AWS documents multi-turn RL reward design on Nova Forge — and the failure mode where a reward component is silently dead

AWS Machine Learning Blog·medium signalXBlueskyLinkedInCopy link

AWS published a worked example on August 14, 2026 training Amazon Nova Lite 2.0 on 500 programming tasks using GRPO with LoRA on SageMaker HyperPod, with a four-component reward: correctness at weight 1.0, asking-before-coding at 0.6, a guessing penalty at 0.4, and loop detection at 0.2. The most useful line is diagnostic rather than architectural — "a component with near-zero within-group variance contributes nothing to learning" — meaning a reward term can look healthy in aggregate metrics while teaching the model nothing. The post also covers sandboxed execution of model-generated code with resource limits and unforgeable verification, which is the part most homegrown RL harnesses get wrong.