ReForge: LLM-Driven Adaptive ABR Algorithm Design
A new framework called ReForge, detailed in a preprint on arXiv (2608.15138), proposes a method for continually adapting adaptive bitrate (ABR) algorithms to changing network conditions using large language models (LLMs). The framework addresses the limitation of traditional ABR algorithms, which are designed for specific network scenarios and fail when conditions change. ReForge uses an LLM to propose small edits to a fuzzy rule set that routes decisions to a pool of pre-trained policies, with a replay mechanism ensuring that changes do not harm previously served scenarios. This approach aims to redesign ABR algorithms in minutes as new scenarios arise, rather than taking months of manual engineering. The work is authored by researchers and submitted to arXiv, indicating a focus on computer science and networking. The framework's key innovation is the integration of LLMs into the continual learning loop, enabling rapid adaptation while maintaining performance across all encountered scenarios.
Key facts
- ReForge is a continual heuristic learning framework for ABR algorithms.
- It uses a large language model (LLM) to propose edits to a fuzzy rule set.
- The framework adapts to changing network scenarios in minutes.
- Replay over all served networks ensures edits are safe.
- The design fits only the world visible at its birth, motivating ReForge.
- LLMs can design ABR algorithms in hours, matching or beating hand-built designs.
- The paper is available on arXiv with ID 2608.15138.
- The framework routes decisions to a frozen pool of pre-trained policies.
Entities
Institutions
- arXiv