Saved in:
| Main Authors: | , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2510.23334 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866914116366499840 |
|---|---|
| author | Quamar, Mohammad Atif Areeb, Mohammad Sharma, Nishant Shreekumar, Ananth Rosenthal, Jonathan Ozmen, Muslum Ozgur Kuznetsov, Mikhail Celik, Z. Berkay |
| author_facet | Quamar, Mohammad Atif Areeb, Mohammad Sharma, Nishant Shreekumar, Ananth Rosenthal, Jonathan Ozmen, Muslum Ozgur Kuznetsov, Mikhail Celik, Z. Berkay |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2510_23334 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Adaptive Blockwise Search: Inference-Time Alignment for Large Language Models Quamar, Mohammad Atif Areeb, Mohammad Sharma, Nishant Shreekumar, Ananth Rosenthal, Jonathan Ozmen, Muslum Ozgur Kuznetsov, Mikhail Celik, Z. Berkay Computation and Language 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. |
| title | Adaptive Blockwise Search: Inference-Time Alignment for Large Language Models |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2510.23334 |