On the Limits of Test-Time Compute: Sequential Reward Filtering for Better Inference

Test-time compute (TTC) has become an increasingly prominent paradigm for enhancing large language models (LLMs). Despite the empirical success of methods such as best-of-n (BoN) sampling and sequential revision, their fundamental limits remain unclear. We address this gap by analyzing a mixture-of-reference policy model and proving that standard BoN is inherently suboptimal. To move closer to the optimal frontier, we study reward-filtered sequential inference, a simple procedure that selectively incorporates only high-reward generations into the context. This mechanism concentrates computation on superior policy candidates and suppresses inferior ones. On the theoretical side, we show that reward-filtered sequential inference yields strictly stronger guarantees than standard TTC paradigms. On the empirical side, we evaluate such an inference strategy across diverse benchmarks and observe consistent improvements over widely used approaches, demonstrating the practical effectiveness of our framework. How to effectively utilize large language models (LLMs) for solving new tasks has become a central research question. Among the many approaches, Test-Time Compute (TTC) has recently attracted significant attention. The key idea of TTC is to allocate additional computation during inference to improve task performance. Unlike post-training approaches such as fine-tuning or reinforcement learning, TTC requires no additional training of the base model. As a result, inference-time alignment methods provide a lightweight yet powerful alternative that greatly simplifies deployment. Wellknown TTC methods include Best-of-N (BoN) sampling, chain-of-thought (CoT) reasoning, and their many variants (