fact-diagnosis signal used to revise content and supplies the answers that FIFA averages to score the revised content. The T2V improvement claim is therefore self-confirmatory rather than evidence about actual faithfulness. This makes the correction sub-claim partially circular while the main evaluation claim remains independently grounded; hence a moderate score of 6 rather than a higher one.

Assumptions & free parameters 0 free parameters · 4 assumptions · 1 invented entities

The metric itself is not derivational; it is a pipeline of learned components. The core assumptions are that LLM fact extraction is complete, VideoQA yes/no answers are accurate verifiers, the product-rule dependency graph correctly models fact dependence, and human Likert ratings are valid ground truth. No numerical constants are fitted to data, but the dependency rule and the choice of the VideoQA oracle are consequential assumptions.

assumptions (4)
  • domain assumption LLM fact extraction recovers all and only hallucination-relevant descriptive facts from text, with no analytic or subjective content.
    Invoked in Eq. 1 and Section 3.2.1; the few-shot prompt instructs the LLM to exclude non-descriptive content, but the completeness of extraction is validated only on small human-annotated sets.
  • domain assumption VideoQA models give correct yes/no answers to fact-checking questions about video content.
    Section 3.3 states that discriminative QA has low hallucination; Table 3 shows the best model accuracy is about 80%, so the metric inherits a nontrivial error floor.
  • domain assumption If a prerequisite fact is false, all semantically dependent downstream facts are false; the product rule in Eq. 4 is the correct way to propagate this dependency.
    Section 3.4 defines the refined score as the indicator of a yes answer multiplied by the parent scores; the DAG structure and product rule are a modeling choice.
  • domain assumption Human 1-5 Likert ratings are a valid ground truth for faithfulness.
    Used in Section 4.1; only three annotators and 120 samples support the human judgments, and no inter-annotator agreement is reported for the faithfulness ratings.
invented entities (1)
  • Spatio-Temporal Semantic Dependency Graph (STSDG)
    purpose: Represents dependencies between fact-questions so that false prerequisites suppress downstream scores and prevent invalid verification.
    This is a new internal representation introduced in Section 3.2.2; validated only by human annotation of a small set of edge annotations, with no external falsifiable prediction.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FIFA: Unified Faithfulness Evaluation Framework for Text-to-Video and Video-to-Text Generation." pith.science (2026). https://pith.science/paper/OBEFM6XH

@misc{pith2026250706523,
  author       = {Pith},
  title        = {Pith review of: FIFA: Unified Faithfulness Evaluation Framework for Text-to-Video and Video-to-Text Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OBEFM6XH}},
  note         = {Machine review of arXiv:2507.06523}
}
read the original abstract

Video Multimodal Large Language Models (VideoMLLMs) have achieved remarkable progress in both Video-to-Text and Text-to-Video tasks. However, they often suffer fro hallucinations, generating content that contradicts the visual input. Existing evaluation methods are limited to one task (e.g., V2T) and also fail to assess hallucinations in open-ended, free-form responses. To address this gap, we propose FIFA, a unified FaIthFulness evAluation framework that extracts comprehensive descriptive facts, models their semantic dependencies via a Spatio-Temporal Semantic Dependency Graph, and verifies them using VideoQA models. We further introduce Post-Correction, a tool-based correction framework that revises hallucinated content. Extensive experiments demonstrate that FIFA aligns more closely with human judgment than existing evaluation methods, and that Post-Correction effectively improves factual consistency in both text and video generation.

Figures

Figures reproduced from arXiv: 2507.06523 by the authors.

