ReForge uses LLM revisions to keep adaptive bitrate algorithms improving after deployment
Researchers propose a system where a large language model rewrites routing rules for deployed streaming algorithms as new networks arrive, with a simulation check before each change is accepted.
In plain words
Adaptive bitrate (ABR) algorithms decide video quality while you stream, and they are usually fixed once shipped, so they struggle on network types their designers never saw. ReForge lets a large language model — an AI system that writes text and code — rewrite the decision rules after deployment. Each rewrite is only accepted if a simulation shows no existing network gets worse.
What was announced
The arXiv paper's authors note that designing an ABR algorithm for a network scenario can take an engineer months, and that LLMs now do this work in hours, but the result is still fixed at deployment and falls behind on later networks. ReForge keeps a pool of existing ABR policies unchanged and has an LLM revise only the rules that decide which policy picks the next bitrate. When a new network arrives, the model reads how the current rules and every policy performed on networks served so far, then rewrites the rules.
A rewrite is deployed only if a replay in simulation shows it improves the served networks overall while making none worse; a fast GPU simulator that replays every episode exactly makes this check cheap enough to run each time. On nine real-world network datasets with networks arriving one at a time, the authors report mean QoE — a quality-of-experience measure — rising from 1.43 to 1.79, above the best single policy at 1.66 and near the 1.86 of picking the best policy per dataset, with no drop on networks served first. They add that the gain holds in a real video player, and that all code, data and records will be available.
What it does not show
- The evidence is the paper's abstract only; full method details, baselines and peer review are not covered, and the reported QoE results are the authors' own and not independently confirmed.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.