A research report can be wrong in many ways that an exact-answer checker cannot see. The useful idea here is to improve the evaluator alongside the agent: new search evidence and competing answers expose factual gaps that a fixed rubric may miss.
Key Findings
- Search-grounded, evolving criteria give long-form training a reward signal beyond short-answer correctness; the paper tests this across four research benchmarks.
- Positive criteria capture newly useful evidence, while negative criteria penalize recurring failures. Keeping discriminative criteria focuses feedback as the policy improves.
- The released training and evaluation stack makes this approach inspectable, rather than leaving only a hosted demo. The official repository includes agent infrastructure, supervised fine-tuning, and reinforcement-learning code.
Who Should Read It
Great fit if you study rewards for research agents or need an open training stack to adapt. Look elsewhere if you need a minimal retrieval pipeline: this approach adds rubric generation, judge-model scoring, and tool orchestration. Treat benchmark results as evidence under the paper's evaluation setup; research citations still merit checking.