Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

ChainReaction: Causal Chain-Guided Reasoning for Modular and Explainable Causal-Why Video Question Answering

T0 review · 5 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that causal-why video question answering should be split into causal-chain extraction and chain-based answering, and that this modular decoupling beats black-box models while making reasoning inspectable.

desk verdict A promising modular causal-chain VideoQA idea whose main claims are undercut by answer-derived annotations and an unfair SOTA comparison; worth engaging, but needs substantial revision. read the letter →

arxiv 2508.21010 v3 pith:MHDBUY65 submitted 2025-08-28 cs.CV cs.AIcs.CLcs.HCcs.LG

Paritosh Parmar , Eric Peh , Basura Fernando This is my paper
classification cs.CVcs.AIcs.CLcs.HCcs.LG
keywords CausalVideoQuestionAnsweringchainsStructuralmodelsChain-of-thoughtExplainabilityQACauCoscoreLargelanguage
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 argues that causal-why video question answering fails when video understanding, causal reasoning, and answer generation are fused into one black-box model. The proposed fix is to split the task: a Causal Chain Extractor turns a video and a why-question into a short natural-language cause-effect chain, and a separate Causal Chain-Driven Answerer picks the answer from that chain. To train the extractor without annotated reasoning traces, the authors use a strong large language model to recover causal chains from existing human-written question-answer pairs, then verify them with programmatic checks, a second large language model, and human review against the videos, producing 46,024 verified chains. With ground-truth chains the answerer reaches 99.4% average accuracy; in the full pipeline the system outperforms all listed baselines on three benchmarks, and human studies report large gains in explainability, trust, and debuggability. The central bet is that explicit, human-readable causal chains are not just a transparency add-on but the mechanism that carries the reasoning.

What carries the argument

The central machinery is the natural-language causal chain: an ordered sequence of events with arrows, such as "Jerry grabs Tom's tail -> Tom grabs his tail -> Tom pulls himself through the tunnel", treated as a structured mediating variable in a structural causal model. The key move is factoring the monolithic dependency video -> answer <- question into two factorized stages: video -> chain <- question (the Causal Chain Extractor) and question -> answer <- chain (the Causal Chain-Driven Answerer). The data-generation mechanism is oracle large-language-model recovery of chains from question-answer pairs plus a three-stage verification pipeline, and the evaluation mechanism is CauCo, a fine-t

What would settle it

Take a random sample of the 46,024 chains, show independent annotators only the source videos without the gold answers, have them write the causal chain they infer, and measure semantic agreement with the large-language-model-recovered chains; low agreement would show the chains are artifacts of the answer rather than video-grounded causes.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that causal reasoning in video QA should be an explicit, editable artifact rather than a latent process. The authors define a causal chain as a short, ordered list of events joined by arrows, and insert it as a mediating variable between video/question and answer in a structural causal model. They factor this into two supervised modules: a Causal Chain Extractor (video + question -> chain) and a Causal Chain-Driven Answerer (question + chain + options -> answer). The enabling move is that human annotators of existing QA datasets already reasoned through such chains when writing answers, so a powerful large language model can recover the chains from

Load-bearing premise

The load-bearing premise is that causal chains recovered by a large language model from question and gold-answer pairs genuinely match the reasoning a human would use and can be grounded in the video, rather than being plausible answer-shaped text.

Editorial extensions

If this is right

  • If causal chains are the right intermediate representation, answer selection with perfect chains is near-ceiling (99.4% average accuracy), so the practical bottleneck shifts to chain extraction.
  • Decoupling reasoning from answer generation beats monolithic strong vision-language models on all three benchmarks, suggesting that explicit reasoning reduces shortcut reliance.
  • The Causal Chain Extractor transfers from a cartoon dataset to a real-world dataset with substantially better chain quality than a zero-shot baseline, making it a candidate reusable reasoning component.
  • Users prefer chain-based outputs over black-box outputs in over 85% of cases, trust them more, and can attribute failures to perception versus reasoning, aiding system debugging.
  • CauCo provides a causality-focused score for evaluating generated chains, complementing standard captioning metrics that do not assess causal coherence.

