From Point Predictions to Hypothesis Sets: MH-GraphLENS-JEPA for Identifiability-Aware Latent Mechanism Characterization

This paper introduces MH-GraphLENS-JEPA, a multi-hypothesis extension of GraphLENS-JEPA for identifiability-aware latent mechanism characterization. The architecture combines relational graph encoding, learned hypothesis slots, permutation-invariant set matching, constrained categorical decoding, and a joint-embedding predictive objective. In the solver-labeled BioWorld benchmark, the JEPA variant improved mean held-out hypothesis F1 from 0.731 to 0.778 and exact-set accuracy from 0.548 to 0.638 across three matched seeds. The manuscript emphasizes that the multi-hypothesis relational formulation is the strongest result, while the incremental JEPA advantage is promising but preliminary. The record includes the 13-page paper and the v0.3.0 source archive with raw results and reproduction code.