arxiv: 2509.08016 · v2 · submitted 2025-09-09 · 💻 cs.CV · cs.LG

Video Parallel Scaling: Aggregating Diverse Frame Subsets for VideoLLMs

Hyungjin Chung , Hyelin Nam , Jiyeon Kim , Hyojun Go , Byeongjun Park , Junho Kim , Joonseok Lee , Seongsu Ha

show 1 more author

Byung-Hoon Kim

This is my paper

Pith reviewed 2026-05-18 18:31 UTC · model grok-4.3

classification 💻 cs.CV cs.LG

keywords Video Parallel ScalingVideoLLMsdisjoint frame subsetsinference-time scalingChinchilla scaling lawtemporal reasoningparallel inferenceprobability aggregation

0 comments

The pith

Aggregating output probabilities from parallel inferences on disjoint video frame subsets improves VideoLLM performance without extra training or longer context.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents Video Parallel Scaling, an inference-time method that runs several parallel streams, each on a different disjoint subset of frames from the same video, then combines the resulting output probabilities. This lets the model draw on a broader set of visual evidence than any single pass through a fixed context window could provide. The authors claim that because the evidence from these subsets tends to be uncorrelated, the aggregation produces a net gain that effectively contracts the Chinchilla scaling law. Experiments on models ranging from 2B to 32B parameters across standard video benchmarks show consistent accuracy lifts. A reader would care because the method promises better temporal reasoning at no additional training cost and with only modest extra inference compute.

Core claim

Video Parallel Scaling contracts the Chinchilla scaling law at inference time by processing disjoint frame subsets in parallel streams and aggregating their output probabilities, thereby integrating richer uncorrelated visual evidence and raising performance on video understanding tasks without any additional training or expansion of the context window.

What carries the argument

Aggregation of output probabilities across multiple parallel inference streams, each operating on a unique disjoint subset of the input video frames.

If this is right

Performance improves consistently on Video-MME and EventHallusion for models from 2B to 32B parameters.
The method scales more favorably than self-consistency and remains complementary to other decoding strategies.
Temporal reasoning capabilities of VideoLLMs increase without raising memory usage from longer context windows.
No retraining is required, so the technique applies directly to existing deployed models.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the uncorrelated-evidence premise holds for other modalities, the same parallel-subset aggregation could be tested on image or audio models that currently hit context limits.
The approach suggests a general way to trade parallel inference compute for effective data scaling, which could be quantified on controlled synthetic videos where frame correlation is known in advance.
Memory-efficient inference becomes feasible for hour-long videos by keeping each stream short while still harvesting diverse evidence through aggregation.

Load-bearing premise

The visual evidence obtained from the different disjoint frame subsets is sufficiently uncorrelated that combining their probabilities produces a performance gain equivalent to contracting the Chinchilla scaling law.

What would settle it

Measure the statistical correlation between the per-token output distributions produced by different frame-subset streams; if gains vanish or reverse once measured correlation exceeds a modest threshold, the central claim is falsified.

Figures

Figures reproduced from arXiv: 2509.08016 by Byeongjun Park, Byung-Hoon Kim, Hyelin Nam, Hyojun Go, Hyungjin Chung, Jiyeon Kim, Joonseok Lee, Junho Kim, Seongsu Ha.

**Figure 1.** Figure 1: Conceptual Illustration of VPS. VideoLLMs take as input subsampled frames from the original video1 , limiting their understanding capabilities. Increasing the sampled frames within context leads to computation/memory issues or decrease in performance. In contrast, VPS keeps the number of subsampled frames, and scales the number of streams in parallel, with each stream attending to different frames. By aggr… view at source ↗

**Figure 2.** Figure 2: VPS consistently improves performance across all dimensions. Across 3 different model classes (Qwen-2.5-VL, InternVL3, Gemma3), 3 different size (2B - 32B), and number of frames used in context, VPS offers improved results with clearer trends with larger models. y-axis denotes the accuracy in the EventHallusion binary QA. 4 EXPERIMENTS Experimental settings We test our method on 3 different model classes, … view at source ↗