Reading between the lines

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

  • Editorial inference: the oracle large-language-model recovery pipeline implies causal-chain supervision can be produced cheaply for any QA dataset with high-quality human answers; a useful test is whether chains recovered from other domains pass the same verification rates.
  • Editorial inference: a cleaner test of whether the extractor learned video causality rather than dataset language would be to keep the answer options fixed and swap the video; if the extracted chain barely changes, the model is exploiting textual bias.
  • Editorial inference: if the near-perfect ground-truth-chain result extends to open-ended answer generation, the answerer could be reused as a domain-agnostic causal reasoning module, making the paper's 'reusable causal reasoning engine' claim literal.
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 / 5 minor

Summary. The paper proposes ChainReaction, a two-stage modular framework for Causal-Why VideoQA. A Causal Chain Extractor (CCE, based on VILA-3B) generates natural-language causal chains from a video-question pair, and a Causal Chain-Driven Answerer (CCDA, based on LLaMA-3.1-8B) selects the answer from candidate options using the chain, question, and options. Since no causal-chain annotations exist, the authors generate chains from gold (Q,A) pairs using GPT-4o, with programmatic checks, cross-LLM verification via Gemini 2.5, and a small human verification stage, collecting 46,024 chains across NextQA, CausalVidQA, and CausalChaos!. They also introduce a CauCo coherence metric for chain quality. Experiments report near-perfect accuracy when ground-truth (generated) chains are used (Table 1), SOTA-beating accuracy in end-to-end settings (Table 2), human-study gains in explainability/trust/debuggability (Tables 4-5), and OOD chain-generation results (Table 6).

Significance. The modular decoupling of causal reasoning from answer generation, with explicit natural-language causal chains, is a timely and potentially impactful idea. The construction of a large chain-augmented training set and the introduction of CauCo are useful contributions, and the paper does include human studies and OOD experiments. However, the current evidence for the core claims is weakened by (i) the fact that chains are generated from gold answers rather than video, making Table 1 a near-circular 'upper bound'; (ii) an apples-to-oranges comparison in Table 2, where baselines appear to be used zero-shot while the proposed model is fine-tuned on target datasets; and (iii) human studies that cannot distinguish video-grounded explanations from answer-shaped paraphrases. These issues are load-bearing for the paper's headline claims of SOTA performance and explainability.

major comments (5)
  1. [§3, Corollary I and Table 1] Causal chains are generated by prompting GPT-4o with only (Q, A); the video V is never used during generation. Human verification checks consistency of the chain with (Q,A) and V, but it is performed on only ~1,000 samples per dataset (~3,000 of 46,024). Therefore Table 1's 99.40% accuracy is not an upper bound on reasoning from video-grounded causal chains; it largely demonstrates that CCDA can map answer-derived chains back to the answer. The same chains serve as SFT targets for CCE, so CCE is trained to emit answer-shaped text, which can create an answer-leakage shortcut independent of genuine video-grounded causality.
  2. [§5, Table 2] The SOTA comparison is not apples-to-apples. The proposed model is supervised fine-tuned on each target dataset (both CCE and CCDA), while the comparison models—especially VLMs such as VILA 1.5, GPT-4o, QwenVL2.5, DeepSeek-VL2—are evaluated in what appears to be a zero-shot setting, as the paper does not state otherwise. The reported gains (e.g., 63.95 vs. 70.75 on NextQA for QwenVL2.5, where Ours is actually lower) and the average superiority could be explained by fine-tuning alone. The authors should either fine-tune the baselines on the same datasets or clearly report zero-shot vs. fine-tuned status for every entry, and include a fine-tuned VILA/LLaMA baseline to control for architecture and fine-tuning effects.
  3. [§5, Studies I–III (Tables 4–5)] The human studies are intended to demonstrate explainability and trust, but they are evaluating causal chains produced by CCE, which was trained on answer-derived chains. Since the chains may be plausible paraphrases of the gold answer rather than video-grounded causal reasoning, the studies cannot distinguish a faithful explanation from a fluent paraphrase. To support the explainability claim, the authors should evaluate chains on video-grounded criteria (e.g., human rating of whether each event is observable in the video, or whether the chain would be the same if the answer were withheld).
  4. [§5, Table 6 and §3] The OOD chain-generation experiment evaluates the CCE against 'groundtruth causal chains' that were themselves generated from (Q,A) with a similar LLM pipeline. Thus the automatic metrics (BLEU, ROUGE, SPICE, CCS) measure similarity to answer-derived paraphrases, not video-grounded correctness. A more meaningful evaluation would compare chains against independently constructed, video-grounded human annotations, at least on a held-out subset. This caveat also applies to Table 3.
  5. [§3, Preliminary II] Preliminary II states that 'LLMs have been demonstrated to accurately recover the intermediate reasoning steps or causal chains' from questions and gold answers, but no reference or supporting experiment is provided. Since Corollary I rests directly on this premise, this is a load-bearing support gap. The authors should cite existing evidence or provide a small controlled study demonstrating that the recovered chains are not merely paraphrases of the answer text.
minor comments (5)
  1. [§5, Results] The text says 'See Figure 3 for qualitative results' in the SOTA comparison, but Figure 3 is the training diagram; the qualitative examples appear in Figure 5. The figure reference should be corrected.
  2. [§4 and Tables 3, 6] The abbreviation CCS is used in tables but is not defined. The text introduces 'CauCo score' as the causal coherence score; please use consistent terminology and define CCS explicitly.
  3. [§1, Abstract/Conclusion] There are several typos, e.g., 'explainabiity' in the conclusion and inconsistent use of 'NextQA' vs 'NXQA' in tables. A careful proofreading pass is needed.
  4. [§2, Related Work] The reference list is long but some citations in the related-work discussion appear to be incomplete or informal (e.g., 'Zang et al. (Zang et al. 2023)' is used repeatedly; please ensure all citations follow the same format). The paper also does not cite any work for the claim in Preliminary II.
  5. [§4, SCM factorization] The Structural Causal Model notation V→C←Q and Q→A←C is introduced informally. A precise definition of the structural equations or the assumed independence relations would make the theoretical motivation clearer.

