✗ Failed on benchmark 2026

Patch-Consensus Weak Residual Training

Implementation & benchmark of arXiv:2608.12479 — Identifying changing partial differential equations using Sampled Local WeakIdent

Usefulness7/10

Difficulty5/10

Novelty6/10

Training Stability Accuracy Sample efficiency Loss Moe routing World model Pde Numerical analysis Statistics Optimization

Source paper: Identifying changing partial differential equations using Sampled Local WeakIdent arXiv:2608.12479 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Train a neural PDE surrogate using weak residuals on randomly sampled local patches rather than pointwise derivative residuals. On every patch, identify which candidate differential-operator terms are consistently supported, then aggregate supports across many patches to obtain spatial equation regions and use the resulting consensus as a robust routing or auxiliary supervision signal.

Formulas

$$\int_{\Omega}u_t(x,t)\phi(x,t)\,dx\,dt=\sum_{l=1}^{L}c^{(l)}\int_{\Omega}f_l[u](x,t)\phi(x,t)\,dx\,dt.$$

$$A_j[l,r]=\int_{P_j}f_l[u_\theta](x,t)\phi_r(x,t)\,dx\,dt,\qquad b_j[r]=\int_{P_j}\partial_tu_\theta(x,t)\phi_r(x,t)\,dx\,dt.$$

$$\hat c_j=\arg\min_{c\in\mathbb{R}^{L}}\|A_jc-b_j\|_2^2+\lambda\|c\|_1,\qquad S_j=\{l:|\hat c_j^{(l)}|>\tau\}.$$

$$\hat S_R=\arg\max_{S}\sum_{j\in R}\mathbf{1}[S_j=S],\qquad \Pr(\hat S_R\neq S_R)\lesssim \exp\{-2m_R(p-1/2)^2\}.$$

Mathematical statement

The paper uses the weak formulation $\int_{\Omega}u_t(x,t)\phi(x,t)\,dx\,dt=\sum_{l=1}^{L}c^{(l)}\int_{\Omega}f_l[u](x,t)\phi(x,t)\,dx\,dt$. Here $u$ is the field, $u_t$ is its time derivative, $\phi$ is a compactly supported test function, $f_l[u]$ is candidate operator feature $l$, and $c^{(l)}$ is its coefficient. For a neural prediction $u_\theta$ and patch $P_j$, define $A_j[l,r]=\int_{P_j}f_l[u_\theta](x,t)\phi_r(x,t)\,dx\,dt$ and $b_j[r]=\int_{P_j}\partial_tu_\theta(x,t)\phi_r(x,t)\,dx\,dt$, where $r$ indexes test functions. Estimate local sparse coefficients using $\hat c_j=\arg\min_c\|A_jc-b_j\|_2^2+\lambda\|c\|_1$, with sparsity parameter $\lambda$. The local support is $S_j=\{l:|\hat c_j^{(l)}|>\tau\}$, where $\tau$ is a coefficient threshold. For a spatial region $R$, select the modal support $\hat S_R=\arg\max_S\sum_{j\in R}\mathbf{1}[S_j=S]$. If patch decisions are independently correct with probability $p>1/2$, the majority-error probability decreases approximately as $\exp[-2m_R(p-1/2)^2]$, where $m_R$ is the number of patches in region $R$. The transferable property is noise reduction through weak integration followed by statistical support consensus.