Figure 1
Figure 1. a) Illustration of the V2T and T2V tasks. The content in red font denotes hallucinated content. b) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of our proposed FIFA metric. The blue arrows represent the information flow for T2V, the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Illustration of STSDG-based Question Gener [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: The proposed Post-Correction method consists of three key stages: Claim Extraction, Claim Verification, [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: UI for faithfulness evaluation of human annotation. [PITH_FULL_IMAGE:figures/full_fig_p031_5.png]
Figure 6
Figure 6. Figure 6: UI for question quality evaluation of human annotation. [PITH_FULL_IMAGE:figures/full_fig_p032_6.png]
Figure 7
Figure 7. Figure 7: UI for dependency verification of human annotation. [PITH_FULL_IMAGE:figures/full_fig_p033_7.png]
Figure 8
Figure 8. Figure 8: UI for the fact-to-question task of human evaluation. [PITH_FULL_IMAGE:figures/full_fig_p034_8.png]
Figure 9
Figure 9. Figure 9: UI for Video Question Answering of human annotation. [PITH_FULL_IMAGE:figures/full_fig_p035_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 13 canonical work pages

  1. [1]

    URL: " 'urlintro :=

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. 2025. Qwen2.5-vl technical rep...

    arXiv 2025

  4. [4]

    Satanjeev Banerjee and Alon Lavie. 2005. METEOR: an automatic metric for MT evaluation with improved correlation with human judgments. In Summarization@ACL, pages 65--72. ACL

    work page 2005

  5. [5]

    Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. 2024. Video generation models as world simulators. OpenAI Blog, 1:8

    work page 2024

  6. [7]

    Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. 2024 b . Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271

    arXiv 2024

  7. [8]

    Baldridge, Roopal Garg, Peter Anderson, Ranjay Krishna, Mohit Bansal, Jordi Pont - Tuset, and Su Wang

    Jaemin Cho, Yushi Hu, Jason M. Baldridge, Roopal Garg, Peter Anderson, Ranjay Krishna, Mohit Bansal, Jordi Pont - Tuset, and Su Wang. 2024. https://openreview.net/forum?id=ITq4ZRUT4a Davidsonian scene graph: Improving reliability in fine-grained evaluation for text-to-image generation . In The Twelfth International Conference on Learning Representations, ...

    work page 2024

  8. [9]

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. 2021. Clipscore: A reference-free evaluation metric for image captioning. In EMNLP, pages 7514--7528. ACL

    work page 2021

Show all 46 references
  1. [10]

    Gritsenko, William Chan, Mohammad Norouzi, and David J

    Jonathan Ho, Tim Salimans, Alexey A. Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. 2022. http://papers.nips.cc/paper\_files/paper/2022/hash/39235c56aef13fb05a6adc95eb9d8d66-Abstract-Conference.html Video diffusion models . In Advances in Neural Information Pro...

    2022

  2. [11]

    Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A. Smith. 2023. https://doi.org/10.1109/ICCV51070.2023.01866 TIFA: accurate and interpretable text-to-image faithfulness evaluation with question answering . In IEEE/CVF International Con...

    2023

  3. [12]

    Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. 2023. https://doi.org/10.48550/ARXIV.2311.05232 A survey on hallucination in large language models: Principles, taxonomy, challen...

    2023 doi arXiv

  4. [13]

    Liqiang Jing, Ruosen Li, Yunmo Chen, and Xinya Du. 2024. https://aclanthology.org/2024.findings-emnlp.290 Faithscore: Fine-grained evaluations of hallucinations in large vision-language models . In Findings of the Association for Computational Linguistics: EMNLP 2024, Miami, F...

    2024

  5. [14]

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, Kathrina Wu, Qin Lin, Junkun Yuan, Yanxin Long, Aladdin Wang, Andong Wang, Changlin Li, Duojun Huang, Fang Yang, Hao Tan, Hongmei Wang, Jacob Song, Jiawang Bai, J...

    2024 doi arXiv

  6. [15]

    Seongyun Lee, Sue Hyun Park, Yongrae Jo, and Minjoon Seo. 2024. https://doi.org/10.18653/V1/2024.NAACL-LONG.23 Volcano: Mitigating multimodal hallucination through self-feedback guided revision . In Proceedings of the 2024 Conference of the North American Chapter of the Associ...

    2024 doi

  7. [16]

    Chaoyu Li, Eun Woo Im, and Pooyan Fazli. 2024 a . https://doi.org/10.48550/ARXIV.2412.03735 Vidhalluc: Evaluating temporal hallucinations in multimodal large language models for video understanding . CoRR, abs/2412.03735

    2024 doi arXiv

  8. [17]

    Yanwei Li, Chengyao Wang, and Jiaya Jia. 2024 b . https://doi.org/10.1007/978-3-031-72952-2\_19 Llama-vid: An image is worth 2 tokens in large language models . In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, ...

