✗ Failed on benchmark 2026

Tangent-Cone Score Target

Implementation & benchmark of arXiv:2607.04514 — Boundary-layer asymptotics for Gaussian-smoothed singular measures

Usefulness7/10

Difficulty6/10

Novelty6/10

Geometry Accuracy Stability Sample efficiency Diffusion Diffusion sampling Loss Geometry Geometry Differential geometry Probability Asymptotic analysis

Source paper: Boundary-layer asymptotics for Gaussian-smoothed singular measures arXiv:2607.04514 ⓘ · analyzed Aug 30, 2026

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 Not peer-reviewed.

Idea description

Use the Gaussian mass of the local inward tangent cone to construct an analytic score target for noisy points lying within O(\sigma) of a support boundary or corner. This prevents a score network from learning an incorrect full-manifold or Euclidean approximation in the region where diffusion sampling is most sensitive to support truncation.

Formulas

$$\Psi(\zeta;a,x)=\frac{1}{2}\left[\|\mathbf{L}_{\mathcal{S}}(x)\,\zeta\|^{2}+\|a_{\mathcal{C}}-\mathbf{L}_{\mathcal{C}}(x)\,\zeta\|^{2}+\|a_{\mathcal{N}}\|^{2}\right].$$

$$p_\sigma\!\left(x+\sigma\mathbf{C}(x)a_{\mathcal C}+\sigma\mathbf{N}(x)a_{\mathcal N}\right)\approx (2\pi\sigma^2)^{-d/2}\sigma^m\rho(x)J(x)\int_{\mathbb{R}^{m-c}\times[0,\infty)^c}e^{-\Psi(\zeta;a,x)}\,d\zeta.$$

$$\boldsymbol{s}_\sigma(y)=\frac{\mathbb{E}[X\mid Y=y]-y}{\sigma^2},\qquad \widehat{\boldsymbol{s}}_\sigma(y)=\frac{1}{\sigma}\left(\mathbb{E}_{w}[v(\zeta)]-\mathbf{C}(x)a_{\mathcal C}-\mathbf{N}(x)a_{\mathcal N}\right).$$

$$w(\zeta\mid a,x)=\frac{\exp[-\Psi(\zeta;a,x)]}{\int_{\mathbb{R}^{m-c}\times\mathbb{R}_{+}^c}\exp[-\Psi(\zeta';a,x)]\,d\zeta'},\qquad v(\zeta)=\mathbf{S}(x)\mathbf{L}_{\mathcal S}(x)\zeta+\mathbf{C}(x)\mathbf{L}_{\mathcal C}(x)\zeta.$$

Mathematical statement

For a support manifold \mathcal{M}\subset\mathbb{R}^{d} of dimension m, let x lie on a codimension-c stratum, let k=d-m, and let \zeta=(\zeta_{\mathcal S},\zeta_{\mathcal C})\in\mathbb{R}^{m-c} imes[0,\infty)^c. Let \mathbf{S}(x) and \mathbf{C}(x) be orthonormal tangent frames along the stratum and its c inward transverse tangent directions, let \mathbf{N}(x) be an ambient normal frame, and let \mathbf{L}_{\mathcal S}(x),\mathbf{L}_{\mathcal C}(x) be the coordinate Jacobian blocks of the local chart. For a boundary-layer observation y=x+\sigma\mathbf{C}(x)a_{\mathcal C}+\sigma\mathbf{N}(x)a_{\mathcal N}, the paper gives the limiting exponent \Psi(\zeta;a,x)=\frac12[\|\mathbf{L}_{\mathcal S}(x)\zeta\|^2+\|a_{\mathcal C}-\mathbf{L}_{\mathcal C}(x)\zeta\|^2+\|a_{\mathcal N}\|^2]. If \rho(x) is the support density and J(x) the adapted chart Jacobian, then the leading smoothed density is proportional to \sigma^{m-d}\rho(x)J(x)\int_{\mathbb{R}^{m-c} imes\mathbb{R}_{+}^{c}}e^{-\Psi(\zeta;a,x)}d\zeta. The exact Gaussian posterior identity is
abla_y\log p_\sigma(y)=(\mathbb{E}[X\mid Y=y]-y)/\sigma^2 for Y=X+\sigma\varepsilon. Replacing the posterior by the cone posterior gives the adapted score approximation \widehat{\boldsymbol{s}}_\sigma(y)=\sigma^{-1}[\bar v-\mathbf{C}a_{\mathcal C}-\mathbf{N}a_{\mathcal N}], where v(\zeta)=\mathbf{S}\mathbf{L}_{\mathcal S}\zeta+\mathbf{C}\mathbf{L}_{\mathcal C}\zeta and \bar v is its expectation under weights proportional to e^{-\Psi}.

