According to PANews, the deployment of AI large language models (LLMs) in sectors such as finance, healthcare, and law faces significant challenges due to the "hallucination" problem, where AI outputs may not match the precision required in these fields. Mira Network has introduced a public test network offering a solution to this issue.
AI LLMs often exhibit hallucinations due to two main reasons: incomplete training data, which leads to creative completions in niche or specialized areas, and reliance on probabilistic sampling, which identifies statistical patterns rather than true understanding, causing inconsistencies in handling high-precision factual queries.
A study published on Cornell University's ArXiv platform proposes a method to enhance the reliability of LLM outputs through multi-model verification. This approach involves generating results with a primary model and then using multiple verification models for majority vote analysis, reducing hallucinations and increasing accuracy to 95.6%.
Mira Network has developed a distributed verification platform to manage and verify interactions between primary and verification models. This middleware network provides a reliable validation layer between users and foundational AI models, enabling services such as privacy protection, accuracy assurance, scalable design, and standardized API interfaces. By minimizing AI LLM hallucinations, Mira Network expands AI's applicability across various specialized scenarios.
Examples of Mira Network's application include:
1) Gigabrain, a trading platform, integrates Mira to verify market analysis and prediction accuracy, filtering unreliable suggestions and enhancing AI trading signals' reliability in DeFi scenarios.
2) Learnrite uses Mira to validate AI-generated standardized exam questions, allowing educational institutions to utilize AI-generated content without compromising test accuracy, maintaining strict educational standards.
3) The blockchain Kernel project incorporates Mira's LLM consensus mechanism into the BNB ecosystem, creating a decentralized verification network (DVN) that ensures the accuracy and security of AI computations on the blockchain.
Mira Network offers a middleware consensus network service, which is not the sole method to enhance AI application capabilities. Other options include data-driven training enhancements, multimodal model interactions, and privacy computing enhancements through cryptographic technologies like ZKP, FHE, and TEE. However, Mira's solution is notable for its rapid implementation and immediate effectiveness.