Circularity Check

1 steps flagged · score 6.0 of 10

Answer-derived causal chains make the Table 1 upper bound circular; the end-to-end SOTA claim remains partially independent.

  1. self definitional [Section 3 (Causal Chain Generation Process) and Section 5 (Experimental Upper bound on Performance, Table 1)]
    "We prompt the Oracle LLM with question ( Q) and the corresponding human written correct, gold answer ( A) and ask it to generate a causal chain ( C) in natural language in a specified format. ... we use groundtruth causal chains instead of predicted ones, training & testing the Causal Chain-Driven Answerer on these annotations. This setting simulates an ideal scenario where causal chain generation is perfectly accurate."

    The causal chain C is generated from the gold answer A, so C is defined in terms of A. Table 1 then 'predicts' A from C, yielding near-perfect accuracy (99.40 avg) by construction: the chain is an LLM paraphrase/expansion of A, not an independently video-grounded reasoning trace. The paper's own example illustrates this: A is 'Jerry wanted to save Uncle Pecos from Tom by shooting toothpaste at Tom' while C ends with 'Uncle Pecos is saved from Tom.' Thus the upper-bound experiment does not establish that causal chains are effective intermediate representations; it demonstrates A→C→A round-tripping. This is the self-definitional pattern: X (chain) is derived from Y (answer), and then Y is 'predicted' from X.

full rationale

The paper's central end-to-end claim (Table 2) is not fully circular: at inference the CCE must produce a chain from video and question without access to the gold answer, and the reported improvements over baselines are not forced by construction. However, the causal-chain annotations themselves are produced by prompting GPT-4o with (Q,A), so they are answer-shaped paraphrases rather than independent, video-grounded reasoning traces. Consequently, Table 1's 'upper bound'—near-perfect accuracy when the CCDA is fed these 'ground truth causal chains'—is a round-trip: A→C via LLM, C→A via CCDA. The paper's own example confirms the chain entails the answer. Because the same A-derived chains are the SFT targets for the CCE, the explainability and trust studies (Studies I–III) cannot separate faithful video-grounded explanations from plausible answer paraphrase. Preliminary II is asserted without a supporting reference, and Corollary I is an unsupported load-bearing premise; these are correctness risks rather than circularity steps per se. Since one key supporting experiment reduces by construction while the main SOTA comparison retains independent content, a partial-circularity score of 6 is appropriate.

