[1Win4] Poster session 1

[1Win4-39] Enhancing Abductive Reasoning in LLMs Through Premises Retrieval

〇Yuanyi Wang¹, Ichiro Kobayashi¹ (1.Ochanomizu University)

Keywords:Large language model, Reasoning

We propose a novel algorithmic pipeline that utilizes Monte Carlo Markov Chain (MCMC) to enhance the reasoning capabilities of large language models (LLMs), with a particular focus on abductive reasoning. To efficiently and cost-effectively identify hypotheses that best explain observations and their supporting premises, we introduce a fully unsupervised MCMC algorithm. By prioritizing the arrangement of highly relevant premises, the proposed method effectively constructs valid hypotheses within LLMs, leading to improved text generation accuracy and increased recall of premise knowledge. When tested on the Entailment Bank dataset, the approach increases the premise retrieval recall rate, enabling LLMs to generate more plausible answers that closely align with the ground truth.