Venue: Lecture Theatre 7, Diamond Building.

Vibe research - are we close? 

Programming supported by large language models has existed for several years. However, recent improvements in model capabilities have significantly accelerated their use, giving rise to what is often called “vibe coding”: programs can be produced without a deep understanding of the underlying code. In many cases, agentic systems can not only generate the program but also design test cases and iteratively refine the implementation.

AI is now being explored in many other areas, including research itself. Several commercial tools now offer a “research mode,” and media reports suggest that systems already exist that can propose plausible research directions. More recently, an organisation has demonstrated a system that can propose research questions, conduct computational experiments, write the code required for the experiments, and produce a paper describing the results.

These developments raise the possibility of “vibe research”, where (computational) research can be performed without detailed knowledge of how the underlying methods are implemented.

This raises several questions, both in general and specifically for the speech and language community:

• Do we already have the components needed for a full “vibe research” system? If not, what pieces are still missing?

• Are large language models sufficient to propose meaningful research questions and conduct the associated work?

• How might this change the role of researchers and the skills expected from them?

• What could be the impact on peer review, publication practices, research funding and the future structure of research?

• How do we evaluate research produced largely by automated systems?

• What should students learn if implementation is increasingly automated?

• How do we prevent socially undesirable research from being carried out?