**Figure 3.** Figure 3: VPS scales better for longer videos. Comparing the results of Qwen2.5-VL-7B on Video-MME for each category, we see a clearer trend in the long video category (15 - 30 min.). 2 4 8 16 Number of Frames 0.425 0.450 0.475 0.500 0.525 0.550 0.575 Accuracy (%) Qwen2.5-VL 2 4 8 16 Number of Frames 0.50 0.52 0.54 0.56 0.58 0.60 0.62 Accuracy (%) InternVL3 2 4 8 16 Number of Frames 0.47 0.48 0.49 0.50 0.51 0.52 Acc… view at source ↗

**Figure 4.** Figure 4: VPS scales favorably compared to Self-consistency. On Video-MME, VPS outperforms Self-consistency under the same budget by being able to incorporate information from different frames, rather than relying on the same information for all the streams. plateau or decrease as one incorporates more frames into the context, as can also be observed in the plot in [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

**Figure 5.** Figure 5: VPS consistently improves performance across all dimensions. Across 3 different model classes (Qwen-2.5-VL, InternVL3, Gemma3), 3 different size (2B - 32B), and number of frames used in context. y-axis denotes the accuracy in the Video-MME multiple-choice QA. Qualitative results We present additional examples of the free form description task in Tab. 8. C EXPERIMENTAL DETAILS C.1 MAIN EXPERIMENT For Video-… view at source ↗

read the original abstract

Video Large Language Models (VideoLLMs) face a critical bottleneck: increasing the number of input frames to capture fine-grained temporal detail leads to prohibitive computational costs and performance degradation from long context lengths. We introduce Video Parallel Scaling (VPS), an inference-time method that expands a model's perceptual bandwidth without increasing its context window. VPS operates by running multiple parallel inference streams, each processing a unique, disjoint subset of the video's frames. By aggregating the output probabilities from these complementary streams, VPS integrates a richer set of visual information than is possible with a single pass. We theoretically show that this approach effectively contracts the Chinchilla scaling law by leveraging uncorrelated visual evidence, thereby improving performance without additional training. Extensive experiments across various model architectures and scales (2B-32B) on benchmarks such as Video-MME and EventHallusion demonstrate that VPS consistently and significantly improves performance. It scales more favorably than other parallel alternatives (e.g. Self-consistency) and is complementary to other decoding strategies, offering a memory-efficient and robust framework for enhancing the temporal reasoning capabilities of VideoLLMs.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

VPS is a practical inference-time trick that aggregates outputs from disjoint video frame subsets and shows consistent gains, but the claimed contraction of the Chinchilla law is not derived from the scaling equations.

read the letter

The main thing to know is that this paper describes Video Parallel Scaling, which splits a video into disjoint frame subsets, runs separate inferences on each, and combines the output probabilities. The goal is to pull in more visual information without lengthening the context window or retraining the model. They test it on models from 2B to 32B parameters and report improvements on Video-MME and EventHallusion, plus better scaling than self-consistency and compatibility with other decoding methods. Those empirical results are the clearest part of the work and look like something practitioners could try directly. The experiments span multiple architectures and scales, which gives the gains some credibility as a general inference-time lever. The soft spot is the theoretical section. The abstract states that the method contracts the Chinchilla scaling law by using uncorrelated visual evidence, yet it supplies no derivation that starts from the loss function L(N,D) and inserts a correlation term or effective-token multiplier to show how logit averaging reduces loss with the original exponents. The assumption that the subsets are independent enough for linear information gain is invoked but not shown from first principles. If the full paper has a clean proof or even a simple effective-D calculation, that would strengthen the claim; otherwise it reads as an empirical observation wrapped in scaling-law language. This paper is for engineers and researchers who already work with VideoLLMs and need cheap ways to handle longer videos at test time. A reader focused on practical performance lifts will find the benchmarks and ablations useful. The work is coherent enough on its own terms to deserve a serious referee, even with the theory gap. I would send it to peer review and ask the authors to expand the derivation and add more controls on how the aggregation is implemented.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces Video Parallel Scaling (VPS), an inference-time technique for VideoLLMs. VPS runs multiple parallel inference streams on disjoint subsets of video frames and aggregates their output probabilities to integrate richer visual information without extending the context window or additional training. The authors claim this approach theoretically contracts the Chinchilla scaling law by leveraging uncorrelated visual evidence from the subsets. They report consistent empirical gains across model scales (2B–32B) on benchmarks including Video-MME and EventHallusion, with more favorable scaling than self-consistency and complementarity to other decoding strategies.