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

The framework rests on the assumption that LLM-generated chains from question-answer pairs are faithful reasoning traces. The SCM language is metaphorical rather than formal. No new physical or mathematical entities are introduced.

assumptions (5)
  • domain assumption Human annotators of base datasets implicitly used causal chains when writing answers.
    Section 3, Preliminary I. This underpins the validity of recovering chains from Q,A pairs. If annotators did not use such chains, the LLM-recovered chains may be artificial.
  • domain assumption LLMs can reliably recover causal chains from (Q,A).
    Section 3, Preliminary II and Corollary I. GPT-4o generates chains, Gemini 2.5 verifies; no evidence beyond a sample of 1000 per dataset that these match human reasoning.
  • ad hoc to paper Causal chains are an effective intermediate representation for VideoQA.
    This is the central hypothesis, supported mainly by the upper-bound experiment which is near-trivially high because chains are derived from A.
  • ad hoc to paper The SCM factorization V->C<-Q and Q->A<-C captures causal structure.
    Section 4. It is a descriptive graph, not a testable causal model; no interventions, counterfactuals, or do-calculus are used.
  • domain assumption CauCo, trained on positive/negative chains from the same distribution, is a valid measure of causal coherence.
    Section 4, CauCo. Negative samples are perturbed positives from the same LLM style, so the metric may reward chains similar to the LLM's output rather than true causality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ChainReaction: Causal Chain-Guided Reasoning for Modular and Explainable Causal-Why Video Question Answering." pith.science (2026). https://pith.science/paper/MHDBUY65

@misc{pith2026250821010,
  author       = {Pith},
  title        = {Pith review of: ChainReaction: Causal Chain-Guided Reasoning for Modular and Explainable Causal-Why Video Question Answering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MHDBUY65}},
  note         = {Machine review of arXiv:2508.21010}
}
read the original abstract

Existing Causal-Why Video Question Answering (VideoQA) models often struggle with higher-order reasoning, relying on opaque, monolithic pipelines that entangle video understanding, causal inference, and answer generation. These black-box approaches offer limited interpretability and tend to depend on shallow heuristics. We propose a novel, modular paradigm that explicitly decouples causal reasoning from answer generation, introducing natural language causal chains as interpretable intermediate representations. Inspired by human cognitive models, these structured cause-effect sequences bridge low-level video content with high-level causal reasoning, enabling transparent and logically coherent inference. Our two-stage architecture comprises a Causal Chain Extractor (CCE) that generates causal chains from video-question pairs, and a Causal Chain-Driven Answerer (CCDA) that derives answers grounded in these chains. To address the lack of annotated reasoning traces, we introduce a scalable method for generating accurate causal chains from existing datasets. We construct human verified causal chains for 46K samples. We also propose CauCo, a new evaluation metric for causality-oriented captioning. Experiments on three large-scale benchmarks demonstrate that our approach not only outperforms state-of-the-art models, but also yields substantial gains in explainability, user trust, and generalization -- positioning the CCE as a reusable causal reasoning engine across diverse domains. Project page: https://paritoshparmar.github.io/chainreaction/

Figures

Figures reproduced from arXiv: 2508.21010 by the authors.

Figure 1
Figure 1. Concept. (1) Existing Video (V) Question (Q) An￾swer (A) approaches through the lens of structural causal models (SCMs), highlighting their monolithic and black￾box nature. (2) In contrast, we propose a principled de￾parture from this paradigm: a mediation mechanism based on natural language causal chains (C). We factorize this SCM into two SCMs (3,4)—enabling structured video un￾derstanding, reasoning, and inferenc… view at source ↗
Figure 2
Figure 2. Causal chain construction for SFT. (1) Human annotators of base datasets intuitively and implicitly make use of causal chains when writing correct answers. (2) We propose to recover these causal chains with the help of LLM using questions and correct gold answers. (3,4) Our robust causal chain generation and verification pipeline. 3 Causal Chain Construction for SFT In this section, we propose a novel, scalable and … view at source ↗
Figure 3
Figure 3. Training of both modules in stages. on the questions, and express them in a detailed yet con￾cise form. Generating causal chains from videos conditioned on questions constitutes a complex reasoning task. Towards that end, we propose to leverage the representational capac￾ity of foundation models. However, generating causal chains is non-trivial task that extends beyond the native capabilities of existing vision-lang… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Sensitivity to causal chains. (1) plot of per￾formance drop vs links removed. (2) Qualitative example. OCC: original causal chains, MCC: masked chains, SA: se￾lected answer. (3) Qualitative examples of SA changes based on hypothetical chains (HCC). Please zoom-in. grou…
Figure 5
Figure 5. Figure 5: Qualitative results. GT AO: Groundtruth Answer Option; BM: Baseline Model; SA: Selected Answer; CC: Causal Chain. Only a few frames per video are shown. Green and red boxes indicate success and failure cases. In the first example, actor masks come from the CausalVidQA …
Figure 6
Figure 6. Figure 6: Qualitative examples of Out-of-Domain gener￾ated causal chains (OODGCC). Please zoom-in. agnosis. Six researchers with expertise in computer vision, VLMs, and LLMs each analyze 20 failure cases from both systems, identifying fault locations. For our model, they view th…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Causal-Audit: Explicit and Auditable Graph-based Reasoning via Target-Aware Causal Chain Construction

