基于多视角的大语言模型一阶逻辑推理增强
Multi-View-Based Enhancement of First-Order Logical Reasoning with Large Language Models
顾铭贤 1邓赵红1
作者信息
- 1. 江南大学人工智能与计算机学院，无锡 214122
- 折叠
摘要
面向规则密集且包含否定例外与量词作用域的自然语言推理任务，针对大语言模型推理过程难以机检、结论易波动以及错误成因难定位的问题，提出一种由"一阶逻辑形式化-求解器反馈"驱动的推理增强框架。该框架首先将原始文本进行证据片段单元化，并通过多视角并行抽取提升语义覆盖，统一映射为结构化中间表示；随后采用确定性规范化与重写聚合缓解同义表达与符号不一致，并在 Z3 约束下执行一致性融合与增量可满足性检查，从而构造与答案绑定的可执行理论。当理论出现冲突或不可满足时，求解器反馈用于定位关键冲突来源并触发局部修复，形成闭环的生成-验证-修正流程，同时以结构化日志保留证据来源与诊断信息以支持可追溯分析。实验结果表明，在五个标准数据集上，该方法整体优于 Naive、CoT、Logic-LM 与 SymbCoT；消融分析进一步验证多视角覆盖与求解器引导的一致性融合是性能提升的主要来源。结果说明，将正确性压力下沉至可执行符号层并利用求解器反馈进行闭环校正，能够在开销可控的前提下提升推理准确性、鲁棒性与可解释性。
Abstract
For natural language reasoning tasks that are rule-intensive and involve negation exceptions and quantifier scope, this work addresses the challenges that large language models face in making reasoning processes machine-checkable, producing stable conclusions, and localizing the causes of errors. A reasoning enhancement framework driven by "first-order logic formalization-solver feedback" is proposed. The framework first segments the input text into evidence spans and performs multi-view parallel extraction to improve semantic coverage, mapping the results into a structured intermediate representation. It then applies deterministic normalization and rewrite-based aggregation to mitigate paraphrasing variance and symbol inconsistencies, and conducts solver-constrained consistency fusion with incremental satisfiability checking in Z3 to construct an executable theory aligned with the final answer. When conflicts or unsatisfiability arise, solver feedback is used to identify key sources of inconsistency and trigger localized repairs, forming a closed loop of generation-verification-correction, while structured logs retain evidence provenance and diagnostic signals for traceable analysis. Experiments on five benchmarks aligned with the SymbCoT protocol show consistent improvements over Naive, CoT, Logic-LM, and SymbCoT, and ablation studies further confirm that multi-view coverage and solver-guided consistency fusion are the primary contributors to performance gains. The results indicate that shifting correctness pressure to an executable symbolic layer and leveraging solver feedback for closed-loop correction can improve reasoning accuracy, robustness, and interpretability under controlled cost.关键词
计算机科学与技术/大语言模型/神经符号推理/规则驱动/思维机制Key words
Computer Science and Technology/Large Language Models/Neuro-Symbolic Reasoning/Rule-Driven/Reasoning Mechanisms引用本文复制引用
顾铭贤,邓赵红.基于多视角的大语言模型一阶逻辑推理增强[EB/OL].(2026-02-02)[2026-02-03].http://www.paper.edu.cn/releasepaper/content/202602-2.学科分类
计算技术、计算机技术
评论