文章背景与核心概要
在处理复杂的现实世界问题时，非演绎推理（包括归纳和溯因推理）扮演着至关重要的角色。现实世界中往往存在多种能够合理解释观察结果的假设，而“奥卡姆剃刀原理”（即倾向于选择最简单的有效解释）则是科学和人类决策的核心准则。然而，过去针对大语言模型（LLM）非演绎推理能力的评估大多忽视了对“简约性”的考察。
为了填补这一空白，本文作者引入了一个全新的综合框架，用于生成结合了归纳与溯因任务的一阶逻辑推理问题，并提出了一个能够量化评估假设简单性的自动化指标。研究结果表明，尽管当前最先进的大语言模型能够在简单场景下完成基本的归纳与溯因推理，但在面对复杂的世界模型时依然力不从心，且即便使用了上下文学习（In-Context Learning）或基于可验证奖励的强化学习（RLVR）等主流增强技术，它们也难以持续生成高质量、符合奥卡姆剃刀原理的简约假设。
Do Language Models Follow Occam's Razor? An Evaluation of Parsimony in Inductive and Abductive Reasoning
arXiv: 2509.03345 [cs.AI]
Authors: Yunxin Sun, Abulhair Saparov
Accepted at: EMNLP 2026 Main Conference
arXiv: 2509.03345 [cs.AI]
Authors: Yunxin Sun, Abulhair Saparov
Accepted at: EMNLP 2026 Main Conference
📌 Summary
📌 Summary
本文探讨了大语言模型（LLM）在执行非演绎推理时是否遵循奥卡姆剃刀原理（即偏爱最简单的有效解释）。虽然现实世界的解题过程往往会产生多个有效的假设，但以往对LLM的评估在很大程度上忽略了简约性（Parsimony）这一维度。
This paper investigates whether Large Language Models (LLMs) adhere to Occam's Razor (favoring the simplest valid explanation) when performing non-deductive reasoning. While real-world problem-solving often yields multiple valid hypotheses, prior evaluations of LLMs have largely ignored parsimony.
为了弥补这一研究空白，作者引入了一个新颖的合成框架，用于生成结合了归纳任务与溯因任务的一阶逻辑推理问题，同时配备了评估假设简单性的自动化指标。研究发现，尽管顶尖的LLM能够应付基础场景，但在面对复杂的世界模型时却一筹莫展，且即便采用上下文学习或RLVR等先进技术，它们也无法稳定生成高质量、符合简约原则的假设。
To address this gap, the authors introduce a novel synthetic framework for generating first-order logic reasoning questions that combine inductive and abductive tasks, alongside an automated metric to evaluate hypothesis simplicity. Their findings reveal that while state-of-the-art LLMs manage basic scenarios, they consistently struggle with complex world models and fail to generate high-quality, parsimonious hypotheses—even when using advanced techniques like in-context learning or RLVR.
📋 Metadata & Submission Details
📋 Metadata & Submission Details
- 学科领域： 人工智能（
cs.AI）；计算与语言（cs.CL） - 首次提交： 2025年9月3日
- 最新修订： 2026年8月26日（第三版 v3）
- DOI： 10.48550/arXiv.2509.03345
- Subjects: Artificial Intelligence (
cs.AI); Computation and Language (cs.CL)- First Submitted: September 3, 2025
- Latest Revision: August 26, 2026 (Version v3)
- DOI: 10.48550/arXiv.2509.03345
📝 Abstract
📝 Abstract
非演绎推理涵盖归纳推理与溯因推理，对于解决复杂的现实世界问题至关重要。归纳和溯因推理的一个关键特征在于：通常存在许多有效的假设，而其中最简单（遵循奥卡姆剃刀原理）的假设往往最有用。然而，近来评估大语言模型（LLM）非演绎推理能力的工作却忽略了这一维度。本研究填补了这一空白，重点探讨LLM的归纳与溯因推理能力是否遵循奥卡姆剃刀原理，同时对其推理的正确性进行了检验。为此，我们引入了一个框架，能够综合生成满足以下条件的推理问题：(a) 同时要求归纳推理与溯因推理；(b) 能够轻松扩展以生成任何可由一阶逻辑表达的溯因/归纳推理问题。智能体的任务是在给定的世界模型下，产出能够解释观察结果的假设。我们还提出了一种新的自动化指标，以定量评估假设是否符合奥卡姆剃刀原理；只有既正确又最简单的假设才被认为是高质量的。我们对当前最先进的LLM进行的评估表明，LLM能够在简单场景下执行归纳和溯因推理，但在复杂的世界模型以及生成高质量假设方面表现挣扎，哪怕使用了诸如上下文学习和RLVR等广受欢迎的推理增强技术也难以克服这一缺陷。
Non-deductive reasoning, encompassing inductive and abductive reasoning, is essential in addressing complex real-world questions. One key feature of inductive and abductive reasoning is that there are many valid hypotheses; the simplest ones (those that adhere to Occam's Razor) are often most useful. However, this aspect is ignored in recent work that evaluates the non-deductive reasoning capabilities of large language models (LLMs). This work fills this gap, focusing on understanding whether the inductive and abductive reasoning capabilities of LLMs adhere to Occam's Razor, while also examining the correctness of their reasoning. To accomplish this goal, we introduce a framework to synthetically generate reasoning questions that (a) require inductive reasoning and abductive reasoning simultaneously; (b) is readily extended to produce any abductive/inductive reasoning question expressible in first-order logic. The task for the intelligent agent is to produce hypotheses to explain observations under a given world model. We also propose a new automated metric to assess whether hypotheses quantitatively adhere to Occam's Razor; those hypotheses that are correct and simplest are considered high-quality. Our findings on state-of-the-art LLMs suggest that LLMs can perform inductive and abductive reasoning in simple scenarios, but struggle with complex world models and with producing high-quality hypotheses, even with popular reasoning-enhancing techniques such as in-context learning and RLVR.
🔗 Full-Text & Resources
🔗 Full-Text & Resources
