Comparing Learning Paradigms for Egocentric Video Summarization

Daniel Wen

REVIEW 3 major objections 5 minor 19 references

Comparing Learning Paradigms for Egocentric Video Summarization

T0 review · 3 major / 5 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 A prompt-tuned GPT-4o beats specialized video summarizers on egocentric clips.

desk verdict A novel empirical comparison that overclaims on thin evidence: one annotator's scores on 21 videos cannot support the GPT-4o margin or the first- versus third-person claim. read the letter →

arxiv 2506.21785 v1 pith:MMTOYSPE submitted 2025-06-26 cs.CV cs.AI

Daniel Wen This is my paper

classification cs.CVcs.AI

keywords egocentricvideosummarizationpromptfine-tuningGPT-4oShotluckHolmesTAC-SUMEgo-Exo4Dlearningparadigms

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 asks which learning paradigm—supervised, unsupervised, or prompt fine-tuning—best summarizes first-person (egocentric) video. The author evaluates three systems on 21 egocentric videos drawn from the Ego-Exo4D dataset: Shotluck Holmes (supervised state of the art), TAC-SUM (unsupervised state of the art, in CLIP and DINO variants), and a GPT-4o model steered by chain-of-thought prompting and history-aware frame-by-frame narration. Headline result: the prompt-tuned GPT-4o scores 64.95 on a human-scored quality measure, ahead of Shotluck Holmes at 61.19 and the TAC-SUM variants at 56.24 and 58.43. The paper reads this as evidence that current specialized models, designed for third-person footage, handle the camera shake, varying perspectives, and task-focused framing of body-worn video poorly, and it treats the comparison as a proof of concept rather than a definitive benchmark.

What carries the argument

The central machinery is the prompt fine-tuning protocol rather than a new model architecture: GPT-4o receives a four-step chain-of-thought prompt that directs it to segment the video, identify key activities, remove redundant segments, and assemble a chronological summary; in the alternative frame-wise variant, a short prompt asks for narration of each frame given the timestamp and previous narrations. The comparison rests on a quality score defined as the mean of a single human evaluator's accuracy, clarity, and relevance ratings for each summary, averaged over all videos. That scoring rubric is the measurement instrument that orders the models.

What would settle it

Blind re-scoring of the same 84 summaries by a larger panel of independent evaluators would test whether GPT-4o's lead over Shotluck Holmes survives; if average scores reorder or the gap narrows below noise, the paper's central claim fails. A more direct test would run the three models on a larger, fixed set of Ego-Exo4D videos with ground-truth summaries and compare against standard automatic metrics.

Watch

Extended reading notes

Core claim

The paper's central discovery is that a general-purpose multimodal model, GPT-4o, adapted to the task through prompt fine-tuning, outperforms both a supervised state-of-the-art summarizer (Shotluck Holmes) and an unsupervised one (TAC-SUM) on egocentric video summarization, with mean quality scores of 64.95 versus 61.19, 56.24, and 58.43. The paper additionally asserts that state-of-the-art models are less effective on first-person video than on third-person video, attributing the gap to dynamic camera motion and shifting object perspectives; however, no third-person videos were tested in the reported evaluation. The winning system uses a chain-of-thought prompt instructing the model to segment the video by motion and scene changes, identify key activities within each segment, discard repetitive or irrelevant intervals, and output a chronological summary timeline; the frame-wise variant grounds each narration in the previous narrations and a timestamp.

Load-bearing premise

The conclusions rest on the assumption that one person's subjective ratings of 21 selected first-person videos are a dependable and representative measure of summarization quality; if the ratings are noisy or the videos atypical, both headline claims collapse.

Editorial extensions

If this is right

If GPT-4o's advantage holds under broader evaluation, general-purpose multimodal models should be considered a strong baseline for egocentric summarization, not just specialized video architectures.
The small observed gaps imply that none of the three paradigms solves egocentric summarization, so the problem remains open for models that explicitly handle dynamic egocentric camera motion.
Because the strongest system used no training on egocentric data, prompt-time structure (segmentation, key-activity detection, redundancy removal) may substitute for some supervision in video tasks.
Future benchmarks for first-person video should separate model capability from prompt design, since the winning score depends on the prompt protocol as much as on the model.

