Thomas Schmidhuber

Large Language Models are meanwhile a standard-tool in technical documentation. But still, hallucinations occur and the content provided by the model is not fully reliable. This talk addresses, how and why hallucinations are generated by language models and what users can do, to reduce the amount of unreliable information.

Besides a good prompt, the design of context for the model becomes vital to achieve the results the user expects. No matter whether LLMs are used for research or generation of new content, the proper setting of the context will improve the quality of the model answers. Understanding how inputs lead to hallucinations and how the user input can be design to reduce the amount of false information improves the efficiency when using language models in the daily work-flow.