The quest for optimizing adaptive bitrate (ABR) algorithms has been propelled to new heights with the introduction of LLM-ABR, which stands for Large Language Model – Adaptive Bitrate. This novel system harnesses the generative potential of ( ) to automatically design ABR algorithms that are tailored to diverse network conditions. Developed by researchers at
Microsoft Research, University of
Texas at Austin, and Peking University, LLM-ABR marks a significant leap forward in the ABR algorithm landscape by consistently outperforming standard ABR algorithms across varied network environments like broadband, satellite, and 4G/5G networks. The field of adaptive bitrate streaming has seen significant advancement over recent years, with research and development focused on improving video streaming quality and efficiency. Historically, constructing ABR algorithms was a laborious process involving heuristic reasoning, techniques, and extensive empirical testing. This complexity created a bottleneck in the development of ABR solutions suitable for the ever-evolving network conditions. Until now, the majority of the focus has been on enhancing existing ABR techniques rather than creating new ones from scratch, which is what sets LLM-ABR apart as it leverages the of LLMs to generate innovative designs. LLM-ABR’s approach involves using input prompts and existing algorithm source code to produce a variety of new ABR designs. Unlike traditional methods that rely on manually crafted algorithms, this LLM-generated approach includes a normalization check which ensures the output is scaled correctly for neural network inputs. The process generates thousands of potential designs, which are then whittled down after a compilation check to filter out non-viable options. The designs that pass this initial screening are subjected to further evaluation to determine their performance in improving video Quality of Experience (QoE). The evaluation of LLM-ABR’s generated algorithms demonstrates a wide range of performance improvements, from 1.4% to 50%, depending on the network scenario. Notably, the greatest improvements were observed with satellite internet services like Starlink, where default algorithms often overfit to specific conditions. Even for mobile networks such as 4G and 5G, the newly crafted architectures delivered modest but consistent enhancements over the baseline across all tested periods. These findings underscore the potential of LLM-ABR in revolutionizing ABR algorithm development. Corroborating the research findings, a scientific paper published in the Journal of Networking and Computer Applications, titled “Adaptive Bitrate Streaming for Next-Generation Networks,” highlights the industry’s growing need for adaptive and scalable ABR solutions. While the scientific paper focuses on the current challenges in ABR technology, LLM-ABR emerges as a response, offering a glimpse into the future of ABR algorithms, where AI-driven designs could become the norm. In sum, LLM-ABR’s introduction is a game-changer in the development of adaptive bitrate algorithms. It represents a new way of thinking where AI-generated solutions can be tailored to specific network environments, promising a future where streaming quality is significantly improved. While the cost and complexity of using LLMs can be a barrier, the benefits of LLM-ABR are clear. The system’s flexibility and adaptability could make it an invaluable tool in the ongoing effort to enhance the streaming experience amidst the growing demand for high-quality video content.