△ Mechanism confirmed, baseline not beaten 2026

Hypoelliptic transport-diffusion layer

Implementation & benchmark of arXiv:2608.29813 — Boundary Harnack inequalities for Kolmogorov equations in asymptotically cylindrical Lipschitz domains

Usefulness7/10

Difficulty5/10

Novelty7/10

Architecture Accuracy Stability Sample efficiency World model Ssm Attention Pde Geometry Dynamical systems Harmonic analysis

Source paper: Boundary Harnack inequalities for Kolmogorov equations in asymptotically cylindrical Lipschitz domains arXiv:2608.29813 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Replace an isotropic local mixing layer with a kinetic layer that smooths features in x and transports them in y along the characteristic direction x. The layer should be useful for phase-space data, learned simulators, and world models in which positions or transported quantities evolve through coupled drift and diffusion rather than independent Euclidean motion.

Formulas

$$\mathcal{K}u=\Delta_xu+x\cdot\nabla_yu-\partial_tu,$$

$$\delta_r(x,y,t)=(rx,r^3y,r^2t),$$

$$D_hf(x,y)=\mathbb{E}_{\xi\sim\mathcal{N}(0,2I_m)}\left[f\left(x+\sqrt{h}\,\xi,y\right)\right]\approx e^{h\Delta_x}f(x,y),$$

$$f_{k+1}(x,y)=f_k(x,y)+g_k(x,y)\odot\left[D_hf_k\left(x,y+h x\right)-f_k(x,y)\right],\qquad 0\leq g_k\leq 1.$$

Mathematical statement

The paper studies the Kolmogorov operator $\mathcal{K}=\Delta_x+x\cdot\nabla_y-\partial_t$, where $x\in\mathbb{R}^m$ is the diffusive variable, $y\in\mathbb{R}^m$ is the higher-order transported variable, $t$ is time, $\Delta_x=\sum_{i=1}^m\partial_{x_i}^2$, and $x\cdot\nabla_y=\sum_{i=1}^m x_i\partial_{y_i}$. Its intrinsic dilation is $\delta_r(x,y,t)=(rx,r^3y,r^2t)$, so x has homogeneous degree 1, t degree 2, and y degree 3. The proposed layer uses the two constructive pieces of $\mathcal{K}$: diffusion in x, approximated by $e^{h\Delta_x}$, and characteristic transport in y, generated by $x\cdot\nabla_y$. Here $h>0$ is a layer step size, $D_h$ is an x-only diffusion operator, $S_hf(x,y)=f(x,y+hx)$ is transport, and $g_k$ is a learned gate constrained to $[0,1]$. The residual update is a stable interpolation between the current feature and the transported-diffused feature.

Implementation notes

Integrate the layer as a replacement for one local convolution, MLP-mixer block, or state-space transition in a model whose features are indexed by phase-space coordinates (x,y), with optional time conditioning. Let the input be a tensor f of shape [batch, spatial-x locations, spatial-y locations, channels], together with coordinate tensors x and y. First choose a positive step h. Implement D_h as a depthwise separable convolution over x axes only, using a normalized Gaussian kernel with variance 2h; on irregular coordinates, use K nearest x-neighbors with weights w_ij proportional to exp(-||x_i-x_j||^2/(4h)), normalized over j. Next implement characteristic transport by evaluating the diffused feature at y_i+h*x_i using bilinear interpolation or a differentiable grid sampler. Compute delta=f_shift-f, and update f_new=f+sigmoid(G(f))*delta, where G is a pointwise linear layer and the sigmoid gate supplies g in the formula. Optionally follow this with a pointwise channel-mixing MLP and LayerNorm. For a multiscale stack, use x receptive-field radius r_l, y radius r_l^3, and time or layer step proportional to r_l^2, directly matching delta_r. Pseudocode is: f_diff=DiffusionX(f,h); f_shift=InterpolateY(f_diff,y+h*x); delta=f_shift-f; f=f+sigmoid(G(f))*delta. The paper-derived quantities are the Kolmogorov operator, anisotropic dilation, Gaussian x diffusion, and x-directed y transport; h, kernel truncation radius, channel width, and gate initialization are empirical hyperparameters. The first experiment should use a small damped-particle or Lorenz-style phase-space prediction dataset. Compare a six-layer MLP or Transformer local mixer against an equal-parameter, equal-FLOP model replacing two mixers with kinetic layers. Measure one-step MSE, long-horizon rollout error, gradient norm variance, and performance as training-set size decreases. Success means lower rollout error and fewer unstable trajectories, especially on systems where y evolves through x. Include ablations removing the transport term, using isotropic diffusion, and replacing h*x by a learned unconstrained displacement.

