Search-G1: Grounded Search Agents via Representation-Based Intrinsic Rewards

arXiv:2608.07531v3 proposes Search-G1, a representation-based intrinsic reward framework. This framework uses prompt state readings to predict the adequacy of closures to define retrieval necessity and uses answer submission readings to evaluate sensitivity to evidence deletion. After calibration, Search-G1 can provide additional credit for correct search trajectories and penalize repeated searches without the need for process annotations or LLMs as judges in reinforcement learning optimization. Experiments show that this method improves the balance between grounding and search cost, generating shorter response-side trajectories with competitive task accuracy across multiple search-based question-answering benchmarks and at both model sizes. The related code is available on the GitHub repository.