Reading between the lines

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

If egocentric videos were matched against third-person videos from the same Ego-Exo4D activities, the claimed first-person shortfall could be tested directly; the paper currently offers no third-person baseline.
An inter-annotator agreement study (e.g., several raters scoring the same summaries) would reveal whether the 3.76-point lead is meaningful relative to rating noise.
The chain-of-thought prompt's structure resembles the temporal-context mechanism in TAC-SUM, suggesting that prompting may reproduce inductive biases that unsupervised methods learn from clustering.
A testable extension: applying the same prompt to a frame stream without timestamps would isolate whether temporal grounding or the model's world knowledge drives the win.

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

The new thing here is the direct comparison itself: Shotluck Holmes, TAC-SUM, and prompt-engineered GPT-4o on egocentric video summarization, with GPT-4o scoring 64.95 against Shotluck's 61.19 and the TAC-SUM variants in the high 50s. That specific ranking is new, as far as I know, and it is worth a second look. The paper is also transparent about its resource limits, calls itself a proof-of-concept, and gives enough detail on the TAC-SUM and Shotluck reimplementations to be useful to someone trying to run them on Ego4D-style data. The Future Work section admits the manual quality ranking is a limitation, which is honest.

The soft spots are exactly where the stress-test note lands. The headline claims rest on one human evaluator scoring 21 videos per model, with no inter-rater reliability, no confidence intervals, and no significance tests. The 3.76-point gap between GPT-4o and Shotluck Holmes is within plausible noise: with 21 videos, even a moderate across-video standard deviation of 20 yields a standard error above 4 points, mean the margin could easily flip. And the abstract's claim about first-person versus third-person performance is not tested at all—there are no third-person videos in the evaluation. The conclusion that models designed for third-person videos struggle with egocentric ones is plausible, but this paper does not demonstrate it.

Minor issues: 'prompt fine-tuning' is really just prompting with a few templates, not any gradient update, and the 21-video sample is small even for a proof-of-concept. These are not fatal on their own, but they compound the evidence problem.

Who gets value from this? Someone working on egocentric summarization might find the implementation details and the described failure modes interesting, and it could serve as a cautionary example of how easy it is to overstate results from a small human evaluation. But as a standalone empirical claim, it does not hold up. If it goes to peer review, a serious referee could push for a larger annotated set, multiple annotators, a third-person control, and significance testing—that would be a useful outcome. I would accept it for review with that expectation, but I would not cite it in its current form.

Recommendation: send it to review with a clear major-revision pathway, or desk-reject and invite resubmission with the missing baselines. Either way, the present version does not establish its headline claims.

Referee Report

3 major / 5 minor

Summary. The paper compares three learning paradigms for egocentric video summarization: Shotluck Holmes (supervised), TAC-SUM (unsupervised, with CLIP and DINO variants), and a prompt fine-tuned GPT-4o model. Evaluation is performed on 21 first-person videos from Ego-Exo4D, with summaries scored by a single human evaluator on accuracy, clarity, and relevance. The paper reports that GPT-4o achieves the highest quality score (64.95) and concludes that current state-of-the-art models are less effective on first-person videos than on third-person videos, while also identifying challenges such as dynamic camera movement. The manuscript is framed as a proof-of-concept under resource constraints.

