SHELDON harness models separate hypothesis and observation spaces for scientific agents
shi_weiyan · x · 2026-10-16
SHELDON is a scientific reasoning harness addressing a key flaw: AI agents usually start with a hypothesis and hunt for supporting observations, risking missed unexpected discoveries. It models separate hypothesis and observation spaces so agents actively explore both, generalizing across medical diagnosis, system-log investigation, ARC-AGI 2, and autoresearch.
More from Research
- Agentic sandbox WorldBox doubles GPT-6 Astra's MazeBench score, 23 to 61 gems — ZhitingHu · 2026-10-16
- U-Mich Kozik Lab case study: SeqHub untangles the understudied respiratory microbiome — anshulkundaje · 2026-10-16
- Arc Institute's Patrick Hsu on CRISPR, AI-designed organisms and betting on people — thesephist · 2026-10-16
- SHOES: 10 minutes of real-robot RL lets quadrupeds walk in shoes, CoRL 2026 paper — kscottz · 2026-10-16
- Post-trained NVIDIA Nemotron-3.5-Super-VL hits 35% on AutoResearchExam, matching a 20× larger model — AlexGDimakis · 2026-10-16
- U of T team mines 122,194 patient reviews across 45 Ontario hospitals with NLP to expose satisfaction and equity patterns — ishtiaqueSIA · 2026-10-16