On the Expressive Power of GNNs to Solve Linear SDPs
Paper • 2604.27786 • Published
VC-2-FMPNN checkpoints — reproduction of On the Expressive Power of GNNs to Solve Linear SDPs
Trained ppgn (VC-2-FMPNN) checkpoints from an independent reproduction of Qian & Morris
(ICML 2026, arXiv:2604.27786, OpenReview bjvLKXsTMK).
Both use the authors' unmodified GNN4SDP architecture
(hidden=64, num_conv_layers=8). Trained on CPU at reduced scale (GPU was unavailable).
| file | dataset | role | best loss |
|---|---|---|---|
ppgn_maxcut_best.pt |
synthetic Max-Cut SDP (n=30) | Claim 3 expressivity | test MSE 5.8e-3 |
ppgn_sdplib_best.pt |
SDPLIB MCP (n≤124) | Claim 4 overfit + Claim 5 warm-start | train loss 7.2e-4 |
Load with GNN4SDP.models.get_model(cfg) then load_state_dict. Full code, data pipeline, and
results are in the companion dataset
repro-gnn-linear-sdp-bundle.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support