    cs.AI 2026-04 conditional novelty 6.0 of 10

    A four-stage LLM framework that constructs, audits, and aggregates explicit causal chains outperforms prompting baselines on three context-free causal QA benchmarks.

Reference graph

Works this paper leans on

48 extracted references · 33 canonical work pages · cited by 1 Pith paper

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  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]

    Anderson, P.; Fernando, B.; Johnson, M.; and Gould, S. 2016. Spice: Semantic propositional image caption evaluation. In Computer Vision--ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, 382--398. Springer

    work page 2016

  4. [4]

    L.; and Parikh, D

    Antol, S.; Agrawal, A.; Lu, J.; Mitchell, M.; Batra, D.; Zitnick, C. L.; and Parikh, D. 2015. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, 2425--2433

    work page 2015

  5. [5]

    Ayazoglu, M.; Yilmaz, B.; Sznaier, M.; and Camps, O. 2013. Finding causal interactions in video sequences. In Proceedings of the IEEE International Conference on Computer Vision, 3575--3582

    work page 2013

  6. [6]

    Bai, S.; Chen, K.; Liu, X.; Wang, J.; Ge, W.; Song, S.; Dang, K.; Wang, P.; Wang, S.; Tang, J.; Zhong, H.; Zhu, Y.; Yang, M.; Li, Z.; Wan, J.; Wang, P.; Ding, W.; Fu, Z.; Xu, Y.; Ye, J.; Zhang, X.; Xie, T.; Cheng, Z.; Zhang, H.; Yang, Z.; Xu, H.; and Lin, J. 2025. Qwen2.5-VL Technical Report. arXiv preprint arXiv:2502.13923

    arXiv 2025

  7. [7]

    Banerjee, S.; and Lavie, A. 2005. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, 65--72

    2005

  8. [8]

    Bi, J.; Liang, S.; Zhou, X.; Liu, P.; Guo, J.; Tang, Y.; Song, L.; Huang, C.; Sun, G.; He, J.; et al. 2025. Why reasoning matters? a survey of advancements in multimodal reasoning (v1). arXiv preprint arXiv:2504.03151

    arXiv 2025

Show all 48 references
  1. [9]

    Chen, G.; Li, Y.; Liu, X.; Li, Z.; Al Suradi, E.; Wei, D.; and Zhang, K. 2023. LLCP: Learning Latent Causal Processes for Reasoning-based Video Question Answer. In The Twelfth International Conference on Learning Representations

    2023

  2. [10]

    Chen, T.; Liu, H.; He, T.; Chen, Y.; Ma, X.; Zhong, C.; Zhang, Y.; Wang, Y.; Lin, H.; Lin, W.; et al. 2025. MECD: Unlocking multi-event causal discovery in video reasoning. Advances in Neural Information Processing Systems, 37: 92554--92580

    2025

  3. [11]

    Chen, Z.; Wang, W.; Cao, Y.; Liu, Y.; Gao, Z.; Cui, E.; Zhu, J.; Ye, S.; Tian, H.; Liu, Z.; et al. 2024. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271

    2024 arXiv

  4. [12]

    Fan, C.; Zhang, X.; Zhang, S.; Wang, W.; Zhang, C.; and Huang, H. 2019. Heterogeneous memory enhanced multimodal attention model for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 1999--2007

