Self-Evolving Deep Research via Joint Generation and Evaluation

Researchers introduce SCORE, a self-evolving co-evolutionary framework that jointly trains evaluation and generation models for deep research report generation. The approach addresses limitations in LLM-based research agents by enabling evaluators to dynamically adapt standards as solver performance improves, demonstrating consistent quality improvements over static evaluation methods.

Deep research capabilities represent a frontier challenge for Large Language Models, requiring agents to generate comprehensive reports without traditional ground-truth validation. The research community has struggled with reward design in this domain, as evaluating research quality involves subjective dimensions that resist standardization. Previous attempts using LLM-as-a-judge and query-dependent rubrics achieved only partial success because evaluators remained static, unable to raise standards proportionally as generators improved.

The SCORE framework fundamentally restructures this problem by treating evaluation and generation as interdependent rather than isolated tasks. By coupling these components in a shared-parameter learning process, the system enables mutual improvement where better generation pushes evaluator standards higher, which in turn incentivizes deeper research capabilities. The meta-harness component introduces algorithmic discipline by dynamically controlling evaluation conditions based on solver performance, preventing both trivial optimization and evaluator drift.

This advancement carries broader implications for AI development beyond research applications. Open-ended task generation—where ground-truth evaluation is inherently ambiguous—appears across numerous domains including creative writing, strategic planning, and scientific discovery. The co-evolutionary approach provides a template for addressing similar challenges where traditional supervised learning fails. For developers building LLM-based agents, this research validates the principle that evaluation and generation should evolve together rather than being treated as sequential pipeline stages.

Future research directions include scaling SCORE to larger models and testing on increasingly complex research domains. The framework's effectiveness at preventing optimization saturation suggests potential applications in iterative AI training processes where human oversight becomes increasingly sparse.

• →Co-evolutionary training framework enables evaluators to dynamically adapt standards as generator performance improves, solving the saturation problem in deep research tasks.
• →Shared-parameter learning between evaluation and generation modules creates mutual optimization pressure unavailable in isolated component architectures.
• →Meta-harness mechanism controls evaluation environment to prevent both trivial optimization and evaluator drift, improving training stability.
• →Approach generalizes beyond research reports to any open-ended task domain where ground-truth evaluation is subjective or ambiguous.
• →Experimental results demonstrate consistent quality improvements over static evaluation methods on deep research benchmarks.