Title: From Last Layer Logits to Logic: Enhancing LLMs with Logic-Consistent Structured Knowledge Reasoning

Abstract:

Large Language Models (LLMs) have demonstrated remarkable proficiency in natural language reasoning by leveraging pre-training on extensive unstructured text corpora. This foundation allows them to comprehend linguistic logic and produce responses that are logically sound. However, the inherent disparity between unstructured and structured knowledge formats causes LLMs to frequently struggle with maintaining logical coherence. This issue manifests as Logic Drift, particularly in structured knowledge reasoning tasks like Knowledge Graph Question Answering (KGQA). Current solutions attempt to mitigate this limitation by embedding intricate, prompt-based workflows to steer the reasoning process. Yet, these methods offer only superficial, input-level guidance and do not resolve the root cause of Logic Drift in the model's output. Furthermore, their rigid reasoning structures lack the flexibility to adapt across diverse tasks and knowledge graphs.

To bolster the logical consistency of LLMs when handling structured knowledge, this study focuses on the logits generated during the final stage of autoregressive decoding. We introduce the Logits-to-Logic framework, a novel approach that utilizes logits strengthening and logits filtering as its primary mechanisms to rectify logical flaws in LLM outputs. Our comprehensive experiments demonstrate that this method substantially enhances the logical consistency of LLMs in structured knowledge reasoning, securing state-of-the-art results across multiple KGQA benchmarks.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC