aes24 International Evaluation Conference

Authors: Gerard Atkinson (ARTD Consultants)

Last year saw the rise of large language models, with words like Chat-GPT and Bard becoming part of common discussion. The evaluation community was not immune to this trend, and papers were published that looked at just how well machine learning approaches could do against human evaluators on topics such as qualitative analysis and evaluative judgement. The answer? Not as well as you would think (but you could get wrong answers faster than ever!)

But the designers of these models took on the feedback and created newer and more sophisticated tools. In addition, there have been innovations in hybrid models which combine the best features of different methods while minimising their weaknesses. Coupled to this is the growing field of standalone models that can be run on a desktop computer but produce responses that match or exceed cloud-based models, and models that can draw on rich contextual information (such as documentation or full interview transcripts) to make decisions.

This presentation provides an update on the state of machine learning in 2024 and presents new findings in relation to the performance of machine learning models on tasks including topic classification and rubric analysis.