Blockchain and LLMs: Secure Knowledge Sharing for Reliable AI Responses.

The increasing prevalence of large language models (LLMs) necessitates reliable access to external knowledge to mitigate the issue of ‘hallucination’, where models generate factually incorrect or nonsensical outputs. However, valuable data often resides in isolated systems, creating barriers to effective knowledge integration due to privacy and security protocols. Researchers at Shanghai Jiao Tong University and China Telecom Research Institute, including Zhaojiacheng Zhou, Hongze Liu, Shijing Yuan, Hanning Zhang, Jiong Lou, Chentao Wu, and Jie Li, address this challenge in their work, “BLOCKS: Blockchain-supported Cross-Silo Knowledge Sharing for Efficient LLM Services”. They propose a blockchain-based framework designed to coordinate knowledge across disparate systems, enabling secure and efficient access to foundational information for LLMs, while incentivising participation through a reputation and cross-validation system.

Large language models (LLMs) increasingly exhibit a tendency to generate inaccurate information, commonly termed ‘hallucinations’, prompting investigation into methods of augmenting them with dependable external knowledge. This work addresses the challenge of fragmented, siloed knowledge, often inaccessible due to privacy and security restrictions, by proposing a novel blockchain-based framework designed to coordinate multiple disparate knowledge sources. The system actively distills local data into concise prompts, meticulously recording all transactions on a blockchain to ensure secure and verifiable knowledge retrieval, thereby enhancing LLM performance.

The framework centres on a robust reputation mechanism, implemented to incentivise participation and maintain the quality of shared knowledge. This system assesses contributions from various sources and assigns corresponding reputation scores, directly influencing the trustworthiness of information presented to the LLM. Cross-validation techniques further enhance reliability by rigorously verifying knowledge consistency across independent sources, reducing the potential for misinformation. Researchers designed a sophisticated query generation framework, providing an application programming interface (API) for efficient knowledge retrieval, streamlining the process of accessing and integrating external data. An API is a set of rules and specifications that software programs can follow to communicate with each other.

Experiments demonstrate the framework’s ability to facilitate efficient knowledge sharing within a secure blockchain environment, significantly improving LLM performance. The system leverages blockchain’s inherent security features to protect sensitive data while simultaneously enabling access to a broader knowledge base for LLMs, fostering more reliable outputs. Results indicate that the proposed approach effectively mitigates the hallucination problem by grounding LLM responses in verified, external knowledge, leading to more trustworthy and informative AI applications.

The work builds upon existing research in several key areas, integrating advancements in blockchain technology, reputation systems, and LLM prompting techniques. Studies on blockchain interoperability, such as Belchior et al. (2021), inform the design of a system capable of seamlessly integrating diverse knowledge silos, overcoming compatibility challenges. Furthermore, the incorporation of robust reputation systems, explored by Hu et al. (2018), directly addresses the critical need for incentivising accurate and reliable knowledge contributions, fostering a collaborative environment. The framework also draws upon advancements in LLM prompting techniques, aiming to effectively translate distilled knowledge into a format readily usable by the language model, maximising information transfer.

Future work will focus on scaling the framework to accommodate larger datasets and more complex knowledge domains, exploring different consensus mechanisms to optimise performance and security, and investigating the potential for integrating the framework with other AI technologies. Consensus mechanisms are methods for achieving agreement within a distributed system, such as a blockchain. Researchers also plan to develop a user-friendly interface to facilitate the contribution and validation of knowledge, fostering a broader community of contributors and ensuring the long-term sustainability of the framework. Ultimately, this research aims to create a more reliable and trustworthy AI ecosystem, empowering users with access to accurate and verifiable information.