✓✓ Beats tuned baseline 2026

Dephasing-Controlled Transport Layer

Implementation & benchmark of arXiv:2607.22240 — Fermions on a 1D lattice: localized sources and sinks with dephasing

Usefulness8/10

Difficulty7/10

Novelty7/10

Dynamics Stability Accuracy Scalability Ssm Rnn Graph nn Training dynamics Dynamical systems Pde Spectral theory Linear algebra

Source paper: Fermions on a 1D lattice: localized sources and sinks with dephasing arXiv:2607.22240 ⓘ · analyzed Aug 30, 2026

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 Not peer-reviewed.

Idea description

Replace repeatedly applied unconstrained message passing or recurrent transition maps with a transport layer containing a coherent hopping branch and an explicit dephasing operator. Small dephasing preserves sharp, oscillatory propagation, whereas large dephasing suppresses inter-position correlations and produces stable diffusion-like receptive-field growth, which should reduce long-horizon ringing and exploding sensitivities.

Formulas

$$H=-J A,\qquad A_{ij}=\mathbf{1}_{|i-j|=1}$$

$$\frac{dC}{dt}=-\mathrm{i}[H,C]-\gamma\bigl(C-\operatorname{diag}(C)\bigr)+S(C),\qquad n_i=C_{ii}$$

$$\frac{dn_i}{dt}\simeq D_\gamma(n_{i+1}-2n_i+n_{i-1}),\qquad D_\gamma\simeq\frac{2J^2}{\gamma}\quad(\gamma\gg J)$$

$$\widehat n_k(t)\simeq\widehat n_k(0)\exp[-D_\gamma k^2t],\qquad C_{ij}^{\mathrm{off}}(t)\propto e^{-\gamma t}$$

Mathematical statement

The paper uses the nearest-neighbor tight-binding Hamiltonian H=-J A, where A is the one-dimensional adjacency matrix and J is the hopping amplitude. Represent a hidden-state correlation matrix C, with C_ij measuring interaction or coherence between positions i and j. Use the Lindblad-inspired evolution dC/dt=-i[H,C]-gamma(C-diag(C))+S(C), where gamma is the dephasing rate and S contains optional source or sink terms. The propagated neural feature is n_i=C_ii. In the strong-dephasing limit, adiabatic elimination of nearest-neighbor off-diagonal correlations gives dn_i/dt approximately equal to D_gamma(n_(i+1)-2n_i+n_(i-1)), with D_gamma approximately 2J^2/gamma in lattice units. Therefore off-diagonal correlations decay at rate gamma, while low-frequency diagonal modes decay diffusively at rate D_gamma k^2.

Implementation notes

Integrate this mechanism at the propagation step of a one-dimensional sequence model, graph neural network, or state-space model. Use a hidden feature tensor n of shape L by d and maintain either an explicit small correlation tensor C of shape L by L by d for an MVP or a low-rank factorization for larger sequences. Initialize C_ii from the hidden features and initialize local off-diagonal entries from neighboring feature inner products. For each propagation step, compute the nearest-neighbor commutator using left and right shifts, damp off-diagonal entries by exp(-gamma times dt), and read out the diagonal as the propagated state. Pseudocode is: C = C - dt*(1j*(H@C-C@H) + gamma*(C-diag(C))); n = real(diag(C)); h = MLP(n). The paper supplies the hopping Hamiltonian, local dephasing term, and diffusion-limit equation; the nonlinear readout and learnable parameters are adaptations. The first cheap experiment should use a small SSM or RNN on Copy, Selective-Copy, and long-range adding tasks, comparing a vanilla recurrent transition, a diffusion convolution, and this layer. Sweep gamma/J from 0.01 to 100. Measure Fourier-mode decay, gradient norms, and loss. The prediction is that oscillatory propagation changes to exponential damping near gamma comparable to J, and for gamma at least 5J the measured diffusion coefficient should satisfy D_gamma gamma divided by 2J squared within 20 percent of one.

Verification

Beats tuned baseline

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Not reproduced ✗

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

How to read this: The implementation beat the baseline, but the claimed mechanism did not reproduce at benchmark scale. The gain may come from another effect; treat the cause as unresolved.

Stage 1 — Mechanism check agent confidence 9/10

Built a self-contained Lindblad-inspired dephasing-controlled transport MVP in experiment.py. The mechanism checks passed: off-diagonal coherence decay matched gamma within numerical precision, measured diffusion coefficients matched 2J^2/gamma within 0.1%, and variance scaling transitioned from ballistic (power 2.00 at gamma=0) to diffusive (power 1.05 at gamma=20). The effect is real in the toy transport model, although this does not establish a win on a trained neural-network task.

Agent confidence: 9/10

Baseline: Coherent hopping at t=3: variance=18.000, mass=1.000, variance power=2.00.
Idea: Dephasing gamma=10 at t=3: variance=1.160, mass=1.000, variance power=1.10; measured D=0.200 versus predicted 0.200. Across gamma=2,5,10,20,40, coherence-rate and diffusion-coefficient ratios were approximately 1.000.

Limitations: Only the finite one-dimensional correlation-matrix transport mechanism and impulse-response proxy were tested. No Copy, Selective-Copy, adding task, recurrent neural-network training, FLOP/speed comparison, gradient-sensitivity experiment, low-rank scalability implementation, or GPU benchmark was performed.

How to run: python3 experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.22240) Sep 2, 2026

Beats the tuned baseline by 79.5% (p=0.0081, wins 8 / 8 paired seeds; metric: lower is better).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0007
Idea mean: 0.0002
Effect (Δ): -0.0006 (−79.5%; negative = idea better)
Wins: 8 / 8 paired seeds
p-value: 0.0081 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±35.9%
Mechanism: Not confirmed ✗
Practical effect: Helps
Baseline tuning: swept over 3 configs

Limitations:

Only the built-in dynamics track was tested. The explicit correlation tensor is O(L^2) with L=8, so scalability and speed were not benchmarked. The trained-model mechanism signature did not quantitatively match the analytical decay prediction.

How to run: python3 bench_experiment.py

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

Stage 2 — Benchmark Aug 30, 2026 latest Worked ✓

Agent confidence: 8/10

Implemented an end-to-end dephasing-controlled transport system on the structurally matched dynamics track, with the same task, training budget, input projection, hidden width, and readout as the GRU baseline. The idea achieved significantly lower test MSE: 0.0001529 versus 0.0007449, with all 8 paired seeds improving and permutation p=0.0081. The trained-model mechanism signature was not confirmed: observed coherence ratio 0.7528 versus predicted 0.4493.

Baseline: mean test MSE 0.0007449479, std 0.0002542016, best lr 0.01
Idea: mean test MSE 0.0001529217, std 0.0000820060, best lr 0.01, paired delta -0.0005920262, p=0.0081, 8/8 wins

Benchmark result