arXiv:2604.08703v1 [cs.MM] 9 Apr 2026
QoS-QoE Translation with Large Language Model Yingjie Yu
Mingyuan Wu
Ahmadreza Eslaminia
[email protected] University of Illinois Urbana-Champaign Urbana, Illinois, USA
[email protected] University of Illinois Urbana-Champaign Urbana, Illinois, USA
[email protected] University of Illinois Urbana-Champaign Urbana, Illinois, USA
Lingzhi Zhao
Kaizhuo Yan
Klara Nahrstedt
[email protected] University of Illinois Urbana-Champaign Urbana, Illinois, USA
[email protected] University of Illinois Urbana-Champaign Urbana, Illinois, USA
[email protected] University of Illinois Urbana-Champaign Urbana, Illinois, USA
Abstract
quality of the delivered service [13]. Understanding the QoS-QoE relationship is important for multimedia applications because it supports system design, adaptive streaming, network optimization, and user-centered quality prediction [1, 4]. A large body of prior work has studied QoS-QoE relationships in multimedia applications, especially video streaming, by modeling how QoS factors map to perceived QoE, and in some cases how QoE targets guide adaptation decisions, using methods such as subjective experiments, heuristic rules, analytical modeling, and machine learning-based prediction [1, 5, 20, 33]. These studies have clarified how factors such as bitrate adaptation, stalling, startup delay, resolution changes, and network impairments affect perceived quality. However, many of these approaches are developed for particular setups and validated under specific conditions, which makes their findings and models difficult to generalize across scenarios. Applying them to new settings often requires substantial re-modeling, additional measurements, or new subjective studies. These limitations motivate a more unified QoS-QoE translation capability that can support both forward translation from QoS to QoE and reverse translation from QoE targets to QoS conditions across diverse scenarios. Achieving this goal requires both strong models and high-quality data. LLM-based systems are a promising foundation because they have shown strong potential in multimedia-related tasks such as video understanding, audio processing, and multimodal agent-style decision making, while also supporting flexible reasoning and structured prediction [12, 15, 17, 26]. At the same time, constructing suitable source-grounded data is challenging because QoS-QoE relationships are scattered across the literature, reported in heterogeneous forms such as text, tables, figures, and equations, and often accompanied by incomplete or implicit contextual metadata. To address this gap, we present QoS-QoE Translation dataset, a source-grounded dataset of structured QoS-QoE relationships from the literature, with a current focus on video streaming. Our goal is to transform prior studies into a reusable data resource for the multimedia community. Each entry preserves the extracted relationship together with supporting evidence and contextual metadata, enabling interpretability and reproducibility. Because reported QoS-QoE relationships rarely appear in a single uniform format and often must be recovered from multiple forms of source evidence together with their surrounding context, we construct the dataset through a pipeline that combines paper curation, QoS-QoE
QoS-QoE translation is a fundamental problem in multimedia systems because it characterizes how measurable system and network conditions affect user-perceived experience. Although many prior studies have examined this relationship, their findings are often developed for specific setups and remain scattered across papers, experimental settings, and reporting formats, limiting systematic reuse, cross-scenario generalization, and large-scale analysis. To address this gap, we first introduce QoS-QoE Translation dataset, a source-grounded dataset of structured QoS-QoE relationships from the multimedia literature, with a focus on video streaming related tasks. We construct the dataset through an automated pipeline that combines paper curation, QoS-QoE relationship extraction, and iterative data evaluation. Each record preserves the extracted relationship together with parameter definitions, supporting evidence, and contextual metadata. We further evaluate the capability of large language models (LLMs) on QoS-QoE translation, both before and after supervised fine-tuning on our dataset, and show strong performance on both continuous-value and discrete-label prediction in bidirectional translation, from QoSQoE and QoE-QoS. Our dataset provides a foundation for benchmarking LLMs in QoS-QoE translation and for supporting future LLM-based reasoning for multimedia quality prediction and optimization. The complete dataset and code are publicly available at https://yyu6969.github.io/qos-qoe-translation-page/, for full reproducibility and open access.
CCS Concepts • Information systems → Multimedia databases; • Computing methodologies → Information extraction; • Networks → Network performance analysis.
