Reproduce “Adapting Evidential Neural Networks to Test-Time Neighbor Fusion Improves Molecular Property Prediction” on OpenResearch

About this paper

Cameron Gruich, Weichi Yao, Yixin Wang, Bryan Goldsmith

PG-EVIKAL refines molecular-property predictions by combining an evidential model’s estimate with labels from relevant molecular neighbors, without updating the model. Its uncertainty estimates determine how much neighbor evidence can change a prediction, while property-guided selection identifies informative neighbors. Across 16 datasets, it reduces baseline RMSE on 14, with a median reduction of 19.4%; results are not universal.

Key insights

• Evidential uncertainty decomposition provides an active mechanism to leverage epistemic uncertainty as a resource for refining predictions at inference time, rather than merely a passive diagnostic.
• The quality and property relevance of selected neighbors are paramount for effective neighbor fusion, with property-guided re-ranking offering significantly amplified performance gains over purely structural similarity.
• Test-time neighbor fusion offers a computationally efficient, retraining-free alternative to active learning for continuous model refinement, which is especially valuable for deployed models or in workflows with scarce, sequentially acquired data.

Summary by alphaXiv.