Last Layer Logits to Logic: Empowering LLMs with Logic-Consistent Structured Knowledge Reasoning

Name: Last Layer Logits to Logic: Empowering LLMs with Logic-Consistent Structured Knowledge Reasoning
Uploaded: 2026-10-06T14:39:44.774164+00:00
Duration: 446 s
Description: The Logits-to-Logic framework addresses 'Logic Drift' in LLMs when answering Knowledge Graph questions by aligning their output probabilities with the question and KG logic. It introduces three stages—compiling legal paths into an NFA, strengthening relevant logits, and filtering irrelevant ones—to ensure the model generates logically consistent reasoning paths.

The Logits-to-Logic framework addresses 'Logic Drift' in LLMs when answering Knowledge Graph questions by aligning their output probabilities with the question and KG logic. It introduces three stages—compiling legal paths into an NFA, strengthening relevant logits, and filtering irrelevant ones—to ensure the model generates logically consistent reasoning paths.

Open the full episode