✓✓ Beats tuned baseline 2026

Fermionic circuit message passing

Implementation & benchmark of arXiv:2607.27198 — Mixed partition functions are exactly the graph parameters of exponentially bounded edge-connection rank

Usefulness7/10

Difficulty6/10

Novelty8/10

Architecture Accuracy Generalization Scalability Graph nn Embedding Architecture Algebra Tensor decomposition Combinatorics Representation theory

Source paper: Mixed partition functions are exactly the graph parameters of exponentially bounded edge-connection rank arXiv:2607.27198 ⓘ · analyzed Aug 31, 2026

AI-generated research hypothesis, automatically tested. Not peer-reviewed.

Idea description

Augment every graph-neural-network edge message with an even commuting channel and a low-dimensional odd anticommuting channel. Contracting odd channels around an edge circuit gives a sign determined by the number of odd edges, while local states with odd incident degree are forced to vanish; this supplies a built-in parity and cycle constraint that ordinary GNNs must learn implicitly.

Formulas

$$V=V_0\oplus V_1,\qquad h^{\mathrm{RS}}=h^{\mathrm{cat}}\circ J,$$

$$J\big|_{\operatorname{Sym}V_0\otimes\Lambda^jV_1}=(-1)^{j(j-1)/2}\,\operatorname{id},$$

$$\prod_v(-1)^{j_v/2}=(-1)^{\frac12\sum_v j_v}=(-1)^{|F|},\qquad j_v=\deg_F(v),$$

$$\left(\bigotimes_v\varpi_{\deg_F v}\right)\circ\mathrm{Regroup}_{F,\kappa}(\widehat C_1^{\otimes |F|})=(-1)^{c(\kappa)}\sum_{\varphi:F\to[2\ell]}\bigotimes_v\bigwedge_{(a_1,a_2)\in\kappa_v}\xi_{\varphi(a_1)}\wedge\eta_{\varphi(a_2)}.$$

Mathematical statement

The construction uses a super vector space V=V_0\oplus V_1, where V_0 is the even sector and V_1 is the odd sector. The paper defines an involution J on \operatorname{Sym}V_0\otimes\Lambda V_1 by J=(-1)^{j(j-1)/2}\operatorname{id} on exterior degree j, and transforms a local tensor by h^{\mathrm{RS}}=h^{\mathrm{cat}}\circ J. Here j is the number of odd incident half-edges and h^{\mathrm{cat}} is the ordinary categorical contraction. Odd local degrees vanish, so only even j contribute. For an odd-edge subset F, parity transfer contributes \prod_v(-1)^{j_v/2}=(-1)^{|F|}, where j_v=\deg_F(v) and |F| is the number of selected odd edges. The explicit circuit formula uses \widehat C_1=-\sum_i\xi_i\otimes\eta_i, odd basis vectors \xi_i,\eta_i, exterior contraction \varpi_j, and c(\kappa), the number of circuits in the pairing \kappa.

Implementation notes

Integrate this into a message-passing GNN immediately after edge message construction and before aggregation. For each directed edge (u,v), produce an even vector e_uv in R^k and an odd vector o_uv in R^{2ell}; use separate linear maps, with ell=2 or 4 initially. Maintain the odd state as an exterior-algebra coefficient tensor rather than an unrestricted vector. A practical low-order implementation stores only degree-0 and degree-2 coefficients: degree-0 is a scalar and degree-2 is a skew-symmetric matrix A in R^{2ell x 2ell}, representing the exterior sum \sum_{i<j}A_ij\xi_i\wedge\xi_j. At each node, collect incident odd messages and compute degree-2 coefficients by skew-symmetrized pair products; discard configurations with odd incident degree, implementing the paper's vanishing rule. For a selected odd-edge mask F, multiply the contribution by (-1)^{|F|}; if explicit circuit regrouping is used, multiply each circuit contribution by (-1)^{c(\kappa)}. Pseudocode: m_e=MLP_e(h_u,h_v); m_o=MLP_o(h_u,h_v); e_v=sum_u m_e; A_v=sum_{u<w}(m_o[u]m_o[w]^T-m_o[w]m_o[u]^T); h_v=MLP([h_v,e_v,vec(A_v)]). The paper supplies the parity algebra and sign rule; neural maps and truncation degree are empirical. First test on cycle-versus-tree classification and molecular graph classification against a 3-layer GIN with matched width and FLOPs. Success means higher accuracy on cycle-parity tasks, better extrapolation to longer cycles, and no degradation as message-passing depth increases.

