The Optimizer Is the Agent: Reasoning-Driven Search across Prompts, Programs, and ML Workflows
ReASearch tests whether an agent can absorb the search loop usually handled by separate optimizers.
The paper presents a tool-using agent that chooses evaluations, diagnoses failures, edits candidates, and decides when to verify or restart. The same scaffold is used across prompt, program, and ML workflow optimization. Across 14 tasks, the authors report 2% to 40% gains over strong domain-specific baselines, with some results exceeding prior human best-known solutions. HF Daily Papers' note
The paper presents a tool-using agent that chooses evaluations, diagnoses failures, edits candidates, and decides when to verify or restart. The same scaffold is used across prompt, program, and ML workflow optimization. Across 14 tasks, the authors report 2% to 40% gains over strong domain-specific baselines, with some results exceeding prior human best-known solutions. HF Daily Papers' note
score 5