Introducing the ERAs Project: Advancing AI Error-Reasoning in Science

Recent advances in artificial intelligence have fuelled the dream of building “AI scientists” – systems that can autonomously perform research and thus, potentially at least, accelerate the scientific discovery process. To realise this dream the novel AI agents need the ability to engage in the kind of reasoning that scientists mostly engage in, i.e., error-reasoning. Experiments often fail and the average researchers spends the majority of their time trying to figure out what has gone wrong.

It is, however, uncertain to what extent existing AI agents can reason about error. Underlying this uncertainty is the fact that scientific error-reasoning has not been widely or deeply datafied: scientists work through errors in weekly laboratory meetings, on whiteboards, or in the hallways of a conference venue. These discussions rarely find their way into published materials and are thus underrepresented in the data on which AI models are trained.

To address this uncertainty, we need benchmarks that allow us to assess the extent to which AI models can reason about scientific error. However, even our most sophisticated benchmarks for AI agents don’t currently test for this type of reasoning. Instead, they focus on the retrieval of textbook knowledge or on individual research tasks, such as the extraction of information from a figure in a paper.

Developing effective benchmarks requires a good understanding of error-reasoning in science: what are the types of errors scientists encounter and what are the strategies they usually deploy to address them? Unfortunately, we still lack a systematic and comprehensive theory of error in science. This means that we are not yet in a good position to build the benchmarks we need.

The ERAs Project will assemble an interdisciplinary team of researchers to address this problem from the bottom up. In the first phase we will develop an error framework for a specific scientific discipline, the life sciences (ErrorTheory). The narrow focus on one discipline will allow us to build a systematic and comprehensive account of error for a field that is currently a focal point of the development and deployment of AI Scientists. In the second phase of the project, we will use this ErrorTheory to build a systematic dataset which captures how researchers in the biology laboratory deal with different types of error (ErrorData). In the third phase we will then use ErrorData to construct a novel benchmark that can assess the error-reasoning ability of existing and future AI agents (ErrorBench). Looking into the future, ErrorData will also allow us to develop novel Error-Reasoning Agents (ERAs) by using it for supervised finetuning of frontier models. Ultimately, our project will help build the foundations for the development of reliable and powerful novel AI Scientists.

Details of project structure:

Duration: 48 months

Expected start date: September 2026

Staff included:  Two postdocs (Philosophy, 4 years and Machine Learning, 3 years)

                                Two fully funded PhD positions (3 ½ years each)

Key events:   Three 2-day conferences (Nov 2027; Nov 2028; Spring 2030)