Verification

Mechanism confirmed, baseline not beaten

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

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 a working hypoelliptic transport-diffusion layer with x-only Gaussian diffusion, characteristic y transport, and sigmoid-gated residual interpolation. The mechanism checks confirmed Fourier attenuation within 6.2%, linear displacement scaling with h exactly, and nonexpansive sampled gated updates. The idea improved one-step validation MSE (0.00219 vs 0.00890) with fewer parameters, although its five-step rollout error was worse, so broad long-horizon superiority was not demonstrated.

Agent confidence: 8/10

Baseline: 44 parameters; one-step validation MSE 0.008903; five-step rollout MSE 0.005904
Idea: 12 parameters; one-step validation MSE 0.002194; five-step rollout MSE 0.010434; diffusion prediction max relative error 6.16%; transport displacement/h 1.15498 for all tested h

Limitations: The CUDA convolution backend failed and the experiment used CPU fallback. The toy task used smooth random phase-space fields rather than Lorenz or damped-particle trajectories; the comparison was not equal-parameter or rigorously equal-FLOP, and irregular coordinates, multiscale layers, systematic ablations, gradient variance, and larger long-horizon studies were not tested.

How to run: python3 experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.29813) Sep 2, 2026

Significantly WORSE than the tuned baseline (+29.7%, p=0.0496, wins 1 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small_kinetic_vs_raw
Paired seeds: 8
Baseline mean: 0.001
Idea mean: 0.0013
Effect (Δ): 0.0003 (+29.7%; negative = idea better)
Wins: 1 / 8 paired seeds
p-value: 0.0496 (permutation test, 20 000 shuffles)
Smallest detectable effect: n/a
Mechanism: Confirmed ✓
Practical effect: Hurts
Baseline tuning: swept over 3 configs

Limitations:

Only the built-in dynamics track was tested; long-horizon rollout error, irregular-coordinate diffusion, multiscale stacks, and additional tracks were not tested. The transport mechanism signature was confounded by the learned gate and channel mixing and should not be treated as a clean independent confirmation.

How to run: python3 stage2_bench.py

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

Stage 2 — Benchmark Sep 2, 2026 latest Failed ✗

Agent confidence: 9/10

Implemented a phase-space kinetic mixing layer and evaluated it against a raw-token GRU using the canonical trainer, parity-matched learning-rate sweep, and eight paired seeds. The kinetic model was significantly worse: test MSE 0.00132353 versus 0.00102014 for baseline, paired delta +0.00030339, permutation p=0.04965. The diffusion attenuation check was close, but the transport displacement measurement was confounded by learned gating and channel mixing; regardless, the trained benchmark metric is a clear loss.

Baseline: dynamics raw-token GRU; best lr=0.006; 8-seed test MSE mean=0.00102014, std=0.00017824
Idea: dynamics kinetic GRU; h=0.08, best lr=0.006; 8-seed test MSE mean=0.00132353, std=0.00057783
Idea wins: 1
N pairs: 8
Paired delta: 0.0003033862 (idea minus baseline; positive is worse)
Parameter counts: baseline 3585, idea 3609
Permutation p value: 0.04965

Benchmark result