Adaptive ML, developer of the world's first Reinforcement Learning Operations (RLOps) platform, today published an open-source RL glossary at dev.adaptive-ml.com — the first structured, full-stack reference for Reinforcement Learning applied to large language models in production.

The glossary covers 37 entries across 8 categories: training, data, rewards, optimization, agents, inference, evaluation, and RLOps, organized as a progressive learning path from supervised fine-tuning through reward modeling, chain-of-thought prompting, LLM-as-judge evaluation, and full RLOps pipelines. Unlike existing references, which are either academic in scope, course-embedded, or limited to narrow segments of the stack, the Adaptive ML glossary covers the complete lifecycle: the concepts an organization needs to move a specialized LLM from foundation model to production asset. Entries include academic citations and explicit notes on limitations — including failure modes like reward hacking, benchmark gaming, and Goodhart's Law in practice. It is a living resource; additional entries and interactive visualizers are in development.

The glossary was written by the team running RL at enterprise scale. Adaptive ML's platform powers active deployments at AT&T, where models process 900,000 call summaries daily, SK Telecom, where a 4-billion-parameter specialized model outperforms closed frontier alternatives in content moderation, Manulife, Deloitte, and others — with trillions of tokens in production. The release is a deliberate act of category definition: as enterprise interest in reinforcement learning spikes — driven by reasoning model breakthroughs and growing pressure to reduce dependency on closed APIs — Adaptive ML is establishing the reference vocabulary for the field it is building. 

The company designed the glossary not as a marketing asset but as an open repository aimed at nurturing a community with deep, shared knowledge of RL. It gives ML engineers, infrastructure leads, and executives a common language to evaluate, build, and deploy specialized models with confidence. It also serves a practical function inside Adaptive ML's own client engagements: a tool to onboard broader organizational teams — beyond core ML — onto a shared vocabulary before and during production deployments.