Paper Type
ERF
Abstract
As large language models (LLMs) take on increasingly consequential roles in scientific and applied reasoning, their capabilities demand rigorous examination. While current Natural Language Processing (NLP) research extensively accounts for deductive and inductive logic, abductive reasoning, which is the critical ability to generate plausible explanations from incomplete observations, is an underexplored frontier. This research aims to address this gap by developing a comprehensive evaluation and enhancement framework for generative abductive reasoning in LLMs that moves beyond the discriminative selection approach. First, we will construct a novel dataset that operationalizes genuine hypothesis generation. Second, we will introduce multi-dimensional metrics that assess the plausibility and coherence of generated explanations. Finally, we will systematically evaluate advanced inference strategies and alignment techniques in terms of abductive performance. By shifting the paradigm from selection to generation, this work provides researchers with robust tools to measure and improve LLM reasoning, ultimately fostering verifiable trust in emergent systems.
Paper Number
1468
Recommended Citation
Venkatakrishnan, Radhakrishnan, "Abductive Reasoning with LLM" (2026). AMCIS 2026 Proceedings. 19.
https://aisel.aisnet.org/amcis2026/conftheme/conftheme/19
Abductive Reasoning with LLM
As large language models (LLMs) take on increasingly consequential roles in scientific and applied reasoning, their capabilities demand rigorous examination. While current Natural Language Processing (NLP) research extensively accounts for deductive and inductive logic, abductive reasoning, which is the critical ability to generate plausible explanations from incomplete observations, is an underexplored frontier. This research aims to address this gap by developing a comprehensive evaluation and enhancement framework for generative abductive reasoning in LLMs that moves beyond the discriminative selection approach. First, we will construct a novel dataset that operationalizes genuine hypothesis generation. Second, we will introduce multi-dimensional metrics that assess the plausibility and coherence of generated explanations. Finally, we will systematically evaluate advanced inference strategies and alignment techniques in terms of abductive performance. By shifting the paradigm from selection to generation, this work provides researchers with robust tools to measure and improve LLM reasoning, ultimately fostering verifiable trust in emergent systems.
When commenting on articles, please be friendly, welcoming, respectful and abide by the AIS eLibrary Discussion Thread Code of Conduct posted here.
Comments
NEXTTRANS