ARC: Dynamic Agent Configuration for LLM Systems
A new hierarchical policy called ARC (Agentic Resource & Configuration learner) has been developed by researchers to enhance LLM-based systems by selecting agent configurations tailored to specific queries. This innovation tackles the limitations of static templates and manually adjusted heuristics that utilize the same setup regardless of the complexity of the query. ARC treats agent configuration as a semi-Markov decision process (SMDP), with each configuration functioning as a temporally extended option. Compared to budget-matched tool-augmented LLMs, ARC shows significant improvements across reasoning, tool-use, and agentic benchmarks, boosting average reasoning accuracy by 31.3%, increasing tool-use accuracy by 13.95%, and achieving a twofold success rate in τ-Bench (Airline) Pass^1. The method enhances workflows, tools, token budgets, and prompts from a vast combinatorial design space.
Key facts
- ARC is a lightweight hierarchical policy for dynamic agent configuration.
- Agent configuration is formulated as a semi-Markov decision process (SMDP).
- ARC improves reasoning accuracy by 31.3% on average.
- Tool-use accuracy increases by 13.95% with ARC.
- τ-Bench (Airline) Pass^1 success is doubled.
- Current methods use fixed templates or hand-tuned heuristics.
- Configuration involves workflows, tools, token budgets, and prompts.
- The approach reduces brittle behavior and wasted compute.
Entities
—