HPEC 2026 — do multidimensional architectures earn their compute?
Trained checkpoints for a controlled benchmark on a US Census port lattice: 30,087 series over 192 months, shaped State (14) x HS6 commodity (1343) x Flow (2), 9 channels.
Every run here is one cell of a declared matrix. A run is only present if it finished with a completion record whose SHA-256 matches the checkpoint bytes and whose provenance fingerprint matches the manifest it was declared under.
What the directories are
| prefix | experiment |
|---|---|
exp0_* |
learning-rate search, 5 rates over two decades per arm |
exp0_mech_* |
the fa mixer's own rate search — see below |
*_aggregate_s* |
Test 1, fixed aggregate split |
*_aggregate_roll_* |
Test 1.1, rolling origin |
*_asa_*, *_fa_*, *_grid_* |
Tests 2/3/6, N-D grid arms |
*_id_* |
Test 4, identity-aware axial |
manifest.json is the declared matrix. Each run carries best.pth,
logs/metrics.json (per-epoch train/val loss, lr, wall time, gradient norms),
cost.json (measured FLOPs), and its training log.
The fa learning rate is not the N-D learning rate
The factorized-attention arm was originally given the rate the axial probe selected. Probing it directly showed every backbone prefers a different rate, by up to 33x:
| backbone | axial-probe rate | fa's own rate |
|---|---|---|
| gru | 3e-4 | 1e-2 |
| lstm | 1e-4 | 1e-3 |
| mamba2 | 3e-3 | 1e-4 |
| mamba3 | 1e-4 | 1e-3 |
| transformer | 1e-3 | 3e-4 |
At 3-D, mamba3 reaches 0.634 at its own rate and 1.024 at the inherited one — which is worse than the axial arm it was being compared against, and would have reversed that cell's conclusion.
Status
Snapshot taken while the sweep is still running. Tests 1 and 1.1 are complete;
the N-D grid arms are partial. Counts per experiment are in manifest.json.