✗ Failed on benchmark 2026

Capacity-Preserving Transient Message Passing

Implementation & benchmark of arXiv:2607.25065 — Topology and dynamics of unimodular random hyperbolic manifolds

Usefulness7/10

Difficulty6/10

Novelty7/10

Dynamics Stability Accuracy Scalability Graph nn Ssm Regularization Dynamical systems Probability Graph theory Geometry Functional analysis

Source paper: Topology and dynamics of unimodular random hyperbolic manifolds arXiv:2607.25065 ⓘ · analyzed Aug 31, 2026

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 Not peer-reviewed.

Idea description

Replace ordinary graph convolution with a Markov diffusion whose edge transition probability is proportional to a geometric distance weight d(v,u)^(n−2), following the paper's Delaunay discretization theorem. Use graph Dirichlet capacity on expanding node shells to detect whether information escapes through the graph or repeatedly returns to local neighborhoods, and use this signal to select propagation depth or add a regularizer.

Formulas

p(v,u) = d(v,u)^(n−2) / sum over neighbors z of v of d(v,z)^(n−2)

L_w f(v) = sum over neighbors u of w(v,u) [f(v) − f(u)], with w(v,u) = d(v,u)^(n−2), and P = D^(-1)W

Cap_G(A,B) = infimum over f of (1/2) sum over ordered pairs (u,v) of w(u,v)[f(u) − f(v)]^2, subject to f=1 on A, f=0 outside B, and 0≤f≤1

H_(t+1) = sigma(((1−alpha)H_t + alpha P H_t Theta_t + b_t)); L_cap = (1/K) sum over k of max(0, tau − Cap_hat(A_k,B_k))^2

Mathematical statement

Let M be an n-dimensional hyperbolic manifold, let m be a good point configuration, and let G=(V,E) be its Delaunay graph. For an edge {v,u}, let d(v,u) be its hyperbolic distance and define w(v,u)=d(v,u)^(n−2). The paper uses the transition probability p(v,u)=w(v,u) divided by the sum of w(v,z) over neighbors z of v, and proves that M is transient if and only if this weighted random walk on G is transient. For neural propagation, define the weighted Laplacian Lf(v)=sum over neighbors u of w(v,u)(f(v)−f(u)), the weighted degree q(v)=sum over neighbors u of w(v,u), and the row-stochastic diffusion matrix P with entries P(v,u)=w(v,u)/q(v). For node sets A contained in B, define graph capacity as the minimum Dirichlet energy over functions f taking value 1 on A, value 0 outside B, and values in [0,1] elsewhere. Large capacity across expanding shells indicates an energetically available route for escape; capacity that vanishes rapidly with shell radius indicates recurrent, locally trapped propagation.

Implementation notes

Integrate this mechanism into a graph neural network layer wherever a normalized adjacency matrix is currently used. Assume that each node has a hyperbolic coordinate or another geometry from which edge distances can be computed. First construct a sparse neighborhood graph using a Delaunay approximation, k-nearest neighbors, or the existing graph topology. For every edge, compute d(v,u), set W[v,u] = d(v,u) raised to the power n−2, add a small epsilon to avoid zero degree, and row-normalize W to obtain P. Use the update H_next = activation((1−alpha)H + alpha times P H Theta + bias), where H is the node-feature matrix, Theta is a trainable weight matrix, and alpha is a residual diffusion strength. Pseudocode is: compute edge distances; assign W; normalize each row; for each layer compute message = P @ H; compute H = activation((1−alpha) * H + alpha * message @ Theta + bias). Every K layers, estimate capacity for shell pairs A_k and B_k by solving the constrained quadratic Dirichlet problem with conjugate gradients. A cheap approximation is to initialize f=1 on A_k and f=0 outside B_k, optimize the Dirichlet energy, and evaluate the final energy. The distance-weighted transition and capacity energy are taken from the paper; shell definitions, alpha, K, and threshold tau are engineering choices. Run first on a synthetic node-classification task generated from hyperbolic random graphs, comparing GCN, residual GCN, and this layer at equal parameter count. Also test a citation or molecular graph benchmark with available node coordinates. Measure validation accuracy, gradient norms, pairwise feature similarity, and the influence norm ||P^t e_v||_2 for multiple t. The quantitative prediction is that graphs with slowly decaying expanding-shell capacity will preserve long-range influence and delay oversmoothing, while graphs whose capacity decays rapidly will exhibit recurrent local mixing and little benefit from added depth. Sweep the hyperbolic curvature or global distance scale and look for a transition where shell-capacity decay changes regime; the transition should coincide with a sharp change in influence survival and the depth at which feature variance collapses.