    2024 doi

  9. [18]

    Rensis Likert. 1932. A technique for the measurement of attitudes. Archives of psychology

    1932

  10. [19]

    Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. 2024. https://aclanthology.org/2024.emnlp-main.342 Video-llava: Learning united visual representation by alignment before projection . In Proceedings of the 2024 Conference on Empirical Methods in Natural...

    2024

  11. [20]

    Chin-Yew Lin. 2004. ROUGE : A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74--81. ACL

    2004

  12. [21]

    Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang. 2024 a . https://openreview.net/forum?id=J44HfH4JCg Mitigating hallucination in large multi-modal models via robust instruction tuning . In The Twelfth International Conference on Learning Represen...

    2024

  13. [22]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual instruction tuning. CoRR, abs/2304.08485

    2023 arXiv

  14. [25]

    Muhammad Maaz, Hanoona Abdul Rasheed, Salman Khan, and Fahad Khan. 2024. https://doi.org/10.18653/V1/2024.ACL-LONG.679 Video-chatgpt: Towards detailed video understanding via large vision and language models . In Proceedings of the 62nd Annual Meeting of the Association for Co...

    2024 doi

  15. [26]

    Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen - tau Yih, Pang Wei Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2023. https://doi.org/10.18653/V1/2023.EMNLP-MAIN.741 Factscore: Fine-grained atomic evaluation of factual precision in long form text genera...

    2023 doi

  16. [27]

    Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. 2022. https://doi.org/10.1007/978-3-031-19772-7\_1 Expanding language-image pretrained models for general video recognition . In Computer Vision - ECCV 2022 - 17th E...

    2022 doi

  17. [28]

    Yingwei Pan, Zhaofan Qiu, Ting Yao, Houqiang Li, and Tao Mei. 2017. https://doi.org/10.1145/3123266.3127905 To create what you tell: Generating videos from captions . In Proceedings of the 2017 ACM on Multimedia Conference, MM 2017, Mountain View, CA, USA, October 23-27, 2017 ...

    2017

  18. [29]

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei - Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In ACL, pages 311--318. ACL

    2002

  19. [30]

    Sheth, and Amitava Das

    Vipula Rawte, Sarthak Jain, Aarush Sinha, Garv Kaushik, Aman Bansal, Prathiksha Rumale Vishwanath, Samyak Rajesh Jain, Aishwarya Naresh Reganti, Vinija Jain, Aman Chadha, Amit P. Sheth, and Amitava Das. 2024. https://doi.org/10.48550/ARXIV.2411.10867 Vibe: A text-to-video benc...

    2024 doi arXiv

  20. [31]

    Nasib Ullah and Partha Pratim Mohanta. 2022. https://doi.org/10.1007/978-3-031-26316-3\_37 Thinking hallucination for video captioning . In Computer Vision - ACCV 2022 - 16th Asian Conference on Computer Vision, Macao, China, December 4-8, 2022, Proceedings, Part IV , volume 1...

    2022 doi

  21. [32]

    A \" a ron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. 2017. https://proceedings.neurips.cc/paper/2017/hash/7a98af17e63a0ac09ce2e96d03992fbc-Abstract.html Neural discrete representation learning . In Advances in Neural Information Processing Systems 30: Annual Conferen...

    2017

  22. [33]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2023. http://arxiv.org/abs/1706.03762 Attention is all you need

    2023 arXiv

  23. [34]

    Ruben Villegas, Mohammad Babaeizadeh, Pieter - Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. 2023. https://openreview.net/forum?id=vOEXS39nOF Phenaki: Variable length video generation from open domain textua...

    2023

  24. [35]

    Yuxuan Wang, Yueqian Wang, Dongyan Zhao, Cihang Xie, and Zilong Zheng. 2024. https://doi.org/10.48550/ARXIV.2406.16338 Videohallucer: Evaluating intrinsic and extrinsic hallucinations in large video-language models . CoRR, abs/2406.16338

    2024 doi arXiv

  25. [36]

    Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. 2022 a . https://doi.org/10.1007/978-3-031-19787-1\_41 N \" u wa: Visual synthesis pre-training for neural visual world creation . In Computer Vision - ECCV 2022 - 17th European Conference, Tel ...