Significance. If the central comparison were reliable, the result that a general-purpose prompted model outperforms specialized video summarization models on egocentric data would be a useful proof-of-concept and would support the need for more egocentric-specific model development. The qualitative examples (rock climbing, cooking, dancing) are illustrative and potentially informative. However, the two headline claims are not supported by the evidence as presented: the first-person-versus-third-person comparison is untested, and the model ranking rests on a single evaluator's subjective ratings without any reliability analysis. The paper's transparency about its limitations is a positive feature, but it does not compensate for the missing baseline and missing statistical support.

major comments (3)

[Abstract and Section 5 (Conclusion)] The central claim that 'current state-of-the-art models perform less effectively on first-person videos compared to third-person videos' is not tested anywhere in the manuscript. Section 4.1 describes evaluation on 21 first-person videos only; no third-person videos are summarized or scored. The Nvidia example in Figure 1 is an anecdotal third-person illustration, not a comparison. This claim is load-bearing in the abstract and conclusion and must be either supported by a systematic third-person baseline or removed.
[Section 4.1 (Evaluation Methodology) and Table 1] The Quality Score is defined as the mean of one human evaluator's 0-100 ratings across accuracy, clarity, and relevance for 21 videos per model, but the paper reports no inter-rater reliability, confidence intervals, significance tests, or per-video scores. The headline margin between GPT-4o (64.95) and Shotluck Holmes (61.19) is only 3.76 points, and Section 4.4 itself describes the result as 'marginally surpasses.' Without any estimate of evaluator noise, the ranking is not established. The paper's own Future Work admits that the 'manual quality ranking system' is a limitation, yet the abstract states the ranking as a definitive finding.
[Section 4.1 (Evaluation Methodology), steps i-vi] Although Section 4.1 states that the videos 'are annotated with ground-truth video summaries and narrations,' the evaluation procedure never uses these ground-truth annotations. The human evaluator only compares each model summary to the original video, so the Quality Score measures subjective impression rather than agreement with a reference summary. This is a missed opportunity to ground the comparison in an objective benchmark and further weakens the reliability of the central comparison.

minor comments (5)

[Throughout] There are several typos and formatting issues, including 'it's' for 'its' in Sections 2.3 and 3.2, 'Shotluck Holmes' with an unwanted space in Section 4.3, and inconsistent spacing in 'TinyLLaV A.'
[Results and GitHub/Google Drive mention] The paper says the video summaries can be found in the 'provided Google Drive in our Github,' but no URL is given, so the availability of the summaries for verification cannot be confirmed.
[References] Several listed references are never cited in the body: [3] (VideoMamba), [4] (Mamba), [5] (Space-Time Attention), and [12] (Chain of Density). Either cite them or remove them from the bibliography.
[Section 4.2] The claim that TAC-SUM-DINO outperforms TAC-SUM-CLIP (58.43 vs. 56.24) is based on the same single-evaluator protocol and would also require reliability evidence; the difference is even smaller than the GPT-4o-vs-Shotluck-Holmes margin.
[Section 4.1] The statement that the 21-video subset 'comprehensively represents all activities within the Ego-Exo4D dataset' is a strong generalization claim that is not quantitatively supported, especially given the diversity of activities in the full dataset.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the comparison is empirical, with human ratings as an external benchmark and no fitted parameters or self-citation chains driving the conclusions.

full rationale

The paper contains no mathematical derivation or parameter-fitting loop that could make a prediction reduce to its inputs by construction. The central comparison, GPT-4o versus Shotluck Holmes and TAC-SUM variants, is evaluated by a human scorer against original Ego-Exo4D videos, an external benchmark not derived from any of the models' outputs. The paper does not fit any parameter to the evaluation subset and then rename that fit as a prediction; the quality score is a direct mean of human ratings. The references are to prior external works (Shotluck Holmes, TAC-SUM, Ego-Exo4D, GPT-4o) and not to the author's own prior results, so there is no self-citation chain that forces the conclusion. The abstract's statement that models perform less effectively on first-person than third-person videos is not tested in the paper, but that is an unsupported or under-supported empirical claim, not a circularity. Similarly, the small evaluation set and single evaluator raise validity concerns, but those are correctness or reliability issues, not instances of the paper's conclusions being equivalent to its inputs by definition.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The central comparison relies on no fitted numeric parameters. The main epistemic load is carried by assumptions about the validity of the human evaluation and the representativeness of the video subset; these are listed as domain assumptions. No invented entities are introduced.

