<ul><li>Recent advancements in large language models have improved reasoning abilities through search and backtracking techniques.</li><li>Sequential search, enabled by backtracking, allows linearized exploration via long chain-of-thought generation.</li><li>Parallel sampling with best-of-n selection is an alternative approach to scaling test-time compute.</li><li>Comparative analysis on reasoning tasks shows that while sequential search outperforms parallel sampling on Sudoku, it underperforms on CountDown, indicating the limitations of backtracking.</li></ul>

To Backtrack or Not to Backtrack: When Sequential Search Limits Model Reasoning

Discover more