Frontier large language models are increasingly used to automate scientific research through iterative search, but idea-driven search introduces challenges beyond finding solutions. The paper distinguishes idea-driven from solution-driven search and identifies three problems: organizing evolving ideas, selecting promising directions, and keeping ideas aligned with their implementations. It presents the Agentic Idea Manager (AIM), a fully autonomous framework that uses an Agentic Surrogate and Agentic Acquisition mechanism, inspired by Bayesian optimization, to organize and prioritize research directions. A Solution Auditor checks idea-solution integrity, while a Resource Planner distributes the remaining experimental budget across parallel search branches. Across 10 AutoLab benchmark tasks, AIM exceeds the strongest baseline by 1.6 percentage points on System Optimization and 4.9 points on long-horizon Model Development & CUDA tasks. It also reaches the best baseline performance up to 3.1 times faster in wall-clock time. Theoretical analysis indicates that explicit idea-level allocation makes semantic coverage controllable and becomes increasingly useful when strong research directions are sparse among many plausible alternatives.
AI News
The latest AI releases, research, products, and industry updates.
Loading...