assumptions (3)

domain assumption Human evaluator ratings are a valid and reliable proxy for video summarization quality.
The entire comparison rests on a single human evaluator's subjective scores on accuracy, clarity, and relevance, with no validation against established summarization metrics or ground-truth summaries. Entered in Section 4.1.
domain assumption The 21-video Ego-Exo4D subset is representative of egocentric videos.
The paper claims the subset 'comprehensively represents all activities within the Ego-Exo4D dataset' but provides no selection protocol, coverage statistics, or evidence of representativeness. Entered in Section 4.1.
domain assumption The models were configured equivalently to their original publications.
Shotluck Holmes is fine-tuned on Ego4D rather than used as released, and TAC-SUM's hyperparameters (e.g., epsilon, cluster count) are taken from the original paper. Performance differences could reflect configuration choices rather than paradigm differences. Entered in Sections 3.2 and 3.3.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Comparing Learning Paradigms for Egocentric Video Summarization." pith.science (2026). https://pith.science/paper/MMTOYSPE

@misc{pith2026250621785,
  author       = {Pith},
  title        = {Pith review of: Comparing Learning Paradigms for Egocentric Video Summarization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MMTOYSPE}},
  note         = {Machine review of arXiv:2506.21785}
}

read the original abstract

In this study, we investigate various computer vision paradigms - supervised learning, unsupervised learning, and prompt fine-tuning - by assessing their ability to understand and interpret egocentric video data. Specifically, we examine Shotluck Holmes (state-of-the-art supervised learning), TAC-SUM (state-of-the-art unsupervised learning), and GPT-4o (a prompt fine-tuned pre-trained model), evaluating their effectiveness in video summarization. Our results demonstrate that current state-of-the-art models perform less effectively on first-person videos compared to third-person videos, highlighting the need for further advancements in the egocentric video domain. Notably, a prompt fine-tuned general-purpose GPT-4o model outperforms these specialized models, emphasizing the limitations of existing approaches in adapting to the unique challenges of first-person perspectives. Although our evaluation is conducted on a small subset of egocentric videos from the Ego-Exo4D dataset due to resource constraints, the primary objective of this research is to provide a comprehensive proof-of-concept analysis aimed at advancing the application of computer vision techniques to first-person videos. By exploring novel methodologies and evaluating their potential, we aim to contribute to the ongoing development of models capable of effectively processing and interpreting egocentric perspectives.

Figures

Figures reproduced from arXiv: 2506.21785 by the authors.

**Figure 1.** Nvidia Video Summarization Example This frame is taken from an example video produced by Nvidia’s latest video summarization agent [19]. The model demonstrates strong performance in video summarization and narration on thirdperson videos. However, it is crucial to acknowledge that these results were obtained in a controlled, simulated environment with limited variables and no camera movement, conditions that contri… view at source ↗

**Figure 2.** TAC-SUM model architecture that the reduced sequence effectively captures the essential content of the original video while significantly reducing the number of frames, thereby balancing computational efficiency and data fidelity. For each frame in the sampled sequence ˆIi , the CLIP and DINO feature extractors are used to generate its corresponding visual embedding ei . Similarly to the TAC-SUM paper, we assess th… view at source ↗

**Figure 3.** Shotluck Holmes model architecture To preprocess the Ego4D videos for compatibility with the Shotluck Holmes architecture, each video is converted into a tensor, which is then input into SigLip [16], a vision encoder. To manage the frame limitations of our training hardware, we employ one of two sampling methods: uniform sampling and head-tail sampling. Uniform sampling selects frames at regular intervals across the… view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: TAC-SUM-DINO vs TAC-SUM-CLIP: Ego-Exo4D minnesota_rockclimbing_034_16 [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]

**Figure 5.** Figure 5: Shotluck Holmes: Ego-Exo4D upenn_0711_Cooking_6_4 [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

**Figure 6.** Figure 6: GPT-4o: Ego-Exo4D upenn_0727_Partner_Dance_2_2_5 [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

19 extracted references · 5 canonical work pages

[1]

