2025-2026 Topic: Develop the techniques, procedures, and processes to validate the output of Generative AI models in conducting scientific and technical analysis.

Large Language Models and Generative Artificial Intelligence are significantly altering the workplace and portend a major technological disruption. Their ability to generalize and extract relationships from vast data sets has significant potential with far reaching impact.  While their initial adoption in common tasks, with a knowledgeable human reviewer, have proven effective, their use in conducting more detailed scientific and technical analysis requires further scrutiny.  LLM and neural networks in general will provide a response whether it is factually correct on not.  There are plenty of (and sometimes funny) anecdotes where they have produced responses, that to a human expert, are obviously incorrect.  Yet, they are being used and applied in ever increasing fields and their ability to find hidden relationships in data can exceed that of the human expert.  Therefore, to fully exploit the technology, innovative verification frameworks driven by large language models that incorporate fact-grounding, source traceability and provenance tracking must be developed to provide a measure of technical assurance that the output is correct. A confidence score is not enough; analysts need traceability of information, source reliability assessment, and transparency of analytic reasoning.

Some desirable features:

• Confidence & Risk Indicators: The proposed solution will enable analysts to find ways to measure, test, and validate correctness of S&TI outputs.

• Multi-Source Correlation: The proposed solution will help the analyst in comparing outputs across independent sources, by producing cross-source consistency checks and adversarial alternative hypotheses.

• Entity Relationship Extraction: How do we verify that the extracted entities/links are real and not hallucinations?

• Hypothesis Generation: LLMs possess the ability to suggest hypotheses, but which validation methods are suitable to scrutinize such hypotheses? The proposed solution will help analysts develop hypothesis matrices scored against authoritative sources.