Reproduce “AgentTTS: Large Language Model Agent for Test-time Compute-optimal Scaling Strategy in Complex Tasks” 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

Fali Wang, Hui Liu, Zhenwei Dai, Jingying Zeng, Zhiwei Zhang, Zongyu Wu, Chen Luo, Zhen Li and 3 more

Researchers from The Pennsylvania State University and Amazon developed AgentTTS, an LLM-agent-based framework designed for test-time compute-optimal scaling in complex multi-stage tasks. The system intelligently allocates specific LLMs and computational budgets to individual subtasks, demonstrating superior search efficiency, faster convergence, and enhanced interpretability compared to traditional and other LLM-based optimization methods.

Key insights

  • Different subtasks within a multi-stage workflow exhibit distinct preferences for larger versus smaller models, and each subtask/model configuration has a task-specific optimal compute budget for test-time scaling.
  • Increasing test-time compute (e.g., more samples) initially improves performance but yields diminishing returns or can even degrade performance beyond a certain optimal point.
  • The compute budget allocated to an earlier subtask significantly impacts the scaling dynamics and computational requirements of downstream subtasks, highlighting their strong interdependence.

Summary by alphaXiv.

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