Reproduce “The Shadow Price of Reasoning: Economic Perspective on Optimal Budget Allocation for LLMs” 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
Xu Wan, Speed Zhu, Jianwei Cai, Guang Chen, XiMing Huang, Wiggin Zhou, Mingyang Sun
Researchers from Zhejiang University and Tencent HY Team developed CLEAR, an economic framework for optimally allocating Large Language Model (LLM) inference tokens across diverse queries to maximize aggregate accuracy. The method leverages a global "shadow price" of computation and achieves substantial accuracy improvements over uniform allocation, particularly under strict computational budgets.
Key insights
- LLM reasoning utility for a given query empirically follows an S-shaped curve, revealing non-linear returns on token investment and the existence of task-specific "emergence thresholds" for meaningful progress.
- An economic framework, utilizing a global "shadow price," provides a theoretical basis for optimal resource allocation by ensuring marginal utility parity across all actively processed tasks.
- "Rational abandonment" of queries that cannot generate sufficient net utility at the prevailing shadow price is critical for maximizing global accuracy, allowing resources to be reallocated to more promising tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.