UniVTG: Towards Unified Video-Language Temporal Grounding

Lin, Kevin Q., et al. "UniVTG: Towards Unified Video-Lanaguage Temporal Grounding." ArXiv (Cornell University), 18 Aug 2023, https://arxiv.org/pdf/2307.16715

work page Pith review arXiv 2023
[2]

VideoLLM-online: Online Video Large Language Model for Streaming Video

Chen, Joya, et al. "VideoLLM-online: Online Video Large Language Model for Streaming Video." ArXiv (Cornell University), 17 Jun 2024, https://arxiv.org/pdf/2406.11816

arXiv 2024
[3]

VideoMamba: State Space Model for Efficient Video Understanding

Li, Kunchang, et al. "VideoMamba: State Space Model for Efficient Video Understanding." ArXiv (Cornell University), 12 Mar 2024, https://arxiv.org/pdf/2403.06977

arXiv 2024
[4]

Mamba: Linear-time Sequence Modeling with Selective State Spaces

Gu, Albert and Dao, Tri. "Mamba: Linear-time Sequence Modeling with Selective State Spaces." ArXiv (Cornell University), 31 May 2024, https://arxiv.org/pdf/2312.00752

arXiv 2024
[5]

Is Space-Time Attention All You Need for Video Understanding?

Bertasius, Gedas, et al. "Is Space-Time Attention All You Need for Video Understanding?" ArXiv (Cornell University), 9 Jun 2021, https://arxiv.org/pdf/2102.05095

arXiv 2021
[6]

Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives

Grauman, Dristen, et al. "Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives." ArXiv (Cornell University), 25 Sep 2024, https://arxiv.org/pdf/2311.18259

arXiv 2024
[7]

Shotluck Holmes: A Family of Efficient Small-Scale Large Language Vision Models For Video Captioning and Summarization

Luo, Richard, et al. "Shotluck Holmes: A Family of Efficient Small-Scale Large Language Vision Models for Video Captioning and Summarization." ArXiV (Cornell University), 21 Oct 2024, https://arxiv. org/pdf/2405.20648

work page Pith review arXiv 2024
[8]

Shot2Story20K: A New Benchmark for Comprehensive Understanding of Multi-shot Videos

Han, Mingfei, et al. "Shot2Story20K: A New Benchmark for Comprehensive Understanding of Multi-shot Videos." ArXiV (Cornell University), 21 Oct 2024, https://arxiv.org/pdf/2312.10300

arXiv 2024

Show all 19 references

[9]

Advancing High-Resolution Video-Language Representation with Large-Scale Video Transcriptions

Xue, Hongwei, et al. "Advancing High-Resolution Video-Language Representation with Large-Scale Video Transcriptions." ArXiV (Cornell University), 8 Jul 2022, https://arxiv.org/pdf/2111.10337

2022 arXiv
[10]

Transnet v2: An effective deep network architecture for fast shot transition detection

Tomáš Souˇcek and Jakub Loko ˇc. "Transnet v2: An effective deep network architecture for fast shot transition detection." ArXiV (Cornell University), 11 Aug 2020, https://arxiv.org/pdf/2008.04838 12

2020 arXiv
[11]

Ego4D: Around the World in 3,000 Hours of Egocentric Video

Grauman, Kristen, et al. "Ego4D: Around the World in 3,000 Hours of Egocentric Video." ArXiV (Cornell University), 11 Mar 2022, https://arxiv.org/pdf/2110.07058

2022 arXiv
[12]

From Sparse to Dense: GPT-4 Summarization with Chain of Density Prompting

Adams, Griffin, et al. "From Sparse to Dense: GPT-4 Summarization with Chain of Density Prompting." ArXiV (Cornell University), 8 Sep 2023, https://arxiv.org/pdf/2309.04269

2023 arXiv
[13]

GPT-4 Technical Report

OpenAI. "GPT-4 Technical Report." ArXiV (Cornell University), 4 Mar 2024, https://arxiv.org/ pdf/2303.08774

2024 arXiv
[14]

