Reproduce “Plan and Budget: Effective and Efficient Test-Time Scaling on Reasoning Large Language Models” 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
Junhong Lin, Xinyue Zeng, Jie Zhu, Song Wang, Julian Shun, Jun Wu, Dawei Zhou
A framework called PLAN-AND-BUDGET enhances the efficiency and effectiveness of Large Language Model (LLM) reasoning by adaptively allocating computational resources at test time. This approach, built upon a theoretical Budget Allocation Model (BAM), improves an efficiency-aware effectiveness score (E³) by up to 193.8% while reducing token usage by up to 39% across various reasoning tasks.
Key insights
- LLM reasoning miscalibration can be effectively reinterpreted as a resource allocation problem under uncertainty, allowing for principled adaptive computation.
- Decomposing complex queries into structured sub-questions and adaptively allocating computational budget to these sub-questions based on their estimated complexity and uncertainty significantly improves reasoning efficiency.
- Front-loaded decay strategies for budget allocation, such as polynomial decay and cosine annealing, are particularly effective for complex reasoning tasks by prioritizing resources for earlier, potentially more uncertain reasoning steps.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.