Implementation notes

1. Integration point: apply this as an auxiliary loss for a PINN, neural operator, or spatiotemporal world model predicting a field $u_\theta(x,t)$. Keep the standard observation loss and add weak-form consistency after the model begins producing reasonable fields. 2. Pseudocode: sample $M$ spatial-temporal patches $P_j$; draw quadrature points and smooth compactly supported test functions $\phi_r$; evaluate $u_\theta$, $\partial_tu_\theta$, and candidate features $f_l[u_\theta]$ by automatic differentiation; compute quadrature approximations to $A_j$ and $b_j$; solve the small LASSO for $\hat c_j$; threshold to produce $S_j$; group patches by spatial cell and compute the modal support $\hat S_R$. Optimize $L=L_{data}+\alpha\sum_j\|A_j\hat c_j-b_j\|_2^2+\beta\sum_j\|\hat c_j-\bar c_{R(j)}\|_2^2$, where $\bar c_R$ is the mean coefficient vector for patches assigned to region $R$. Recompute sparse coefficients and supports every 10-50 neural updates, and stop gradients through the LASSO and discrete support decisions initially. 3. Computed directly from the mathematics: weak integrals, local sparse coefficients, supports, regional modal supports, and confidence estimates. Tune empirically: $\lambda$, $\tau$, test-function width, patch size, and $\alpha,\beta$. Estimate $p$ from repeated patch decisions or held-out synthetic ground truth. 4. First experiment: train an MLP PINN on a 2D advection-diffusion field with two spatial regions and different coefficients, using 1-5% noisy observations. Compare pointwise PINN residuals, weak residuals without consensus, and the complete method at equal model evaluations. Measure relative field error, coefficient/support F1, region-boundary accuracy, and gradient variance. Success is lower noisy-data error, more consistent support recovery across seeds, and faster loss reduction at equal FLOPs.

Verification

Failed on benchmark

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Not reproduced ✗

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

Built a readable MVP implementing weak patch integration, local LASSO support estimation, and modal support consensus. The mechanism manifested: weak-integration noise scaled with log-log slope -0.481 versus the predicted -0.5, majority error decreased with patch count for p>0.5, and increased toward certainty of the wrong decision for p<0.5. In the synthetic sparse-regression check, exact local support accuracy was 0.997 for individual patches and 1.000 for modal consensus, although this secondary setup was relatively easy.

Agent confidence: 8/10

Baseline: Individual patch exact support accuracy: 0.9971 mean over 100 trials. Weak-integration standard deviations for n=[8,16,32,64,128,256] were [0.4245, 0.3385, 0.2174, 0.1623, 0.1191, 0.0816].
Idea: Modal consensus exact support accuracy: 1.000 over 100 trials. For local decision accuracy p=0.7, majority error fell from 0.2181 at m=3 to 0.00783 at m=31, versus Hoeffding bounds 0.7866 and 0.0837. Weak-noise scaling was -0.4811 versus the predicted -0.5.

Limitations: Did not train or compare a full neural PDE surrogate or PINN, and did not measure field error, FLOPs, runtime, boundary accuracy, gradient variance, or GPU performance. The LASSO and patch regression are synthetic MVP approximations, and the consensus theorem test assumes independent binary local decisions rather than correlated neural patches.

How to run: python3 weak_consensus_experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.12479) Sep 2, 2026

No significant difference vs. the tuned baseline (+0.0%, p=0.7982, wins 6 / 8 paired seeds).

Benchmark: Pde_patch_consensus
Model: mlp_tiny
Paired seeds: 8
Baseline mean: 0.2987
Idea mean: 0.2987
Effect (Δ): 0 (+0.0%; negative = idea better)
Wins: 6 / 8 paired seeds
p-value: 0.7982 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±0.0%
Mechanism: Not confirmed ✗
Practical effect: No effect
Baseline tuning: swept over 3 configs

Limitations:

Only the registered small 1D PDE regression track was tested. No 2D PDE, boundary accuracy, operator coefficient F1, FLOP accounting, gradient variance, or broader architectures were evaluated; local sparse estimation was a detached ridge-plus-soft-threshold proxy rather than a full differentiable LASSO.

How to run: python3 bench_run.py

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

Stage 2 — Benchmark Sep 1, 2026 latest Failed ✗

Agent confidence: 9/10

The full stage-2 bench was completed on the registered custom PDE track pde_patch_consensus with matched mlp_tiny systems, a tuned baseline sweep, and 8 paired seeds. Patch-consensus weak residual training produced no significant improvement: test MSE delta was +0.0000288114 with permutation p=0.7982, and the trained-model mechanism signature was not confirmed.

Baseline: Full mean test MSE 0.2986701727; best config lr=0.001, weight=0.15, epochs=20.
Idea: Mean test MSE 0.2986989841; paired delta +0.0000288114; permutation p=0.7982; 6/8 paired wins but no significant win; mechanism signature confirmed=false.

Benchmark result