Significance. If the claimed contraction of the Chinchilla law holds, the work would offer a practical, memory-efficient route to increasing perceptual bandwidth in VideoLLMs at inference time. The broad experimental coverage across architectures and scales, together with the reported complementarity to existing decoding methods, constitutes a concrete strength. The absence of a first-principles derivation for the scaling-law claim, however, limits the current theoretical contribution.

major comments (2)

[Abstract and Theoretical Analysis section] Abstract and Theoretical Analysis section: The central claim that VPS 'effectively contracts the Chinchilla scaling law by leveraging uncorrelated visual evidence' is load-bearing for the paper's novelty yet lacks any derivation. No steps are shown that start from the Chinchilla form L(N,D) ≈ E + A/N^α + B/D^β, introduce a correlation factor ρ between frame-subset outputs, and demonstrate an effective increase in D (or equivalent loss reduction) with the original exponents.
[Method section] Method section: The aggregation procedure itself (whether logits or probabilities are averaged, how ties or final answer selection is handled, and the precise definition of 'disjoint' subsets) is not specified. This detail is required both to reproduce the reported gains and to evaluate whether the uncorrelated-evidence premise actually holds.

minor comments (2)

[Experiments] The manuscript would benefit from error bars or statistical significance tests on the Video-MME and EventHallusion results to strengthen the empirical support.
[Figure 1] Figure 1 or the VPS diagram could more explicitly illustrate the frame-subset partitioning and probability-aggregation step for clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thoughtful review and constructive suggestions. We address each of the major comments below and have prepared revisions to the manuscript to incorporate the requested clarifications and expansions.

read point-by-point responses

Referee: [Abstract and Theoretical Analysis section] Abstract and Theoretical Analysis section: The central claim that VPS 'effectively contracts the Chinchilla scaling law by leveraging uncorrelated visual evidence' is load-bearing for the paper's novelty yet lacks any derivation. No steps are shown that start from the Chinchilla form L(N,D) ≈ E + A/N^α + B/D^β, introduce a correlation factor ρ between frame-subset outputs, and demonstrate an effective increase in D (or equivalent loss reduction) with the original exponents.

Authors: We acknowledge that the derivation in the Theoretical Analysis section could benefit from greater explicitness. Although the manuscript outlines the conceptual basis for contracting the Chinchilla scaling law through uncorrelated visual evidence, we agree with the referee that a detailed step-by-step derivation is necessary. In the revised manuscript, we will add the following derivation in the Theoretical Analysis section: Starting from the Chinchilla form L(N,D) ≈ E + A/N^α + B/D^β, we model the effective data D_eff = D * (1 + (1-ρ) * (k-1)) where k is the number of subsets and ρ is the correlation between subset outputs. When ρ is low due to disjoint frames, D_eff increases, effectively contracting the scaling curve. We will include the full mathematical steps and assumptions. revision: yes
Referee: [Method section] Method section: The aggregation procedure itself (whether logits or probabilities are averaged, how ties or final answer selection is handled, and the precise definition of 'disjoint' subsets) is not specified. This detail is required both to reproduce the reported gains and to evaluate whether the uncorrelated-evidence premise actually holds.