Verification

Beats tuned baseline

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

Stage 2 · Practical benchmark result: Significant win ✓

Methodology: 8 paired seeds, permutation test (p<0.05) against a learning-rate-tuned baseline. How verification works

Stage 1 — Mechanism check agent confidence 8/10

Built a degree-2 truncated fermionic message-passing GNN with separate even and odd edge channels, skew-symmetric wedge aggregation, and odd-degree vanishing. The quantitative checks confirmed zero contribution for odd local degree, exact parity transfer across cycle masks, and the predicted wedge-energy scaling 2d(d-1) (observed/predicted ratios 0.984–1.003). On the tiny cycle-versus-path extrapolation task, baseline GIN scored 0.50 and the fermionic model scored 1.00, so the mechanism manifested and also produced a promising toy-task gain; this is not evidence of broad superiority.

Agent confidence: 8/10

Baseline: GIN: 0.50 accuracy on 10 longer test graphs (nodes 11-15), trained on nodes 6-10
Idea: Fermionic degree-2 odd-channel GNN: 1.00 accuracy on the same test set; odd-degree residual 0.0; parity max error 0.0 for cycle sizes 1-8; wedge-energy ratios observed/predicted: 1.003, 0.984, 0.998, 0.992 for dimensions 2,4,8,12

Limitations: Only a synthetic cycle-versus-path task was tested, with one seed and a very small dataset; no molecular benchmarks, multiple seeds, FLOP-matched timing, explicit circuit regrouping, higher exterior degrees, or CUDA comparison was evaluated. The parity sweep uses closed cycle masks and therefore has only two surviving masks per cycle size.

How to run: python3 fast_mvp.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.27198) Sep 2, 2026

Beats the tuned baseline by 100.0% (p=0.0312, wins 6 / 8 paired seeds; metric: lower is better).

Benchmark: Cycle_parity_graph
Model: graphnet
Paired seeds: 8
Baseline mean: 0.2087
Idea mean: 0
Effect (Δ): -0.2087 (−100.0%; negative = idea better)
Wins: 6 / 8 paired seeds
p-value: 0.0312 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±88.7%
Mechanism: Confirmed ✓
Practical effect: Helps
Baseline tuning: swept over 3 configs

Limitations:

Only the custom cycle-versus-path graph classification task was tested; no molecular, social, or irregular real-world graph benchmark was evaluated. The graph size was fixed at 12 nodes, higher exterior degrees and explicit circuit regrouping were not tested, and FLOP/latency comparisons were not measured.

How to run: python3 bench_fermionic.py

Verdict computed by deterministic test code from paired-seed statistics — not by the language model.

Stage 2 — Benchmark Aug 31, 2026 latest Worked ✓

Agent confidence: 8/10

Implemented a custom graph track because no built-in track contains graph topology. The degree-2 fermionic GNN achieved mean test error 0.0 across 8 paired seeds versus 0.20875 for the swept baseline, with paired delta -0.20875 and permutation p=0.0312, satisfying the benchmark significant-win criterion. The trained-model signature confirmed exact zero wedge output for odd incident degrees; this is evidence on the custom cycle/path task, not broad graph-learning superiority.

Baseline: GraphNet baseline, best lr=0.003 and 12 epochs: mean test err 0.20875, std 0.20704; per-seed [0.0, 0.29, 0.5, 0.015, 0.5, 0.325, 0.0, 0.04]. Sweep means: lr 0.001=0.30750, lr 0.003=0.20125, lr 0.006=0.21625.
Idea: Degree-2 fermionic GraphNet, lr=0.003 and 12 epochs: mean test err 0.0, std 0.0; all 8 per-seed errors were 0.0. Paired delta=-0.20875, 6 wins and 2 ties, permutation p=0.0312.

Benchmark result