Cluster-based Video Summarization with Temporal Context Awareness

Huynh-Lam, Hai-Dang, et al. "Cluster-based Video Summarization with Temporal Context Awareness." ArXiV (Cornell University), 9 Apr 2024, https://arxiv.org/pdf/2404.04511v1

2024 arXiv
[15]

Creating Summaries from User Videos

Gygli, Michael,et al. "Creating Summaries from User Videos." Computer Vision - ECCV 2014,https: //doi.org/10.1007/978-3-319-10584-0_33

2014 doi
[16]

Sigmoid Loss for Language Image Pre-Training

Zhai, Xiaohua, et al. "Sigmoid Loss for Language Image Pre-Training." ArXiV (Cornell University), 27 March 2023, https://arxiv.org/abs/2303.15343

2023 arXiv
[17]

BIRCH: An Efficient Data Clustering Method for Very Large Databases

Zhang, Tian, et al. "BIRCH: An Efficient Data Clustering Method for Very Large Databases." SIGMOD Rec. 25(2), 103–114, 1 June 1996, https://doi.org/10.1145/235968.233324

1996
[18]

Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

Wei, Jason, et al. "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models." ArXiV (Cornell University), 10 Jan 2023, https://arxiv.org/pdf/2201.11903

2023 arXiv
[19]

Nov 4, 2024, https://build.nvidia.com/nvidia/video-search-and-summarization 13

Nvidia. Nov 4, 2024, https://build.nvidia.com/nvidia/video-search-and-summarization 13

2024

Pith tools

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

[1] [1]

UniVTG: Towards Unified Video-Language Temporal Grounding

Lin, Kevin Q., et al. "UniVTG: Towards Unified Video-Lanaguage Temporal Grounding." ArXiv (Cornell University), 18 Aug 2023, https://arxiv.org/pdf/2307.16715

work page Pith review arXiv 2023

[2] [2]

VideoLLM-online: Online Video Large Language Model for Streaming Video

Chen, Joya, et al. "VideoLLM-online: Online Video Large Language Model for Streaming Video." ArXiv (Cornell University), 17 Jun 2024, https://arxiv.org/pdf/2406.11816

arXiv 2024

[3] [3]

VideoMamba: State Space Model for Efficient Video Understanding

Li, Kunchang, et al. "VideoMamba: State Space Model for Efficient Video Understanding." ArXiv (Cornell University), 12 Mar 2024, https://arxiv.org/pdf/2403.06977

arXiv 2024

[4] [4]

Mamba: Linear-time Sequence Modeling with Selective State Spaces

Gu, Albert and Dao, Tri. "Mamba: Linear-time Sequence Modeling with Selective State Spaces." ArXiv (Cornell University), 31 May 2024, https://arxiv.org/pdf/2312.00752

arXiv 2024

[5] [5]

Is Space-Time Attention All You Need for Video Understanding?

Bertasius, Gedas, et al. "Is Space-Time Attention All You Need for Video Understanding?" ArXiv (Cornell University), 9 Jun 2021, https://arxiv.org/pdf/2102.05095

arXiv 2021

[6] [6]

Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives

Grauman, Dristen, et al. "Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives." ArXiv (Cornell University), 25 Sep 2024, https://arxiv.org/pdf/2311.18259

arXiv 2024

[7] [7]

Shotluck Holmes: A Family of Efficient Small-Scale Large Language Vision Models For Video Captioning and Summarization

Luo, Richard, et al. "Shotluck Holmes: A Family of Efficient Small-Scale Large Language Vision Models for Video Captioning and Summarization." ArXiV (Cornell University), 21 Oct 2024, https://arxiv. org/pdf/2405.20648

work page Pith review arXiv 2024

[8] [8]

Shot2Story20K: A New Benchmark for Comprehensive Understanding of Multi-shot Videos

Han, Mingfei, et al. "Shot2Story20K: A New Benchmark for Comprehensive Understanding of Multi-shot Videos." ArXiV (Cornell University), 21 Oct 2024, https://arxiv.org/pdf/2312.10300

arXiv 2024