We have released annotated slides for a talk titled Evaluating LLMs is a minefield. We show that current ways of evaluating chatbots and large language models don't work well, especially for questions about their societal impact. There are no quick fixes, and research is needed to improve evaluation methods.

The challenges we highlight are somewhat distinct from those faced by builders of LLMs or by developers interested in comparing between LLMs for adoption. Those challenges are better understood and tackled by evaluation frameworks such as HELM. 

You can view the annotated slides here.

The slides were originally presented at a launch event for Princeton Language and Intelligence, a new initiative to strengthen LLM access and expertise in academia.

The talk is based on the following previous posts from our newsletter:

• Is GPT-4 getting worse over time?

• Does ChatGPT have a liberal bias?

• Generative AI companies must publish transparency reports

• GPT-4 and professional benchmarks: the wrong answer to the wrong question

• ML is useful for many things, but not for predicting scientific replicability

• OpenAI’s policies hinder reproducible research on language models

• Licensing is neither feasible nor effective for addressing AI risks

• Is the future of AI open or closed?