Process-Reward Tactic Evolution for Long-Horizon Bioinformatics Workflows

The paper introduces Process-Reward Tactic Evolution, a training framework designed for LLM agents to effectively manage long-horizon bioinformatics workflows using Galaxy. This framework utilizes a curriculum-based approach to train agents on workflow execution, incorporating a tactic library derived from verified workflow rollouts, which enhances the agent's ability to construct workflows, monitor execution, and ensure biological correctness. The evaluation demonstrates that this process-supervised tactic accumulation significantly improves workflow completion rates and execution efficiency compared to traditional no-memory and reflection-style approaches, highlighting its relevance for practitioners developing complex bioinformatics applications.