Best-of-$\infty$ -- Asymptotic Performance of Test-Time Compute

Researchers propose 'best-of-∞' approach for large language models that uses majority voting with infinite samples, achieving superior performance but requiring infinite computation. They develop an adaptive generation scheme that dynamically selects the optimal number of samples based on answer agreement and extend the framework to weighted ensembles of multiple LLMs.

• →Best-of-∞ approach with majority voting achieves impressive LLM performance but requires infinite test-time computation budget.
• →Adaptive generation scheme efficiently allocates inference-time computation by selecting sample size based on answer agreement.
• →Weighted ensembles of multiple LLMs can outperform any individual model according to the research.
• →Optimal ensemble weighting is formulated as a mixed-integer linear program for efficient computation.
• →Extensive experiments demonstrate the effectiveness of the proposed adaptive approach.