    2022 doi

  26. [37]

    Haoning Wu, Chaofeng Chen, Jingwen Hou, Liang Liao, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. 2022 b . https://doi.org/10.1007/978-3-031-20068-7\_31 FAST-VQA: efficient end-to-end video quality assessment with fragment sampling . In Computer Vision - ECCV 2022 - 17th E...

    2022 doi

  27. [38]

    Jay Zhangjie Wu, Guian Fang, Haoning Wu, Xintao Wang, Yixiao Ge, Xiaodong Cun, David Junhao Zhang, Jia - Wei Liu, Yuchao Gu, Rui Zhao, Weisi Lin, Wynne Hsu, Ying Shan, and Mike Zheng Shou. 2024. https://doi.org/10.48550/ARXIV.2401.07781 Towards A better metric for text-to-vide...

    2024 doi arXiv

  28. [39]

    Jun Xu, Tao Mei, Ting Yao, and Yong Rui. 2016. https://doi.org/10.1109/CVPR.2016.571 MSR-VTT: A large video description dataset for bridging video and language . In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016...

    2016 doi

  29. [40]

    Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. 2021. http://arxiv.org/abs/2104.10157 Videogpt: Video generation using VQ-VAE and transformers . CoRR, abs/2104.10157

    2021 arXiv

  30. [41]

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. 2024. https://doi.org/10.48550/ARXIV.2408....

    2024 doi arXiv

  31. [42]

    Shukang Yin, Chaoyou Fu, Sirui Zhao, Tong Xu, Hao Wang, Dianbo Sui, Yunhang Shen, Ke Li, Xing Sun, and Enhong Chen. 2024. https://doi.org/10.1007/S11432-024-4251-X Woodpecker: hallucination correction for multimodal large language models . Sci. China Inf. Sci., 67(12)

    2024 doi

  32. [43]

    Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. 2024. https://doi.org/10.1109/CVPR52733.2024.00845 Make pixels dance: High-dynamic video generation . In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA...

    2024

  33. [44]

    Boqiang Zhang, Kehan Li, Zesen Cheng, Zhiqiang Hu, Yuqian Yuan, Guanzheng Chen, Sicong Leng, Yuming Jiang, Hang Zhang, Xin Li, Peng Jin, Wenqi Zhang, Fan Wang, Lidong Bing, and Deli Zhao. 2025. https://doi.org/10.48550/ARXIV.2501.13106 Videollama 3: Frontier multimodal foundat...

    2025 doi arXiv

  34. [45]

    Hang Zhang, Xin Li, and Lidong Bing. 2023. https://doi.org/10.18653/V1/2023.EMNLP-DEMO.49 Video-llama: An instruction-tuned audio-visual language model for video understanding . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, EMNLP 20...

    2023 doi

  35. [46]

    Jiacheng Zhang, Yang Jiao, Shaoxiang Chen, Jingjing Chen, and Yu - Gang Jiang. 2024. https://doi.org/10.48550/ARXIV.2409.16597 Eventhallusion: Diagnosing event hallucinations in video llms . CoRR, abs/2409.16597

    2024 doi

  36. [47]

    Weinberger, and Yoav Artzi

    Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. 2020. https://openreview.net/forum?id=SkeHuCVFDr Bertscore: Evaluating text generation with bert . In International Conference on Learning Representations

    2020

  37. [48]

    Dian Zheng, Ziqi Huang, Hongbo Liu, Kai Zou, Yinan He, Fan Zhang, Yuanhan Zhang, Jingwen He, Wei - Shi Zheng, Yu Qiao, and Ziwei Liu. 2025. https://doi.org/10.48550/ARXIV.2503.21755 Vbench-2.0: Advancing video generation benchmark suite for intrinsic faithfulness . CoRR, abs/2...

    2025 doi arXiv

  38. [49]

    Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. 2022. https://doi.org/10.48550/ARXIV.2211.11018 Magicvideo: Efficient video generation with latent diffusion models . CoRR, abs/2211.11018

    2022 doi arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.