Large language models have extensive pretraining knowledge and strong reasoning skills, yet they struggle to adaptively gather information in experimental design tasks. A new paper on arXiv studies how to amortize sequential Bayesian experimental design to address this gap.
The approach focuses on making the information-gathering process more efficient by amortizing the sequential decision-making steps. Rather than treating each experiment as a fresh inference problem, the method aims to reuse computations across related tasks.
The abstract does not report specific results, only the research direction. It suggests that amortization could help LLMs apply their existing knowledge more effectively when deciding what data to collect next.