✗ Failed on benchmark 2026

Power-Preserving Formation GNN

Implementation & benchmark of arXiv:2609.04158 — Formation Matrix and Energy-based Control of Multi-Agent Systems

Usefulness8/10

Difficulty5/10

Novelty6/10

Dynamics Stability Accuracy Graph nn Training dynamics Optimizer Control theory Dynamical systems Graph theory Optimization Linear algebra

Source paper: Formation Matrix and Energy-based Control of Multi-Agent Systems arXiv:2609.04158 ⓘ · analyzed Sep 4, 2026

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 Not peer-reviewed.

Idea description

Replace an unconstrained graph-message-passing block with a port-Hamiltonian layer whose edge interactions are generated by a skew-symmetric formation-matrix coupling and whose node damping is positive semidefinite. The layer can model relative graph structure while preventing unforced hidden-state energy growth, reducing exploding activations and oversmoothing caused by arbitrary repeated propagation.

Formulas

$$\begin{bmatrix}\dot r\\\dot f\end{bmatrix}=\begin{bmatrix}0&-\mathcal{F}(r)^T\\\mathcal{F}(r)&0\end{bmatrix}\begin{bmatrix}\dot r\\f\end{bmatrix}+\begin{bmatrix}\nu\\0\end{bmatrix},\qquad \zeta=\mathcal{F}(r)\dot r$$

$$\zeta^Tf+\dot r^T\dot p=0$$

$$\dot h=[J(h)-R(h)]\nabla H(h),\quad J^T=-J,\quad R\succeq0,\qquad \dot H=-\nabla H(h)^TR(h)\nabla H(h)\leq0$$

$$h_{k+1}=h_k+\Delta t\,[J(h_k)-R(h_k)]\nabla H(h_k),\qquad \Delta t<\frac{2}{\lambda_{\max}(M^{-1/2}RM^{-1/2})}$$

Mathematical statement

For node state $r\in\mathbb{R}^{2n}$ and edge force vector $f\in\mathbb{R}^{m}$, the paper introduces a formation matrix $\mathcal{F}(r)$ that maps node velocities to edge velocities, $\zeta=\mathcal{F}(r)\dot r$. Its interconnection subsystem uses the skew-symmetric matrix $\begin{bmatrix}0&-\mathcal{F}(r)^T\\\mathcal{F}(r)&0\end{bmatrix}$, implying the power identity $\zeta^Tf+\dot r^T\dot p=0$. A port-Hamiltonian neural layer should use $\dot h=[J(h)-R(h)]\nabla H(h)$, where $h$ is the concatenated node hidden state, $J(h)^T=-J(h)$ is the graph interconnection, $R(h)\succeq0$ is damping, and $H(h)\geq0$ is the learned energy. Then $\dot H=-\nabla H^TR\nabla H\leq0$ in the absence of external input. For a quadratic energy $H(h)=\frac12h^TM^{-1}h$, $M\succ0$, choosing $J$ from edge incidence or formation operators and $R=L_RL_R^T$ gives a computable stability certificate.

Implementation notes

Integrate this as a replacement for one or more residual message-passing blocks in a graph neural network. Let each node carry hidden state $h_i\in\mathbb{R}^d$ and let each graph edge $(i,j)$ carry an edge state or force $f_{ij}$. Build a learned formation operator $F_\theta(h)$ from relative node features, with one row per edge; its action can approximate edge differences, for example $F_\theta h=\{W_e(h_i-h_j)\}_e$. Construct the node-edge interconnection explicitly as $J_F=\begin{bmatrix}0&-F_\theta^T\\F_\theta&0\end{bmatrix}$, which is skew-symmetric by construction. Use energy $H_\theta(h)=\frac12h^TM_\theta^{-1}h+\sum_e\psi_{e,\theta}(f_e)$ with diagonal positive masses parameterized by softplus, and parameterize damping as $R_\theta=L_\theta L_\theta^T$ rather than learning an unrestricted matrix. The update is: compute $F$ from edges; compute $g=\nabla H$; compute $v=(J_F-R)g$; update $h\leftarrow h+\Delta t v$. Estimate $\lambda_{\max}$ by 5-10 power iterations and clip or adapt $\Delta t$ below $2/\lambda_{\max}(M^{-1/2}RM^{-1/2})$. The paper supplies the power-preserving interconnection and energy argument; the neural energy, edge feature map, and damping factors are learned. First test on Cora or a synthetic spring-graph node-classification task with 20-50 repeated layers, comparing an unconstrained GNN, an antisymmetric GNN, and this layer. Measure hidden energy $H_k$, gradient norms, accuracy, and failure rate while sweeping $\Delta t$. The prediction is that the unforced layer has non-increasing energy up to numerical error, and instability should appear near the predicted explicit-Euler boundary $\Delta t_c\approx2/\lambda_{\max}(M^{-1/2}RM^{-1/2})$; measured divergence should occur within 20% of this value, while the skew-only variant should show bounded but non-decaying energy.

