<ul><li>Large Language Models (LLMs) have shown success in reasoning tasks but are computationally inefficient due to overthinking and underthinking.</li><li>Researchers have introduced the Bayesian Budget Allocation Model (BBAM) to address inefficiencies in reasoning by modeling it as a sequence of sub-questions.</li><li>A test-time framework called Plan-and-Budget has been proposed to decompose complex queries into sub-questions and allocate token budgets efficiently based on estimated complexity.</li><li>Plan-and-Budget has been effective in improving reasoning efficiency, achieving accuracy gains, token reduction, and overall improvement in computation efficiency.</li></ul>

Plan and Budget: Effective and Efficient Test-Time Scaling on Large Language Model Reasoning

Discover more