Keywords Quality of Service, Quality of Experience, Large Language Model, Multimedia Databases, Multimedia Systems, Benchmark Dataset
1
Introduction
Quality of Service (QoS) and Quality of Experience (QoE) are two central concepts in multimedia systems. QoS describes measurable system, network, and service conditions such as bitrate, delay, jitter, packet loss, and rebuffering, while QoE reflects users’ perceived 1
Paper Curation
QoS–QoE Relationship Extraction
Iterative Data Evaluation
metadata-enriched records
LLM data reviewers
core relationship records
(𝑟! , 𝑐! , 𝑚! )
(𝑟̅ ≥ 8) ∧ (𝑐̅ ≥ 3)
(𝑟" , 𝑐" , 𝑚" )
⋮
⋮
(𝑟# , 𝑐# , 𝑚# )
prompts
𝑟̅ , 𝑐̅, {𝑚$ #$%! )
(𝑟̅ ≤ 4)
+ otherwise Re-evaluation
Figure 1: Overview of the QoS-QoE Translation dataset construction pipeline. The pipeline begins with paper curation, followed by QoS-QoE relationship extraction and iterative data evaluation. relationship extraction, and iterative data evaluation, as shown in Figure 1. This design supports large-scale dataset construction while maintaining quality control and traceability. To assess the utility of QoS-QoE Translation, we perform supervised fine-tuning (SFT) of large language models (LLMs) on bidirectional QoS-QoE translation tasks and evaluate both continuous value and discrete label prediction. Results show strong performance gain, with the best fine-tuned model achieving 90.24% Accuracy for discrete label prediction and 8.49% MAPE (Mean Absolute Percentage Error) for continuous value prediction. These findings suggest that QoS-QoE Translation provides a strong foundation for training LLMs to reason about QoS-QoE relationships and opens up new opportunities for applying LLM in multimedia applications. The main contributions of this work are three-fold: 1) We introduce QoS-QoE Translation, a source-grounded dataset of structured QoS-QoE relationships from the literature, with a current focus on video streaming. 2) We develop a reusable dataset construction pipeline for paper curation, relationship extraction, metadata enrichment, and iterative multi-reviewer quality evaluation. 3) We demonstrate that the dataset supports effective SFT of LLMs for bidirectional QoS-QoE translation and are the first to benchmark existing open-source LLMs in this domain.
the practical meaning of reported QoS-QoE relationships over time. This restriction emphasizes recent and practically relevant evidence while reducing noise from older system settings. Using this humanAI curation process, we collect 505 papers related to QoS-QoE relationships in video streaming, which serve as the foundation for the downstream extraction pipeline and improve the relevance and reliability of the final dataset.
2.2
QoS-QoE Relationship Extraction
Figure 1 overviews the QoS-QoE Translation construction pipeline, which includes paper curation, QoS-QoE relationship extraction, and iterative data evaluation. Together, these stages transform curated papers into structured records and improve their quality through iterative review. Although QoS-QoE Translation focuses on video streaming, the pipeline is reusable and can be adapted to other application domains that require extracting source-grounded relationships from the literature.
Starting from the curated paper corpus, we perform QoS-QoE relationship extraction from academic papers. Because the source papers are provided in PDF format, we first use MinerU [28], an OCR-based document parsing tool, to convert them into machinereadable markdown while preserving textual content and document structure for downstream processing. The converted content, together with carefully designed prompts, is then provided to an LLM for structured information extraction. We use GPT-5.2 Thinking [22] as the core extraction model because its reasoning ability and long-context support make it well suited for source-grounded extraction from complex academic papers. During development, it provided a practical balance between extraction quality and cost for large-scale dataset construction. This extraction stage produces two levels of outputs. The first is a set of core relationship records, which capture the fundamental QoS-QoE relationships extracted from source evidence such as equations, tables, and figures. The second is a set of metadataenriched records, which augment the core relationship records with contextual metadata such as protocol, network type, device type, and scenario. By separating core relationship extraction from contextual metadata enrichment, the pipeline keeps the extracted relationships grounded in source evidence while still providing richer context for downstream analysis and reuse.
2.1
2.3
2
Dataset Construction
Paper Curation
We begin by constructing a curated corpus of research papers on QoS-QoE relationships in video streaming. To identify relevant and high-quality studies, we combine human screening with OpenAI deep research-assisted literature search [21]. We focus on papers published between 2017 and 2025, since advances in streaming systems, codecs, devices, and network configurations can change
Iterative Data Evaluation
To improve data quality, we further introduce an iterative data evaluation stage. As shown in Figure 1, each metadata-enriched record is reviewed by multiple LLM-based data evaluators. In the current dataset construction setup, we instantiate this stage with three data reviewers: Gemini-2.5-flash-lite [9], Claude-haiku-4-520251001 [2], and Grok-4.20-0309-reasoning [31]. Each reviewer 2
Table 1: Field definitions of QoS-QoE Translation.