    2019

  5. [13]

    Foss, A.; Evans, C.; Mitts, S.; Sinha, K.; Rizvi, A.; and Kao, J. T. 2025. CausalVQA: A Physically Grounded Causal Reasoning Benchmark for Video Models. arXiv preprint arXiv:2506.09943

    2025 arXiv

  6. [14]

    Gao, D.; Zhou, L.; Ji, L.; Zhu, L.; Yang, Y.; and Shou, M. Z. 2023. MIST: Multi-modal Iterative Spatial-Temporal Transformer for Long-form Video Question Answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 14773--14783

    2023

  7. [15]

    Gao, J.; Ge, R.; Chen, K.; and Nevatia, R. 2018. Motion-appearance co-memory networks for video question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 6576--6585

    2018

  8. [16]

    M.; Schulz, L

    Gopnik, A.; Glymour, C.; Sobel, D. M.; Schulz, L. E.; Kushnir, T.; and Danks, D. 2004. A theory of causal learning in children: causal maps and Bayes nets. Psychological review, 111(1): 3

    2004

  9. [17]

    Grattafiori, A.; Dubey, A.; Jauhri, A.; Pandey, A.; Kadian, A.; Al-Dahle, A.; Letman, A.; Mathur, A.; Schelten, A.; Vaughan, A.; et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    2024 arXiv

  10. [18]

    Hanson, N. R. 1955. Causal chains. Mind, 64(255): 289--311

    1955

  11. [19]

    Jiang, P.; and Han, Y. 2020. Reasoning with heterogeneous graph alignment for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, 11109--11116

    2020

  12. [20]

    M.; Le, V.; Venkatesh, S.; and Tran, T

    Le, T. M.; Le, V.; Venkatesh, S.; and Tran, T. 2020. Hierarchical conditional relation networks for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 9972--9981

    2020

  13. [21]

    Li, J.; Li, D.; Savarese, S.; and Hoi, S. 2023. BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models. In ICML

    2023

  14. [22]

    Li, J.; Niu, L.; and Zhang, L. 2022. From Representation to Reasoning: Towards both Evidence and Commonsense Reasoning for Video Question-Answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2022

  15. [23]

    Li, K.; Wang, Y.; He, Y.; Li, Y.; Wang, Y.; Liu, Y.; Wang, Z.; Xu, J.; Chen, G.; Luo, P.; et al. 2024. Mvbench: A comprehensive multi-modal video understanding benchmark

    2024

  16. [24]

    Li, Y.; Torralba, A.; Anandkumar, A.; Fox, D.; and Garg, A. 2020. Causal discovery in physical systems from videos. Advances in Neural Information Processing Systems, 33: 9180--9192

    2020

  17. [25]

    Lin, C.-Y. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, 74--81

    2004

  18. [26]

    Lin, J.; Yin, H.; Ping, W.; Molchanov, P.; Shoeybi, M.; and Han, S. 2024. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 26689--26699

    2024

  19. [27]

    Ma, J. 2025. Causal Inference with Large Language Model: A Survey. In Findings of the Association for Computational Linguistics: NAACL 2025, 5886--5898

    2025

  20. [28]

    Narendra, T.; Sankaran, A.; Vijaykeerthy, D.; and Mani, S. 2018. Explaining deep learning models using causal inference. arXiv preprint arXiv:1811.04376

    2018 arXiv

  21. [29]

    Read the Room

    Niu, L.; Li, J.; Yu, X.; Wang, S.; Feng, R.; Wu, B.; Wei, P.; Wang, Y.; and Fan, L. 2025. R\^ 3-VQA:" Read the Room" by Video Social Reasoning. arXiv preprint arXiv:2505.04147

    2025 arXiv

  22. [30]

    OpenAI . 2024. Hello GPT-4o. https://openai.com/index/hello-gpt-4o/. [Online; accessed 31-May-2024]

    2024

  23. [31]

    Papineni, K.; Roukos, S.; Ward, T.; and Zhu, W.-J. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, 311--318

    2002

  24. [32]

    E.; Chen, Y.; Tan, E.; and Fernando, B

    Parmar, P.; Peh, E.; Chen, R.; Lam, T. E.; Chen, Y.; Tan, E.; and Fernando, B. 2024. CausalChaos! Dataset for Comprehensive Causal Action Question Answering Over Longer Causal Chains Grounded in Dynamic Visual Scenes. In The Thirty-eight Conference on Neural Information Proces...

