Pairwise human guidance steers AI microscope without scoring every result
bravo_abad · x · 2026-10-12
Bulanadi et al. let scientists guide automated experiments by pairwise comparisons instead of numerical objectives: a neural network processes microscopy images while a pairwise Gaussian process learns a preference score. In a live experiment on a bismuth ferrite film, the system used a surface image plus measurements at 23 locations to locate regions deforming more strongly under voltage, with the learned preference map closely matching measured response.
More from Research
- Proposal: reward 'time to human grokking' in RL environments, especially math — PeterHndrsn · 2026-10-12
- OpenAI's unreleased model cracked most of a famous math list in one prompt, ~3 hours of thinking each — Puzzleheaded-King584 · 2026-10-12
- Decision models show embedding models are ripe for disruption as the Pareto frontier of model form factors stays unexplored — adhamelarabawy · 2026-10-12
- Rebuild the Lab for Robots: Manual Benches Are the Bottleneck in Bioengineering — anshulkundaje · 2026-10-12
- Sequencing Is Biology's Compute: Call for 1000x Cheaper Sequencing to Power Bio x AI — anshulkundaje · 2026-10-12
- Optimizer Sweep Shows Best Choice Varies With Batch Size; Shampoo Variants Lead — zacharynado · 2026-10-12