Complete Cyclic Subtask Graphs for Tool-Using LLM Agents
A study published on arXiv (2604.22820) presents complete cyclic subtask graphs, which represent a highly adaptable multi-agent framework for long-horizon tasks involving tools. This architecture features fully interconnected executable subtask nodes, while a centralized agent for state analysis and routing determines transitions based on natural language criteria, allowing for unlimited revisiting of subtasks for both recovery and exploration purposes. The research assesses task-specific (Spec-Cyc) and benchmark-generic (Gen-Cyc) graphs across TextCraft, ALFWorld, and Finance-Agent benchmarks. The analysis includes variations in planner/executor/router capabilities, tool exposure (generalist versus specialized), n-shot successful trajectory summaries, and random subtask perturbations induced by faults. Findings indicate three distinct operational regimes, with ALFWorld revealing a significant bottleneck.
Key facts
- arXiv paper 2604.22820 introduces complete cyclic subtask graphs.
- Architecture allows unrestricted revisitation of subtasks.
- Unified state-analysis-and-routing agent uses natural-language criteria.
- Evaluated on TextCraft, ALFWorld, and Finance-Agent benchmarks.
- Ablations include planner/executor/router strength and tool exposure.
- Three distinct performance regimes identified.
- ALFWorld highlights a bottleneck.
- Research focuses on flexibility vs cost trade-offs.
Entities
Institutions
- arXiv