Post
351
Our new paper Moose-chem 3 introduces experiment-guided hypothesis ranking, a novel setting where candidate hypotheses are prioritized based on experimental feedback from previously tested hypotheses.
To support research in this area, the work proposes a simulator grounded in three domain-informed assumptions that can generate simulated experimental feedback without requiring costly real-world trials.
The simulator is validated on a curated dataset of 124 chemistry hypotheses, and the resulting method outperforms strong pre-experiment baselines. This enables scalable research on feedback-driven hypothesis discovery strategies in scientific domains where empirical validation is expensive or slow.
MOOSE-Chem3: Toward Experiment-Guided Hypothesis Ranking via Simulated Experimental Feedback (2505.17873)
To support research in this area, the work proposes a simulator grounded in three domain-informed assumptions that can generate simulated experimental feedback without requiring costly real-world trials.
The simulator is validated on a curated dataset of 124 chemistry hypotheses, and the resulting method outperforms strong pre-experiment baselines. This enables scalable research on feedback-driven hypothesis discovery strategies in scientific domains where empirical validation is expensive or slow.
MOOSE-Chem3: Toward Experiment-Guided Hypothesis Ranking via Simulated Experimental Feedback (2505.17873)
