TLDR: ReFactX is a new method that allows Large Language Models (LLMs) to access external knowledge from massive databases (up to 800 million facts) using “constrained generation.” This technique ensures LLMs only generate verified facts, significantly reducing hallucinations and improving accuracy in Question Answering tasks with minimal performance overhead, without needing complex external tools or models.
Large Language Models (LLMs) have revolutionized many areas, but they still face significant hurdles: knowledge gaps and a tendency to “hallucinate” or generate incorrect information. These issues arise when LLMs lack the specific facts needed to answer user questions accurately. While existing solutions like Retrieval-Augmented Generation (RAG) and tool-use attempt to bridge these gaps by incorporating external knowledge, they often introduce complexity, potential errors, and require processing a large volume of data.
A new research paper introduces ReFactX, a scalable method that empowers LLMs to access external knowledge directly, without the need for additional models or complex service pipelines. This innovative approach uses a technique called constrained generation, which is supported by a pre-built prefix-tree index. This index efficiently stores and allows access to a vast collection of facts derived from a Knowledge Graph.
Here’s how ReFactX works: Facts from a Knowledge Graph are converted into text, tokenized, and then organized into a prefix tree for quick retrieval. During the process of generating a response, if an LLM needs external information, ReFactX activates constrained generation. This mechanism ensures that the LLM can only generate sequences of tokens that form an actual, verified fact from the knowledge base. This guarantees the reliability and factual accuracy of the information provided.
The researchers evaluated ReFactX on various Question Answering tasks. The results demonstrated its impressive scalability, handling a massive knowledge base of 800 million facts with ease. It also showed strong adaptability to specialized, domain-specific data. Crucially, these benefits come with a minimal increase in generation time, adding only about 1% overhead.
The paper highlights that LLMs’ internal knowledge is limited to their training data, making them less effective for tasks requiring up-to-date or proprietary information. ReFactX offers a solution by allowing LLMs to integrate external knowledge seamlessly. When the LLM is instructed to find a fact, the constrained generation takes over, guiding the model to construct a valid fact from the knowledge base. Once the fact is complete, the LLM returns to its normal generation mode to continue reasoning.
For example, to answer “When was the director of Slumdog Millionaire born?”, ReFactX first helps the LLM reason about the steps. Then, using the “Fact:” command, constrained generation guides the model to identify “Danny Boyle” as the director and subsequently find his birth date, ensuring a correct and verifiable answer.
Key contributions of this work include ReFactX itself – a versatile wrapper that allows any LLM to tap into very large knowledge bases without external retrievers. It also features an efficient, disk-backed prefix tree that can manage hundreds of millions of facts with negligible latency. Empirical validation across four Question Answering benchmarks shows that ReFactX achieves competitive results and can boost accuracy by up to 20% compared to LLMs relying solely on their pre-trained knowledge, all while maintaining over 90% precision.
The paper also delves into how ReFactX compares to other methods. While input-based methods like RAG and tool-use can be effective, they often involve complex setups and can increase the number of tokens an LLM needs to process. Memory-based approaches, on the other hand, typically require changes to the LLM’s architecture. ReFactX’s use of constrained generation offers a more streamlined and integrated solution.
The scalability to 800 million facts from Wikidata is a testament to ReFactX’s robust design. It processes vast amounts of data, filters for relevant facts, and then tokenizes and indexes them in a PostgreSQL-backed prefix tree. This structure ensures rapid access and even prevents the LLM from generating the same fact repeatedly. The minimal generation-time overhead, just 1.3% for 4000 tokens, underscores its efficiency.
Experimental results consistently show ReFactX outperforming LLM-only models in terms of precision. While LLMs might perform adequately on questions covered by their training data, ReFactX significantly enhances performance on more challenging datasets, such as 2WikiMultiHopQA, where external knowledge is crucial. It also demonstrates strong performance on domain-specific data, like the proprietary financial dataset used in the study.
However, the researchers acknowledge some limitations. Due to the left-to-right nature of LLM generation, ReFactX works best with facts that progress from known to desired information. It also faces challenges with complex “count” questions (e.g., “How many movies has Danny Boyle directed?”) or long enumerations, which might be better handled by specialized tools like SPARQL engines. Future work aims to address these limitations, including fine-tuning LLMs to better utilize ReFactX and expanding its capabilities for more complex query types.
Also Read:
- Integrating Knowledge Graphs for Advanced Multi-hop Question Answering in Language Models
- SSFO: A Self-Supervised Method for Improving LLM Faithfulness in RAG Systems
ReFactX represents a promising advancement in making LLMs more reliable and factually accurate. It offers a lightweight, easy-to-integrate solution for grounding LLM responses in vast external knowledge bases. For those interested in the technical details and to explore the code, the project is available on GitHub.