... { "time_s": 10.83, "qos": [ {"metric": "initial_loading_delay_level", "value": "1"}, ...], "qoe": [ {"metric": "mos", "value": "3.3148"} ] } ... {
metadata "id": "...", "year": 2019, relationship "venue": "IEEE Access", "domain": "video_streaming", "protocol": [ "DASH" ], "network_type": [ "wired" ], "device_type": [ "desktop" ], "video_type": [ "2d_vod" ], "user_preference": "low_rebuffer", "scenario": "A client streams ...", "history_log": [ ... ], "data_type": "equation", "qos_parameter": [ "initial_loading_delay_level", ...], "qos_parameter_definition": [ ... ], "qoe_parameter": [ "mos" ], "qoe_parameter_definition": [ ... ], "relationship": "MOS = 4.23 - 0.0672 L_{ti} - 0.742 L_{fr} - 0.106 L_{tr}", "description": "The equation expresses MOS as a linear function ...", "source": "...",
}
Field
Definition
id year venue domain protocol network_type device_type video_type user_preference scenario history_log data_type qos_parameter qos_parameter_definition qoe_parameter qoe_parameter_definition relationship description source
Unique identifier for each dataset record. Publication year of the source paper. Publication venue of the source paper. Application domain of the record. Streaming or transport protocol used in the study. Access network environment. Client device used for content consumption. Video content category or media format. User preference emphasized in the study. Summary of the experimental or evaluation setting. Temporally ordered QoS-QoE observations. Evidence type, such as equation, table, or figure. QoS variables in the relationship. Definitions of the QoS variables. QoE variables in the relationship. Definitions of the QoE variables. Extracted dependency between QoS and QoE. Natural-language explanation of the relationship. Source-grounded evidence trace.
Figure 2: An example JSON record from the QoS-QoE Translation, showing enriched metadata and core relationship. source information and contextual attributes, and relationship, which stores the extracted QoS-QoE relationship. This design preserves contextual information and source-grounded relationships in a unified machine-readable format for downstream analysis, benchmarking, and modeling.
provides a rating, a confidence score, and written feedback, denoted in the figure as a tuple of the form (𝑟𝑖 , 𝑐𝑖 , 𝑚𝑖 ). The rating score 𝑟𝑖 is assigned on a 0–10 scale, where 0, 2, 4, 6, 8, and 10 denote strong reject, reject, weak reject, weak accept, accept, and strong accept, respectively. The confidence score 𝑐𝑖 is assigned on a 1–5 scale, where higher values indicate stronger reviewer confidence in the judgment. Reviewer comments are also required to describe the identified issues and suggest possible solutions. These reviewer outputs are aggregated into a decision using the ¯ and the collection of average rating 𝑟¯, the average confidence 𝑐, reviewer feedback messages. Based on our empirical inspection of reviewer outputs during dataset construction, we found the following thresholds to provide a reasonable balance between retaining high-quality records and filtering out unreliable extractions. Accept, Decision = Reject, Conditional Accept,
if 𝑟¯ ≥ 8 and 𝑐¯ ≥ 3, if 𝑟¯ ≤ 4, otherwise.
3.1
(1)
Records that satisfy the accept condition are retained as valid JSON entries, while records that satisfy the reject or conditional accept conditions are sent to a re-evaluation stage. In this stage, human guidance and an LLM are jointly used to revise the data before returning it to the evaluation loop. This iterative mechanism reduces unsupported, ambiguous, or low-quality extraction results and improves the consistency of the final dataset.
3
Metadata Diversity and Coverage
Figures 3a–3d show that the dataset is concentrated in mainstream video streaming settings while still preserving cross-setting diversity. DASH accounts for 59.7% of the protocol distribution, followed by HTTP at 20.2%, while WebRTC, QUIC, RTP, and HLS together contribute a non-trivial share of interactive and transport-level settings. For network type, cellular 4G (33.2%), wired (23.0%), and Internet-based environments (19.3%) dominate, indicating that the dataset covers both mobile and fixed-network evaluations. Device type is largely split between desktop (45.2%) and mobile (41.0%), suggesting that the dataset mainly reflects common end-user viewing platforms. For video type, 2D video-on-demand is the largest category at 61.3%, followed by 2D live streaming at 27.8%, while short video and immersive formats such as 360 and VR video remain represented. Overall, these distributions show that the dataset is anchored in dominant real-world streaming scenarios, but still retains enough diversity to support cross-context analysis.
3.2
Temporal and Venue Coverage
Figure 3e shows that the dataset is concentrated in recent years, with 2024 contributing the largest share at 18.2%, while papers from 2017–2025 remain represented. Figure 3f further shows that the dataset is collected from a broad set of publication venues. ACM MM contributes the largest share at 26.0%, followed by arXiv (16.7%), IEEE Access (15.6%), IEEE INFOCOM (9.5%), ACM TOMM (8.2%), NSDI (7.6%), and so on. Together, these distributions show that QoSQoE Translation is grounded in both multimedia and networking communities, while maintaining strong coverage of recent research and reducing noise from older system settings.
Dataset Overview and Analysis
QoS-QoE Translation contains 1026 source-grounded QoS-QoE relationship records extracted from 505 curated papers after extraction and iterative data evaluation. Figure 2 shows an example JSON record, Table 1 summarizes the field definitions, and Figure 3 summarizes the dataset composition in terms of metadata, temporal and venue coverage, and QoS/QoE parameter distributions. Each record contains two main components: metadata, which captures 3