Researchers propose "Logits-to-Logic" framework to fix LLMs' logical inconsistencies in structured knowledge tasks

Large language models excel at understanding and reasoning through natural language, but they struggle when working with structured knowledge like knowledge graphs. This gap creates what researchers call "Logic Drift"—instances where LLMs fail to maintain logical consistency when required to answer questions or perform reasoning tasks on formally organized data.

The Core Problem#

Traditional approaches to this challenge rely on prompt-based workflows—essentially writing better instructions to guide LLMs toward more logical outputs. However, these methods only influence what goes into the model, leaving the underlying logical inconsistencies in the model's outputs unaddressed. Additionally, such rigid workflows often cannot flexibly adapt to different tasks or knowledge structures.

A New Solution: Logits-to-Logic#

Researchers have introduced the Logits-to-Logic framework, which takes a fundamentally different approach. Rather than tweaking prompts, the technique intervenes at the mathematical level where language models generate text. Specifically, it works with the logits—the numerical scores that determine which word the model will output next—using two key mechanisms:

• Logits strengthening: reinforces the model's correct logical choices
• Logits filtering: removes or suppresses outputs that violate logical consistency

Demonstrated Impact#

The framework was evaluated on Knowledge Graph Question Answering (KGQA) benchmarks, where models must answer questions by reasoning over structured knowledge graphs. Results show significant improvements in logical consistency and state-of-the-art performance across multiple benchmark datasets.

By operating at the generation level rather than relying solely on input-level guidance, this approach more directly addresses the representational mismatch that causes LLMs to falter when moving from unstructured text to highly structured knowledge domains.