Stanford SHELDON Lifts ORCA-Bench by 26.67 With Dual-Space Search

Stanford researchers Priyank Shethia and Monica S. Lam released SHELDON, a scientific-reasoning harness that treats hypothesis search and observation search as two spaces to explore jointly, and report a 26.67-point lift over the best baseline on ORCA-Bench: 66.67 against 40.00 for ReAct and Codex.

From the paper's abstract: "SHELDON is a scientific reasoning harness that separately represents and jointly explores hypothesis and observation spaces." Rather than fixing the evidence set and iterating on explanations, the system is designed around "actively seeking new observations alongside new hypotheses" to find data that current hypotheses cannot explain, which the authors argue drives genuinely new hypotheses rather than refinements of the same guess.

The gains show up across the test suite. On ARC-AGI-2, SHELDON scores 59.17, 11.67 points above EvoX at 47.50. On a QNN circuit-topology task scored out of 1.00, SHELDON lands a perfect 1.00 against EvoX's 0.85 and a benchmark reference topology of 0.917. On DiagnosisArena, a medical-diagnosis benchmark, it scores 84.00, six points above the next method.

The ablation is where the dual-space claim gets its support. Removing the observation space drops DiagnosisArena by 14.00 points and ORCA-Bench by 13.34, larger hits than removing comparative analysis or continued exploration. The authors also report that observations produced by the comparative-analysis step are "93-98% cited in subsequent decisions vs. 9-19% for candidate analysis," suggesting the observations SHELDON invents are the ones that actually move the search.

The paper is a single Stanford group's result and the retrieved abstract does not name the backbone model used across benchmarks. It drops into a dense recent run on agent benchmarks in our tracker; another appeared the same day in Tencent's LifeSearch comparison of twelve deep-research systems.