TL;DR
This guide develops The Economics of RL Post-Training: What a Run Actually Costs from its central engineering problem to practical design choices and limitations. It synthesizes the roadmap topics without claiming new experimental results. Reported findings belong to the cited sources.
Target searches: “cost of RL post training”, “how much does RLHF cost”
Why This Topic Matters
Modern LLM and agent training turns familiar reinforcement-learning ideas into systems problems: actions may be tool calls, rewards may come from fallible graders, and trajectories may span many partially observed steps. The classical foundation is distributed RL; this article focuses on what changes in the newer setting.
The Core Ideas
1. The cost model: GPU-hours × price, decomposed into generation cost and training cost
2. Why generation usually dominates — you generate G responses per prompt, each thousands of tokens, every step
3. Worked example: a GRPO run on a 7B model, real numbers, with public GPU pricing
4. The levers: group size G, max response length, dynamic sampling, rollout batch size, KV cache reuse
5. Cost per point of benchmark improvement — the metric nobody publishes
6. Buy vs build: RL-as-a-service vs your own cluster
7. Where the money is wasted (see C3)
Practical Checklist
- Define the state, action, termination condition, and reward before choosing an optimizer.
- Log raw trajectories and every reward component, not only the aggregate score.
- Separate training measurements from held-out evaluation and adversarial tests.
- Treat generated code, tool calls, and environment actions as untrusted.
- Record model, framework, benchmark, and grader versions so comparisons remain reproducible.
Limitations
This area changes quickly. Framework support and benchmark leaderboards are time-sensitive, while many recent methods are preprints rather than settled practice. The article describes mechanisms and reported evidence; it does not present an original training run.
Key Learnings
- The cost model:
GPU-hours × price, decomposed into generation cost and training cost - Why generation usually dominates — you generate G responses per prompt, each thousands of tokens, every step
- Worked example: a GRPO run on a 7B model, real numbers, with public GPU pricing