Search-G1 teaches AI agents when to retrieve—and when not to

Search-augmented language agents often retrieve external information too liberally, wasting compute and risking answers ungrounded in evidence. Existing reward signals are either sparse outcome checks or expensive process annotations and LLM-judge evaluations. Search-G1, proposed by Cheng Ruoxi and nine co-authors in arXiv:2608.07531, introduces a representation-based intrinsic reward framework with two intervention-calibrated readouts. A prompt-state readout predicts whether the model can answer from closed-book knowledge alone; its complement defines retrieval necessity. An answer-commit readout measures evidence reliance by testing how sensitive the final answer is to evidence deletion. Together, these signals reward correct searched trajectories only when retrieval is essential and the answer genuinely depends on evidence, while penalizing redundant searches.

Because reinforcement learning continuously shifts policy representations, Search-G1 periodically refits both readouts using trajectories from the latest checkpoint, letting the reward co-evolve with the policy. This removes the need for manual annotations or LLM-as-judge inference during optimization while still providing graded, dense feedback. Experiments across multiple search-based question-answering benchmarks and two model scales show Search-G1 improves the grounding-search-cost trade-off—cutting response-side trajectory length notably while maintaining competitive task accuracy. The authors released the code publicly, making it straightforward for teams to apply this method to their own RAG or agent pipelines. The approach offers a practical path toward leaner, more trustworthy retrieval-augmented agents.