Amortizing Sequential Experimental Design in Reasoning LLMs

A research paper proposes amortizing sequential Bayesian experimental design (BED) into a large language model by explicitly training its reasoning, addressing the finding that LLMs struggle to adaptively gather information in experimental design tasks despite strong pretraining knowledge and reasoning capabilities. The work frames adaptive information gathering as a trainable reasoning skill rather than one that emerges from pretraining alone.

Despite their extensive pretraining knowledge and strong reasoning capabilities, large language models LLMs struggle to adaptively gather information in experimental design tasks. We study how to amortize sequential Bayesian experimental design BED into an LLM by explicitly training the reasonin