Replace hand-written RL reward functions with RULER: an LLM judge that ranks trajectories relative to each other

OpenPipe ART·medium signal

RULER (Relative Universal LLM-Elicited Rewards), built into OpenPipe's ART framework, scores a batch of agent trajectories 0–1 relative to one another by reading the agent's own system prompt to infer the goal — no labeled data, expert feedback, or custom reward code. It matches or beats handcrafted rewards on 3 of 4 benchmarks and cuts reward-engineering time 2–3x. For solo builders this collapses the hardest part of RL fine-tuning into a single function call, and it now also runs on W&B serverless training.