Authors: We thank the referee for pointing out this ambiguity. In the revised Method section, we will specify that VPS averages the output probabilities (not logits) from each parallel inference stream. The 'disjoint' subsets are constructed by evenly partitioning the total frames into non-overlapping groups, ensuring no frame overlap. For final answer selection, we take the class or token with the maximum aggregated probability; in case of ties, we break them by selecting the first in lexicographical order. We will also add a discussion on how this aggregation leverages the uncorrelated evidence premise, supported by empirical correlation measurements in the experiments. revision: yes

Circularity Check

0 steps flagged

No circularity; theoretical claim rests on stated premise without reduction to inputs

full rationale

The paper claims to theoretically show that VPS contracts the Chinchilla scaling law via aggregation of uncorrelated visual evidence from disjoint frame subsets. The abstract presents this as a first-principles benefit supporting performance gains without additional training, but supplies no equations, fitted parameters, or self-citations that reduce the contraction result to the modeling assumption by construction. The uncorrelated-evidence premise is invoked to justify the scaling benefit and is treated as an input rather than derived within the visible text; empirical results on Video-MME and EventHallusion are reported as separate validation. No load-bearing step matches the enumerated circularity patterns, so the derivation chain remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the unproven premise that frame-subset outputs are uncorrelated enough to contract the Chinchilla scaling law; no free parameters or new entities are explicitly introduced in the abstract.

axioms (1)

domain assumption Outputs from disjoint frame subsets provide uncorrelated visual evidence whose aggregation contracts the Chinchilla scaling law
Invoked in the abstract to explain both theoretical benefit and performance gains

pith-pipeline@v0.9.0 · 5754 in / 1306 out tokens · 36075 ms · 2026-05-18T18:31:06.263400+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

29 extracted references · 29 canonical work pages · 13 internal anchors

[1]

Qwen2.5-VL Technical Report

URL https://openreview.net/forum?id= eoln5WgrPx. Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2.5-VL technical report.arXiv preprint arXiv:2502.13923,

work page internal anchor Pith review Pith/arXiv arXiv
[2]

LongBench v2: Towards Deeper Understanding and Reasoning on Realistic Long-context Multitasks

Yushi Bai, Shangqing Tu, Jiajie Zhang, Hao Peng, Xiaozhi Wang, Xin Lv, Shulin Cao, Jiazheng Xu, Lei Hou, Yuxiao Dong, et al. LongBench v2: Towards deeper understanding and reasoning on realistic long-context multitasks.arXiv preprint arXiv:2412.15204,

work page internal anchor Pith review Pith/arXiv arXiv
[3]

Parallel scaling law for language models.arXiv preprint arXiv:2505.10475,

Mouxiang Chen, Binyuan Hui, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Jianling Sun, Junyang Lin, and Zhongxin Liu. Parallel scaling law for language models.arXiv preprint arXiv:2505.10475,

work page arXiv
[4]

Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

URL https://openreview.net/forum?id= 6PmJoRfdaK. Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261,

work page internal anchor Pith review Pith/arXiv arXiv
[5]

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

URL https: //arxiv.org/abs/2010.11929. Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenenbaum, and Igor Mordatch. Improving factuality and reasoning in language models through multiagent debate. InForty-first International Conference on Machine Learning,

work page internal anchor Pith review Pith/arXiv arXiv 2010
[6]

Framefusion: Combining similarity and importance for video token reduction on large visual language models.arXiv preprint arXiv:2501.01986,

Tianyu Fu, Tengxuan Liu, Qinghao Han, Guohao Dai, Shengen Yan, Huazhong Yang, Xuefei Ning, and Yu Wang. Framefusion: Combining similarity and importance for video token reduction on large visual language models.arXiv preprint arXiv:2501.01986,

work page arXiv
[7]

Exploring hallucination of large multimodal models in video understanding: Benchmark, analysis and mitigation.arXiv preprint arXiv:2503.19622,

Hongcheng Gao, Jiashu Qu, Jingyi Tang, Baolong Bi, Yue Liu, Hongyu Chen, Li Liang, Li Su, and Qingming Huang. Exploring hallucination of large multimodal models in video understanding: Benchmark, analysis and mitigation.arXiv preprint arXiv:2503.19622,

work page arXiv
[8]

Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al

Model card, accessed 25/Jul/2025. Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. Training compute-optimal large language models. InProceedings of the 36th International Conference on Neural Information Processing Systems, pp. ...

work page 2025
[9]

GPT-4o System Card

URL https://openreview.net/forum?id= kIoBbc76Sy. Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Os- trow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card.arXiv preprint arXiv:2410.21276,

work page internal anchor Pith review Pith/arXiv arXiv
[10]

Multi-granular spatio-temporal token merging for training-free acceleration of video llms.arXiv preprint arXiv:2507.07990,

Jeongseok Hyun, Sukjun Hwang, Su Ho Han, Taeoh Kim, Inwoong Lee, Dongyoon Wee, Joon- Young Lee, Seon Joo Kim, and Minho Shim. Multi-granular spatio-temporal token merging for training-free acceleration of video llms.arXiv preprint arXiv:2507.07990,

work page arXiv
[11]

OpenAI o1 System Card

ISSN 2835-8856. URLhttps://openreview.net/forum? id=H4S4ETc8c9. Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card.arXiv preprint arXiv:2412.16720,

work page internal anchor Pith review Pith/arXiv arXiv
[12]

Regularized best-of-n sampling to mitigate reward hacking for language model alignment

Yuu Jinnai, Tetsuro Morimura, Kaito Ariu, and Kenshi Abe. Regularized best-of-n sampling to mitigate reward hacking for language model alignment. InICML 2024 Workshop on Models of Human Feedback for AI Alignment,

work page 2024
[13]

Improving LLM Video Understanding with 16 Frames Per Second

Yixuan Li, Changli Tang, Jimin Zhuang, Yudong Yang, Guangzhi Sun, Wei Li, Zejun Ma, and Chao Zhang. Improving LLM Video Understanding with 16 Frames Per Second. InProceedings of the 42nd International Conference on Machine Learning (ICML), 2025c. Chengzhi Liu, Zhongxing Xu, Qingyue Wei, Juncheng Wu, James Zou, Xin Eric Wang, Yuyin Zhou, and Sheng Liu. Mor...

work page arXiv
[14]

Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation

URLhttps://arxiv.org/abs/2108.12409. Kanchana Ranasinghe, Xiang Li, Kumara Kahatapitiya, and Michael S Ryoo. Understanding long videos with multimodal language models. InThe Thirteenth International Conference on Learning Representations,

work page internal anchor Pith review Pith/arXiv arXiv
[15]

Jingzhe Shi, Qinwei Ma, Hongyi Liu, Hang Zhao, Jeng-Neng Hwang, and Lei Li

URLhttps://openreview.net/forum?id=OxKi02I29I. Jingzhe Shi, Qinwei Ma, Hongyi Liu, Hang Zhao, Jeng-Neng Hwang, and Lei Li. Explaining context length scaling and bounds for language models.arXiv preprint arXiv:2502.01481,

work page arXiv
[16]

RoFormer: Enhanced Transformer with Rotary Position Embedding

URLhttps://arxiv.org/abs/2104.09864. Hanshi Sun, Momin Haider, Ruiqi Zhang, Huitao Yang, Jiahao Qiu, Ming Yin, Mengdi Wang, Peter Bartlett, and Andrea Zanette. Fast Best-of-N Decoding via Speculative Rejection. In The Thirty-eighth Annual Conference on Neural Information Processing Systems,

work page internal anchor Pith review Pith/arXiv arXiv
[17]

Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context

Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context.arXiv preprint arXiv:2403.05530,

work page internal anchor Pith review Pith/arXiv arXiv
[18]

Gemma 3 Technical Report

Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, et al. Gemma 3 technical report.arXiv preprint arXiv:2503.19786,

work page internal anchor Pith review Pith/arXiv arXiv
[19]

Time Blindness: Why Video-Language Models Can't See What Humans Can?

