RoBoN: Routed Online Best-of-n for Test-Time Scaling with Multiple LLMs

TL;DR

RoBoN is a sequential, routing-based method that leverages multiple LLMs at inference time to improve response quality over traditional single-model best-of-$n$ approaches, without additional training.

Contribution

RoBoN introduces a novel online routing mechanism for multiple LLMs, enhancing test-time scaling and response accuracy without extra training or compute overhead.

Findings

RoBoN outperforms standard best-of-$n$ in accuracy by up to 3.4%.

RoBoN improves performance across various reasoning benchmarks.

Diversity among models can be exploited at inference to enhance results.

Abstract

Best-of-$n$ is a widely used test-time scaling approach for LLM inference. Yet despite evidence that LLMs exhibit complementary strengths across tasks, traditionally best-of-$n$ relies on a single model to generate responses. We propose RoBoN (Routed Online Best-of-$n$), a sequential multi-LLM alternative to the prevailing single-model best-of-$n$. Given a suite of models $\{m_i\}_{i=1}^M$, RoBoN sequentially routes generations one-by-one across models, based on scores computed using a reward model and an agreement signal on the predicted responses. This online routing requires no additional training, keeps compute parity, and works with any plug-in reward model. Across reasoning benchmarks (MATH500, OlympiadBench, MinervaMath, GSM8K, MMLU), RoBoN consistently outperforms standard best-of-$n$ applied to each individual model for larger $n$, with gains of up to 3.4\% in absolute accuracy, and also improves over a uniform multi-model portfolio baseline. Our results indicate that diversity across models can be exploited at inference to improve best-of-$n$ performance over any constituent model alone, providing a simple, training-free path to test-time scaling with multiple LLMs.