Verification

Failed on benchmark

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Not tested

Stage 2 · Practical benchmark result: No measurable effect

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

Implemented distance-weighted Markov diffusion, residual message passing utilities, exact shell Dirichlet-capacity solves, and a synthetic comparison in run_experiment.py. The mechanism manifested: 1D capacity fit was -1.000 versus the predicted -1, 2D capacity was strongly affine in 1/log(R) (R²=0.985), 3D capacity remained nonzero (late/early=0.878), and influence norm decayed with exponent -0.249 versus the predicted -0.25. The weighted layer gave a small fixed-readout accuracy improvement (0.911 vs 0.906), but this is not evidence of a robust trained-GNN win.

Agent confidence: 8/10

Baseline: Unweighted residual diffusion: best accuracy 0.9056 at depth 7; depth-20 accuracy 0.9000; variance 0.4231.
Idea: Distance-weighted residual diffusion: best accuracy 0.9111 at depth 7; depth-20 accuracy 0.9000; variance 0.4536. Math checks: 1D exponent -1.000, 2D capacity-vs-1/log(R) R² 0.9845, 3D late/early capacity 0.8776, influence exponent -0.2486.

Limitations: The task comparison uses a fixed coordinate readout rather than training matched-parameter GNNs. Capacity was evaluated on regular lattice shells with unit edge distances, not hyperbolic Delaunay graphs; no citation or molecular benchmark, GPU/CPU speed study, learned capacity regularizer, curvature sweep, or statistical repeats were run.

How to run: python3 run_experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.25065) Aug 31, 2026

Ahead of the baseline (-0.0137 metric points) but not statistically significant (p=0.0868, wins 6 / 8 paired seeds).

Benchmark: Geometric_graph_diffusion
Model: GeometricMessageNet
Paired seeds: 8
Baseline mean: 0
Idea mean: 0
Effect (Δ): -0.0137 (n/a; negative = idea better)
Wins: 6 / 8 paired seeds
p-value: 0.0868 (permutation test, 20 000 shuffles)
Smallest detectable effect: n/a
Mechanism: n/a
Practical effect: No effect
Baseline tuning: swept over 0 configs

Limitations:

The custom task uses small fixed 24-node kNN geometric graphs rather than a real Delaunay or hyperbolic benchmark. Capacity regularization, shell-capacity estimation during training, curvature sweeps, and external citation or molecular graph datasets were not tested.

How to run: python3 bench_graph.py

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

Stage 2 — Benchmark Aug 31, 2026 latest Failed ✗

Agent confidence: 9/10

Built a custom geometric graph regression track and matched residual message-passing systems, using d^2 edge weights for the idea versus unweighted edges for the baseline. Across 8 paired seeds, the idea reduced MSE from 0.0848644 to 0.0712046 with delta -0.0136598, but the permutation p-value was 0.0868, so it was not a significant win. The trained-model mechanism signature was not confirmed: weighted propagation had lower observed gradient influence and lower predicted-observed correlation.

Baseline: Best lr=0.003, alpha=0.5; full 8-seed MSE 0.0848644 ± 0.0295812.
Idea: Best lr=0.001, alpha=0.5; full 8-seed MSE 0.0712046 ± 0.0237890; 6/8 paired wins; delta=-0.0136598; permutation p=0.0868.
Mechanism signature: Baseline observed gradient influence L2=0.267737, predicted-observed correlation=0.878935; idea observed gradient influence L2=0.244880, correlation=0.384099; confirmed=false.

Benchmark result