Ujjwal Upadhyay, Mukul Ranjan, Zhiqiang Shen, and Mohamed Elhoseiny. Time blindness: Why video-language models can’t see what humans can?arXiv preprint arXiv:2505.24867,

work page internal anchor Pith review Pith/arXiv arXiv
[20]

Inference compute-optimal video vision language models.arXiv preprint arXiv:2505.18855, 2025a

Peiqi Wang, ShengYun Peng, Xuewen Zhang, Hanchao Yu, Yibo Yang, Lifu Huang, Fujun Liu, and Qifan Wang. Inference compute-optimal video vision language models.arXiv preprint arXiv:2505.18855, 2025a. Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought r...

work page arXiv
[21]

Inference Scaling Laws: An Empirical Analysis of Compute-Optimal Inference for Problem-Solving with Language Models

Yangzhen Wu, Zhiqing Sun, Shanda Li, Sean Welleck, and Yiming Yang. Inference scaling laws: An empirical analysis of compute-optimal inference for problem-solving with language models.arXiv preprint arXiv:2408.00724,

work page internal anchor Pith review Pith/arXiv arXiv
[22]

Effective long-context scaling of foundation models

Wenhan Xiong, Jingyu Liu, Igor Molybog, Hejia Zhang, Prajjwal Bhargava, Rui Hou, Louis Martin, Rashi Rungta, Karthik Abinav Sankararaman, Barlas Oguz, et al. Effective long-context scaling of foundation models. InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (...

work page 2024
[23]

Slowfast-llava: A strong training-free baseline for video large language models

Mingze Xu, Mingfei Gao, Zhe Gan, Hong-You Chen, Zhengfeng Lai, Haiming Gang, Kai Kang, and Afshin Dehghan. Slowfast-llava: A strong training-free baseline for video large language models. arXiv preprint arXiv:2407.15841,

work page arXiv
[24]

Eventhallusion: Diagnosing event hallucinations in videoLLMs.arXiv preprint arXiv:2409.16597,

Jiacheng Zhang, Yang Jiao, Shaoxiang Chen, Na Zhao, Zhiyu Tan, Hao Li, and Jingjing Chen. Eventhallusion: Diagnosing event hallucinations in videoLLMs.arXiv preprint arXiv:2409.16597,

work page arXiv
[25]

InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Models

Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, et al. InternVL3: Exploring advanced training and test-time recipes for open-source multimodal models.arXiv preprint arXiv:2504.10479,

work page internal anchor Pith review Pith/arXiv arXiv
[26]

13 Preprint A PROOFS In Chen et al. (2025), the inputs to the parallel streams are learnable transformation of the same input x, so that one can assume that each parallel stream follows (7) in an unbiased way, leading to a simplification in the analysis of the parallel scaling law. We start by reviewing the result from Chen et al. (2025). Lemma 1(Chen et ...

work page 2025
[27]

6, we compare the results of logit averaging and probability averaging when implementing VPS

+O(∆ 3)(41) B FURTHERRESULTS Probability and logit averagingIn Tab. 6, we compare the results of logit averaging and probability averaging when implementing VPS. Across different model classes, we find that both approaches lead to similar results. Thus, while we assume probability averaging in the theoretical analysis for simplicity, we resort to logit av...

work page arXiv
[28]

(2024) and use the prompt specified in Tab

when evaluating the free form descriptions of the video, we follow Zhang et al. (2024) and use the prompt specified in Tab

work page 2024
[29]

Let x′ be the frame-dropped version of the sub-sampled video

C.4 INCORPORATING OTHER STRATEGIES For TCD, we construct a negative stream so that the half the frames are zeroed-out in an interleaved fashion. Let x′ be the frame-dropped version of the sub-sampled video. Then, TCD is implemented with ˜pθ(y|x) = (1 +α)p θ(y|x)−αp θ(y|x′),(42) where α∈[0,1) is a constant. Additionally, we set a hyperparameter β∈[0,1] tha...

work page 2023