Beyond Interpretability: Exploring the Comprehensibility of Adaptive Video Streaming through Large Language Models Conference: Proceedings of the 33rd ACM International Conference on Multimedia; October 27–31, 2025; Dublin, Ireland Proceedings of the 33rd ACM International Conference on Multimedia (

Lianchen Jia, Chaoyang Li, Ziqi Yuan, Jiahui Chen, Tianchi Huang, Jiangchuan Liu, Lifeng Sun · 2025

Read Beyond Interpretability: Exploring the Comprehensibility of Adaptive Video Streaming through Large Language Models Conference: Proceedings of the 33rd ACM International Conference on Multimedia; October 27–31, 2025; Dublin, Ireland Proceedings of the 33rd ACM International Conference on Multimedia ( by Lianchen Jia, Chaoyang Li, Ziqi Yuan, Jiahui Chen, Tianchi Huang, Jiangchuan Liu, Lifeng Sun (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-16448.txt.

Introduction

Lianchen Jia and six co-workers wrote a 2025 paper on making video streaming algorithms easier for developers to grasp. Streaming tools pick a bitrate on the fly, often with deep learning. Those black-box choices are hard to read and tune. Past work turned them into decision trees, but a readable tree is not the same as one a person finds clear.

Their frame, ComTree, first builds all trees that meet a performance bar. A large language model then rates the trees for how clear they feel to a developer. The best-rated tree is picked. Tests showed clearer output with like performance. Code was shared.

Catalog record

CREATORLianchen Jia, Chaoyang Li, Ziqi Yuan, Jiahui Chen, Tianchi Huang, Jiangchuan Liu, Lifeng Sun
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE41 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-16448

