LAG: Logic-Augmented Generation from a Cartesian Perspective

TL;DR

This work tackles hallucinations and reasoning failures in large language models by introducing Logic-Augmented Generation (LAG), a Cartesian-inspired pipeline that emphasizes systematic problem decomposition, an atomic memory bank, and controllable, logic-guided reasoning. LAG decomposes complex questions into logically dependent sub-questions, uses prior answers to guide retrieval, and terminates unproductive branches to prevent error propagation, ultimately synthesizing grounded final answers. Empirical results on HotpotQA, MuSiQue, 2WikiMultiHopQA, and GraphRAG-Bench show substantial accuracy gains and reduced hallucinations compared to Vanilla RAG and GraphRAG baselines, with notable improvements in challenging reasoning tasks. The framework highlights the practical impact of integrating structured logic with retrieval for robust knowledge-intensive tasks, while outlining avenues for multimodal extensions and ethical considerations.

Abstract

Large language models (LLMs) have demonstrated remarkable capabilities across a wide range of tasks, yet exhibit critical limitations in knowledge-intensive tasks, often generating hallucinations when faced with questions requiring specialized expertise. While retrieval-augmented generation (RAG) mitigates this by integrating external knowledge, it struggles with complex reasoning scenarios due to its reliance on direct semantic retrieval and lack of structured logical organization. Inspired by Cartesian principles from \textit{Discours de la méthode}, this paper introduces Logic-Augmented Generation (LAG), a novel paradigm that reframes knowledge augmentation through systematic question decomposition, atomic memory bank and logic-aware reasoning. Specifically, LAG first decomposes complex questions into atomic sub-questions ordered by logical dependencies. It then resolves these sequentially, using prior answers to guide context retrieval for subsequent sub-questions, ensuring stepwise grounding in the logical chain. Experiments on four benchmarks demonstrate that LAG significantly improves accuracy and reduces hallucination over existing methods.

Figures (3)

• Figure 1: Comparison of three RAG paradigms. LAG offers a superior balance of efficiency and accuracy, providing a more lightweight solution than GraphRAG while outperforming it and traditional RAG in accuracy.
• Figure 2: The framework of LAG. (I) Adaptive question decomposition splits complex queries into atomic sub-questions using a cognitive load. (II) Logical chain reasoning resolves sub-Q by the order of logical dependency, utilizing an atomic memory bank for recurrent knowledge. Logical terminator halts unreliable chains early. Finally, answers are synthesized via integrated generation (complete chains) or alternative solution (terminated chains).
• Figure 3: Ablation study of proposed LAG on three standard datasets.