Advancing Intelligence
Through Structured
Research

Efficient Online RFT with Plug-and-Play LLM Judges

Reward-model training is the cost bottleneck in modern RLHF pipelines. This paper presents a frozen, instruction-tuned 7B LLM augmented with a one-line JSON rubric and rank-16 LoRA adapter, achieving 96.2% accuracy on RewardBench.