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Beyond Interpretability: Exploring the Comprehensibility of Adaptive Video Streaming through Large Language Models Conference: Proceedings of the 33rd ACM International Conference on Multimedia; October 27–31, 2025; Dublin, Ireland Proceedings of the 33rd ACM International Conference on Multimedia (MM ’25), October 27–31, 2025, Dublin, Ireland DOI: 10.1145/3746027.3755257 ISBN: 979-8-4007-2035-2/2025/10 CCS: Information systems Multimedia streaming Lianchen Jia, Chaoyang Li, Ziqi Yuan, Jiahui Chen, Tianchi Huang, Jiangchuan Liu, Lifeng Sun Abstract . Over the past decade, adaptive video streaming technology has witnessed significant advancements, particularly driven by the rapid evolution of deep learning techniques. However, the black-box nature of deep learning algorithms presents challenges for developers in understanding decision-making processes and optimizing for specific application scenarios. Although existing research has enhanced algorithm interpretability through decision tree conversion, interpretability does not directly equate to developers’ subjective comprehensibility. To address this challenge, we introduce ComTree , the first bitrate adaptation algorithm generation framework that considers comprehensibility. The framework initially generates the complete set of decision trees that meet performance requirements, then leverages large language models to evaluate these trees for developer comprehensibility, ultimately selecting solutions that best facilitate human understanding and enhancement. Experimental results demonstrate that ComTree significantly improves comprehensibility while maintaining competitive performance, showing potential for further advancement. The source code is available at https://github.com/thu-media/ComTree . † † cc-license: by 1. INTRODUCTION With the advancement of networking technologies ( Narayanan et al. 2021 ; Khorov et al. 2018 ) and the growing desire for content creation ( Pires and Simon 2015 ) , video has become an indispensable part of people’s daily lives. According to the 2024 Global Internet Phenomena Report ( sandvine 2024 ) , video applications are the primary contributors to downstream traffic, accounting for 39% of fixed network traffic and 31% of mobile network traffic. Within content categories, on-demand streaming represents 54% of downstream traffic in fixed networks and 57% in mobile networks, making it the largest subcategory. Furthermore, according to the 2025 Global Internet Phenomena Report ( Networks 2025 ) , the top three major video applications have experienced traffic growth of 15%–51% in fixed networks and 7%–38% in mobile networks compared to the previous year. In video streaming technology, Adaptive Bitrate (ABR) has emerged as the mainstream approach for enhancing the Quality of Experience (QoE) of network video traffic. ABR has been evolving for more than a decade ( Sani et al. 2017 ) . These algorithms select the appropriate bitrate level for the next segment based on the current buffer size and historical bandwidth conditions. ABR algorithms aim to enhance the user’s QoE by improving video quality and playback smoothness while reducing stalling events. Traditional heuristic algorithms optimize the bitrate adaptation process by modeling the playback process ( Huang et al. 2014 ; Yin et al. 2015 ; Spiteri et al. 2020 ) . In recent years, deep learning algorithms have shown promising results in bitrate adaptation ( Mao et al. 2017 ; Huang et al. 2019 ; Jia et al. 2023 ; Jia et al. 2024 ; Wu et al. 2024 ) . These algorithms have the potential to automatically adapt to changes in the environment. However, when deploying these algorithms in production environments, the black-box optimization nature of deep learning brings difficulties for developers to understand. Improving the interpretability of learning-based ABR models has become a hot topic of interest. These works either understand the importance of black-box model states through post-hoc analysis ( Dethise et al. 2019 ) or transform the black-box models into decision trees for deployment ( Meng et al. 2019 ; Grüner et al. 2020 ) . However, for ABR algorithms, merely transforming them into white-box algorithms with interpretability for each decision step is insufficient. A practical challenge in production environments is whether developers can comprehend and further optimize the algorithm to adapt to new network conditions - this refers to the algorithm’s comprehensibility ( Huysmans et al. 2011 ; Freitas 2014 ; Martens et al. 2011 ; Freitas 2006 ; Pazzani 2000 ) . Although previous approaches using greedy algorithms to generate decision trees provide reasons for each decision step, complex decision logic ( Grüner et al. 2020 ; Meng et al. 2019 ) may still pose comprehension challenges for developers (e.g., tree depth, node selection/organization). This complexity creates barriers when developers attempt to adjust and optimize the decision tree. Therefore, our objective extends beyond mere interpretability, aiming to enhance algorithm comprehensibility and ultimately maximize the potential for developer-driven improvements. While enhancing the comprehensibility of ABR shows promising potential, two significant challenges emerge. The first challenge lies in improving algorithm comprehensibility without substantially compromising performance. Current decision tree generation algorithms prioritize performance through greedy node selection until growth stops or node limits are reached ( Loh 2011 ) . However, merely simplifying tree structures, such as depth limitations, inevitably leads to performance degradation. The second challenge concerns defining developer comprehensibility of algorithms. From a developer’s perspective, algorithm comprehensibility is inherently subjective and cannot be easily optimized as an objective function ( Martens et al. 2011 ) . Simple metrics such as tree depths and leaf node counts may inadequately capture all factors affecting developer understanding ( Huysmans et al. 2011 ) . Therefore, we present ComTree , the first framework that incorporates comprehensibility into the generation process of bitrate adaptation algorithms. ComTree aims to enhance algorithmic comprehensibility while maintaining performance through a two-phase approach: first constructing a comprehensive set of heterogeneous decision trees with comparable performance (known as the Rashomon set ( Fisher et al. 2019 ; Semenova et al. 2022 ) ), then conducting comprehensibility assessment using Large Language Models (LLMs) ( Zhao et al. 2023b ) . To construct the Rashomon set, we first compress the ABR state space through feature elimination based on black-box model prior knowledge ( McTavish et al. 2022 ) , significantly reducing the time complexity of optimal decision tree generation. Using the compressed state space, we generate training datasets that align with actual playback probabilities through a teacher-student learning framework similar to previous work ( Bastani et al. 2018 ; Meng et al. 2019 ) . Based on this dataset, we employ a thoroughly pruned dynamic programming algorithm ( Xin et al. 2022 ) to obtain a complete set of heterogeneous decision trees within a predetermined performance error range. For the comprehensibility assessment phase, considering that LLMs have demonstrated human-like capabilities in various comprehension tasks through their learned world knowledge ( Achiam et al. 2023 ; Kaddour et al. 2023 ) , we leverage LLMs as algorithm evaluators in place of developers. Inspired by the Human-Feedback methodology ( Ouyang et al. 2022 ) , we implement a pairwise comparison mechanism similar to human preference learning. To enhance evaluation reliability, we utilize an ensemble approach of LLMs ( Liang et al. 2023 ) , leveraging multiple LLMs to compare comprehensibility within the Rashomon set until consensus is reached. To eliminate the influence of similar decision trees, we employ a two-phase comparison method that achieves the theoretical minimum number of comparisons. Our experimental results demonstrate that ComTree achieves superior comprehensibility while maintaining competitive performance compared to existing algorithms. In comparison with existing research ( Meng et al. 2019 ) , ComTree achieves better performance using only

Read the complete text