Decoding on Graphs: Empower LLMs with KGs (MIT)

Decoding on Graphs (DoG) is a novel framework that integrates Large Language Models (LLMs) with Knowledge Graphs (KGs) to enhance question-answering (QA) tasks. The key innovation is the concept of a "well-formed chain", a sequence of interconnected fact triplets from the KG that starts with entities mentioned in the question and leads logically to the answer. To ensure that LLMs generate these chains faithfully, DoG introduces graph-aware constrained decoding, which dynamically restricts the LLM's token generation based on the KG's topology. This is achieved using a trie data structure that maintains valid token sequences corresponding to the KG's paths, ensuring that each generated reasoning step aligns with actual KG relations and entities. Additionally, DoG employs beam search execution at the triplet level to explore multiple plausible reasoning paths simultaneously, enhancing the robustness and accuracy of the answers. By constraining the LLM's decoding process without altering its underlying parameters or requiring additional training, DoG effectively leverages the LLM's inherent reasoning capabilities while grounding its outputs in the structured knowledge of the KG. Experimental results demonstrate that DoG outperforms existing methods on several KGQA benchmarks, particularly in complex multi-hop reasoning scenarios, highlighting its effectiveness and adaptability across different KGs and LLMs. all rights w/ authors: Decoding on Graphs: Faithful and Sound Reasoning on Knowledge Graphs through Generation of Well-Formed Chains https://arxiv.org/pdf/2410.18415 Nice Idea by ‪@mit‬ and Univ of Hong Kong 00:00 Augment LLMs with Knowledge Graphs 00:47 Subgraph retrievers 02:03 Agents for Integrating LLM and KG 04:25 NEW IDEA by MIT & HK Univ 08:33 Example of Decode on Graphs 15:21 Implementation PROMPT DoG 17:20 Linear graph forms 19:43 Graph aware constrained decoding 24:20 Harvard MED Agents for LLM on KG #aiagents #airesearch #artificialintelligence #massachusettsinstituteoftechnology