Agent-UCT: Cost-Aware Tree Search for Agentic Workflow Optimization
A novel algorithm named Agent-UCT (Agent-based Cost-Aware Upper Confidence Bounds Applied to Trees) has been developed to enhance agentic workflows, particularly in retrieval-augmented generation (RAG) pipelines. This algorithm builds upon UCT by incorporating a reuse-aware regularization term, which is based on a bipartite prefix reuse graph. This addition encourages the selection of branches that utilize previously established configuration prefixes, thereby minimizing redundant executions while still promoting exploration. Additionally, the RAGSpace framework integrates various RAG components from LongRAG, LightRAG, and Self-RAG into a comprehensive five-dimensional configuration space. The research is available on arXiv under ID 2607.24162.
Key facts
- Agent-UCT is a tree search algorithm for optimizing agentic workflows.
- It extends UCT with a reuse-aware regularization term from a bipartite prefix reuse graph.
- The algorithm reduces redundant execution by biasing toward previously materialized prefixes.
- RAGSpace unifies components from LongRAG, LightRAG, and Self-RAG.
- The paper is available on arXiv with ID 2607.24162.
- The approach targets combinatorial optimization under tight evaluation budgets.
- Existing methods like heuristic search and black-box optimization do not exploit compositional structure.
- The algorithm is cost-aware, considering evaluation budgets.
Entities
Institutions
- arXiv