Research/
MIT Researchers Develop New Metric to Quantify LLM Reliability
Researchers have introduced 'semantic entropy,' a method that detects when large language models are likely to hallucinate by analyzing the consistency of meanings in generated responses. This technique allows systems to self-flag uncertainty, significantly improving the accuracy of AI-driven tasks without requiring external reference data.