A research framework called JADAI rethinks how machines run scientific experiments — by learning experiment strategy and result interpretation at the same time.
Most adaptive experiment systems treat design and inference as separate problems: first pick the next experiment, then update your beliefs about the underlying parameters. JADAI collapses that pipeline. It trains a policy network, a history network, and an inference network end-to-end, minimizing a single loss that accumulates reductions in posterior error across an entire experimental sequence. The inference side uses diffusion-based estimators, which means it can handle high-dimensional and multimodal posteriors — situations where a simple bell curve would miss the real distribution.
The practical upshot is that the system gets better at choosing experiments precisely because it is simultaneously getting better at reading results, and vice versa. On standard adaptive design benchmarks, JADAI matches or beats existing approaches. That matters for any field — drug discovery, materials science, physics — where experiments are expensive and you want each one to do maximum informational work.
Amortized Bayesian inference is not new, and neither is adaptive experimental design. What JADAI adds is the joint training signal, which is a meaningful architectural choice but still an academic result pending real-world validation outside benchmark conditions.