Verification

Failed on benchmark

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Not reproduced ✗

Stage 2 · Practical benchmark result: Significantly worse ✗

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 an incidence/formation-matrix port-Hamiltonian node-edge layer with explicit skew-symmetric interconnection and positive semidefinite damping. Numerical checks confirmed exact skewness and power cancellation; the continuous energy derivative matched finite differences and was negative, while 50 repeated stable updates reduced total energy and avoided the unconstrained block’s large norm growth. However, the damping-only Euler bound was not sufficient under strong skew coupling: the stated bound was 10.0, while the measured linear Euler limit was about 0.099, so timestep control must account for the full J-R spectrum.

Agent confidence: 8/10

Baseline: Repeated unconstrained graph propagation: node norm 4.094 -> 4223.000 after 50 steps.
Idea: Port-Hamiltonian propagation: total energy 27.475 -> 22.828 after 50 steps; maximum energy remained 27.475; node norm 4.094 -> 4.934. Skew error and interconnection power were 0; derivative finite-difference mismatch was 2.8e-6. Actual Euler limit was 0.099 versus stated damping-only bound 10.0.

Limitations: This was a synthetic linear NumPy toy experiment, not Cora or a trained node-classification GNN. Learned nonlinear energies, adaptive timestep estimation, accuracy, FLOPs, GPU behavior, and generalization across graph families were not tested; the unconstrained baseline intentionally included a gain to expose activation growth rather than representing a tuned modern GNN.

How to run: python3 power_formation_gnn.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2609.04158) Sep 4, 2026

Significantly WORSE than the tuned baseline (+856.9%, p=0.0081, wins 0 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0003
Idea mean: 0.0027
Effect (Δ): 0.0024 (+856.9%; negative = idea better)
Wins: 0 / 8 paired seeds
p-value: 0.0081 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±177.4%
Mechanism: Not confirmed ✗
Practical effect: Hurts
Baseline tuning: swept over 3 configs

Limitations:

Only the registered dynamics track was tested. The mechanism signature used the trained forced recurrence, not a separate unforced rollout or a full Euler spectral-bound sweep. Other built-in tracks were not evaluated.

How to run: python3 bench_power_formation.py

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

Stage 2 — Benchmark Sep 4, 2026 latest Failed ✗

Agent confidence: 10/10

Implemented the power-preserving formation mechanism as a port-Hamiltonian recurrent replacement with explicit skew coupling and softplus PSD damping. On the structurally matched registered dynamics track, the tuned baseline achieved mean test MSE 0.00028120 while the best idea setting achieved 0.00269064; paired delta was +0.00240944 with permutation p=0.0081, indicating a significant loss. The trained-model signature also failed to confirm non-increasing energy under the forced benchmark recurrence.

Baseline: rnn_small dynamics baseline; lr sweep {0.001, 0.003, 0.009}; best lr=0.009; full 8-seed mean MSE 0.0002811976373777725.
Idea: Port-Hamiltonian recurrent system; idea sweep best lr=0.009; full 8-seed mean MSE 0.002690640962100588; idea wins 0/8.
Paired delta: 0.0024094433247228153 (idea minus baseline; lower is better)
Permutation p value: 0.0081

Benchmark result