One Proposal for Every Margin: Zero-Shot Amortized Sequential Importance Sampling for Binary Matrices
MarginFlow learns a single SIS proposal that generalizes across binary-matrix margins.
The paper frames the ideal sequential importance sampling proposal as the policy of a GFlowNet with unit reward on valid matrices. Its model uses a set transformer over remaining margins, treating each partial matrix as a smaller margin problem. Trained on 1,904 margins, it was tested zero-shot on 1,190 held-out synthetic and real cases. It matched or beat the best post-hoc analytical configuration on 1,187 of them, with a median effective sample fraction of 99.8%. HF Daily Papers' note
The paper frames the ideal sequential importance sampling proposal as the policy of a GFlowNet with unit reward on valid matrices. Its model uses a set transformer over remaining margins, treating each partial matrix as a smaller margin problem. Trained on 1,904 margins, it was tested zero-shot on 1,190 held-out synthetic and real cases. It matched or beat the best post-hoc analytical configuration on 1,187 of them, with a median effective sample fraction of 99.8%. HF Daily Papers' note
score 4