Adaptive Blockwise Search: Inference-Time Alignment for Large Language Models

TL;DR

AdaSearch is an adaptive, inference-time alignment method for large language models that allocates computational effort dynamically to critical tokens, significantly improving alignment quality across various tasks.

Contribution

The paper introduces AdaSearch, a novel blockwise search strategy that adaptively allocates computational resources during inference for better alignment.

Findings

AdaSearch outperforms Best-of-N and fine-tuning baselines.

Win-rates improve by over 10% in key alignment tasks.

Effective across eight large language models.

Abstract

LLM alignment remains a critical challenge. Inference-time methods provide a flexible alternative to fine-tuning, but their uniform computational effort often yields suboptimal alignment. We hypothesize that for many alignment tasks, the initial tokens of a response are disproportionately more critical. To leverage this principle, we introduce AdaSearch, a novel blockwise search strategy. It adaptively allocates a fixed computational budget using a sampling schedule, focusing search effort on these critical tokens. We apply AdaSearch to sequential decoding and introduce its tree-search counterpart, AdaBeam. Our comprehensive evaluation across eight LLMs demonstrates that AdaSearch outperforms strong Best-of-N and fine-tuning baselines. Specifically, win-rates improve by over 10% for harmlessness generation, controlled sentiment generation, and for mathematical reasoning tasks relative to Best-of-N.