Researchers have proposed a new framework called Search-G1 to improve the performance of search-augmented language agents. These agents are designed to retrieve external information only when necessary and ground their answers in retrieved evidence. However, existing methods either provide sparse rewards or require costly annotations. Search-G1 uses two intervention-calibrated readouts to measure the operational grounding of an agent's answers: a prompt-state readout that predicts closed-book sufficiency and an answer-commit readout that estimates evidence reliance from answer-stage sensitivity to evidence deletion. The framework requires periodic refitting of both readouts on trajectories from the latest checkpoint, allowing the reward to co-evolve with the policy. Experiments show that Search-G1 improves the grounding-search-cost trade-off, producing shorter response-side trajectories at competitive task accuracy. --- Why it matters: This matters because it addresses a common challenge in AI research: balancing the need for accurate answers with the cost of retrieving external information. By improving this balance, Search-G1 can lead to more efficient and effective search-based question-answering systems. Source: https://arxiv.org/abs/2608.07531