Deliberation on Priors: Trustworthy Reasoning of Large Language Models on Knowledge Graphs
Jie Ma, Ning Qu, Zhitao Gao, Rui Xing, Jun Liu, Hongbin Pei, Jiang Xie, Linyun Song, Pinghui Wang, Jing Tao, Zhou Su
2025-05-22
Summary
This paper talks about a new method called Deliberation on Priors, which helps large language models make more trustworthy decisions by using extra information from knowledge graphs.
What's the problem?
AI models sometimes make mistakes or give answers that don't make sense because they don't always use reliable background knowledge or follow important rules when reasoning.
What's the solution?
The researchers improved the reasoning of language models by having them learn from knowledge graphs, which are like big maps of facts and relationships, and by teaching the models to check their own reasoning using these trusted sources.
Why it matters?
This matters because it makes AI more accurate and dependable, which is really important for things like research, education, and any situation where you need to trust the answers you get.
Abstract
The Deliberation over Priors framework enhances the trustworthiness of LLMs by integrating structural and constraint priors from knowledge graphs through knowledge distillation and reasoning introspection.