Saved in:
Bibliographic Details
Main Authors: Li, Sophie, Huang, Nicholas, Saxena, Nayan, Luo, Nina, Lin, Vincent, Zhu, Kevin, Dev, Sunishchal
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2511.00699
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914133312536576
author Li, Sophie
Huang, Nicholas
Saxena, Nayan
Luo, Nina
Lin, Vincent
Zhu, Kevin
Dev, Sunishchal
author_facet Li, Sophie
Huang, Nicholas
Saxena, Nayan
Luo, Nina
Lin, Vincent
Zhu, Kevin
Dev, Sunishchal
contents Large language models (LLMs) improve reasoning accuracy when generating multiple candidate solutions at test time, but standard methods like Best-of-N (BoN) incur high computational cost by fully generating all branches. Self-Truncation Best-of-N (ST-BoN) mitigates this by truncating unpromising paths early, but its reliance on consistency-based heuristics is a limitation as it does not directly evaluate branch quality. We present KL-Adjusted Pruned Path Algorithm (KAPPA), an inference-time method that combines Kullback-Leibler divergence, confidence, and entropy into a principled scoring function to guide progressive pruning. By promoting diversity during exploration and selectively eliminating low-scoring branches, KAPPA maintains accuracy while substantially reducing memory and token usage. Experiments on GSM8K and MATH500 with DeepSeek-R1-Distill-Qwen-1.5B and Qwen2.5-7B-Instruct demonstrate that KAPPA stabilizes performance in smaller models and achieves up to ~60% reduction in peak memory and ~90% reduction in total token generation relative to BoN, with minimal impact on accuracy.
format Preprint
id arxiv_https___arxiv_org_abs_2511_00699
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Inference-Time Chain-of-Thought Pruning with Latent Informativeness Signals
Li, Sophie
Huang, Nicholas
Saxena, Nayan
Luo, Nina
Lin, Vincent
Zhu, Kevin
Dev, Sunishchal
Machine Learning
Large language models (LLMs) improve reasoning accuracy when generating multiple candidate solutions at test time, but standard methods like Best-of-N (BoN) incur high computational cost by fully generating all branches. Self-Truncation Best-of-N (ST-BoN) mitigates this by truncating unpromising paths early, but its reliance on consistency-based heuristics is a limitation as it does not directly evaluate branch quality. We present KL-Adjusted Pruned Path Algorithm (KAPPA), an inference-time method that combines Kullback-Leibler divergence, confidence, and entropy into a principled scoring function to guide progressive pruning. By promoting diversity during exploration and selectively eliminating low-scoring branches, KAPPA maintains accuracy while substantially reducing memory and token usage. Experiments on GSM8K and MATH500 with DeepSeek-R1-Distill-Qwen-1.5B and Qwen2.5-7B-Instruct demonstrate that KAPPA stabilizes performance in smaller models and achieves up to ~60% reduction in peak memory and ~90% reduction in total token generation relative to BoN, with minimal impact on accuracy.
title Inference-Time Chain-of-Thought Pruning with Latent Informativeness Signals
topic Machine Learning
url https://arxiv.org/abs/2511.00699