    2024

  25. [33]

    Paszke, A.; Gross, S.; Massa, F.; Lerer, A.; Bradbury, J.; Chanan, G.; Killeen, T.; Lin, Z.; Gimelshein, N.; Antiga, L.; et al. 2019. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32

    2019

  26. [34]

    Pearl, J. 2009. Causality. Cambridge university press

    2009

  27. [35]

    Pearl, J.; and Mackenzie, D. 2018. The book of why: the new science of cause and effect. Basic books

    2018

  28. [36]

    S.; Matyasko, A.; Jaiswal, S.; Fernando, B.; and Tan, C

    Rawal, I. S.; Matyasko, A.; Jaiswal, S.; Fernando, B.; and Tan, C. 2024. Dissecting Multimodality in VideoQA Transformer Models by Impairing Modality Fusion. In International Conference on Machine Learning, 42213--42244. PMLR

    2024

  29. [37]

    R.; Kalchbrenner, N.; Goyal, A.; and Bengio, Y

    Sch \"o lkopf, B.; Locatello, F.; Bauer, S.; Ke, N. R.; Kalchbrenner, N.; Goyal, A.; and Bengio, Y. 2021. Toward causal representation learning. Proceedings of the IEEE, 109(5): 612--634

    2021

  30. [38]

    H.; and Chang, S.-F

    Su, H.-T.; Niu, Y.; Lin, X.; Hsu, W. H.; and Chang, S.-F. 2023. Language models are causal knowledge extractors for zero-shot video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 4951--4960

    2023

  31. [39]

    Tang, K.; Niu, Y.; Huang, J.; Shi, J.; and Zhang, H. 2020. Unbiased scene graph generation from biased training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 3716--3725

    2020

  32. [40]

    Trabasso, T.; and Van Den Broek, P. 1985. Causal thinking and the representation of narrative events. Journal of memory and language, 24(5): 612--630

    1985

  33. [41]

    Wang, Y.; Meng, L.; Ma, H.; Wang, Y.; Huang, H.; and Meng, X. 2024. Modeling Event-level Causal Representation for Video Classification. In Proceedings of the 32nd ACM International Conference on Multimedia, 3936--3944

    2024

  34. [42]

    V.; Zhou, D.; et al

    Wei, J.; Wang, X.; Schuurmans, D.; Bosma, M.; Xia, F.; Chi, E.; Le, Q. V.; Zhou, D.; et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35: 24824--24837

    2022

  35. [43]

    Wei, Y.; Liu, Y.; Yan, H.; Li, G.; and Lin, L. 2023. Visual causal scene refinement for video question answering. In Proceedings of the 31st ACM International Conference on Multimedia, 377--386

    2023

  36. [44]

    Wu, Z.; Chen, X.; Pan, Z.; Liu, X.; Liu, W.; Dai, D.; Gao, H.; Ma, Y.; Wu, C.; Wang, B.; Xie, Z.; Wu, Y.; Hu, K.; Wang, J.; Sun, Y.; Li, Y.; Piao, Y.; Guan, K.; Liu, A.; Xie, X.; You, Y.; Dong, K.; Yu, X.; Zhang, H.; Zhao, L.; Wang, Y.; and Ruan, C. 2024. DeepSeek-VL2: Mixture...

    2024 arXiv

  37. [45]

    Xiao, J.; Shang, X.; Yao, A.; and Chua, T.-S. 2021. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 9777--9786

    2021

  38. [46]

    Yang, X.; Feng, F.; Ji, W.; Wang, M.; and Chua, T.-S. 2021. Deconfounded video moment retrieval with causal intervention. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval, 1--10

    2021

  39. [47]

    Zang, C.; Wang, H.; Pei, M.; and Liang, W. 2023. Discovering the real association: Multimodal causal reasoning in video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 19027--19036

    2023

  40. [48]

    Zhang, H.; Li, X.; and Bing, L. 2023. Video- LL a MA : An Instruction-tuned Audio-Visual Language Model for Video Understanding. In Feng, Y.; and Lefever, E., eds., Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, 5...

    2023

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.