Reproduce “Budget-Aware Tool Use Enables Effective Agent Scaling” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Tengxiao Liu, Zifeng Wang, Jin Miao, I-Hung Hsu, Jun Yan, Jiefeng Chen, Rujun Han, Fangyuan Xu and 7 more

The authors introduce Budget Tracker and BATS to help agents use search and browsing tools. Budget Tracker shows used and remaining budgets for each tool; BATS adds planning and verification to guide whether to continue, pivot, or accept an answer. On BrowseComp, Gemini-2.5-Pro with BATS scores 24.6% accuracy versus 12.6% for budget-constrained ReAct, under a limit of 100 uses per tool.

Key insights

  • Explicit budget signals enable large language model (LLM) agents to internalize resource constraints and dynamically adapt their strategies without requiring additional training.
  • Integrating budget awareness into iterative planning and self-verification processes is crucial for strategic resource allocation, allowing agents to effectively explore, verify, and pivot based on remaining capacity.
  • Tool-call budgets represent a critical scaling dimension for tool-augmented agents, and managing them effectively can yield more favorable cost-performance trade-offs than simply increasing internal 'thinking' or tool availability.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.