Implementation notes

Integrate this at the score-target construction stage of a diffusion or denoising-score model, before the neural network sees the noisy sample. First support a known or estimated local boundary representation. For each clean sample x, assign a stratum codimension c, estimate an orthonormal tangent frame by local PCA, identify inward directions C and ambient normals N, and estimate the chart blocks L_S and L_C. Draw y=x+sigma*epsilon in the usual way, then compute tubular coordinates approximately by projection onto the nearest boundary stratum: a_C=C^T(y-x)/sigma and a_N=N^T(y-x)/sigma. If the sample is farther than r*sigma from every boundary, retain the ordinary denoising target (x-y)/sigma^2. Otherwise evaluate the cone posterior using Gauss-Hermite quadrature for unconstrained zeta_S and truncated Gauss-Hermite or importance samples for zeta_C>=0. The pseudocode is: zeta <- samples_from(R^(m-c) x R_+^c); logw <- -0.5*(norm(L_S@zeta)^2 + norm(a_C-L_C@zeta)^2 + norm(a_N)^2); w <- softmax(logw); vbar <- sum(w*(S@L_S@zeta + C@L_C@zeta)); target <- (vbar-C@a_C-N@a_N)/sigma; loss <- ||score_net(y,sigma)-target||^2. The paper supplies the exponent and cone domain; local frames, projection, density, and Jacobians are estimated from data or known geometry. Start with a 2D score network on the uniform distribution over a square or quarter-disk, comparing standard denoising score matching against the cone target at sigma values from 0.01 to 0.2. Measure score error within 2sigma of edges, normal-component error, constraint violations during reverse sampling, and sliced Wasserstein distance. Success means lower boundary score error and fewer samples leaving the support at equal network size and sampling steps.

Verification

Failed on benchmark

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Not reproduced ✗

Stage 2 · Practical benchmark result: Inconclusive

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 analytic tangent-cone score target for the 2D square, with exact half-line cone posterior calculations and a matched CUDA toy score-network experiment. Prediction 1 (cone mass = Phi(a)) matched at a=-1,0,1,2 with maximum error 2e-16; prediction 2 (sigma times score is invariant) matched Mills ratios across sigma=.01,.05,.2; prediction 3 (omitted-cone normal error decays with inward distance) matched exactly from 7.979 at a=0 to 0.0444 at a=3 for sigma=.1. The learned cone target reduced boundary normal RMSE from 11.43 to 7.72 at sigma=.03 and 3.24 to 1.88 at .07, but worsened it at .15 (1.93 to 2.86), so the mechanism is real but not a universal training win.

Agent confidence: 8/10

Baseline: Boundary normal RMSE: sigma=.03 11.428, .07 3.236, .15 1.926; analytic omitted-cone error at sigma=.1: a=0 7.979, .5 5.092, 1 2.876, 2 .552, 3 .0444.
Idea: Boundary normal RMSE: sigma=.03 7.716 (-32.5%), .07 1.876 (-42.0%), .15 2.860 (+48.5%). Cone mass ratios observed/predicted: a=-1 .158655/.158655, 0 .5/.5, 1 .841345/.841345, 2 .977250/.977250; sigma-score invariance errors were at floating-point precision.

Limitations: Only a flat 2D square and boundary/corner-local analytic targets were tested; no learned geometry, curved manifolds, quarter-disk, reverse diffusion sampling, constraint-violation rate, sliced Wasserstein distance, ablations over quadrature/sample estimation, or larger models were evaluated. The high-noise regime showed the local target can be inferior when opposite boundaries interact.

How to run: python3 experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.04514) Sep 2, 2026

Ahead of the baseline (−1.8%) but not statistically significant (p=0.4106, wins 5 / 8 paired seeds).

Benchmark: Tangent_cone_square_score
Model: mlp_tiny
Paired seeds: 8
Baseline mean: 30.7124
Idea mean: 30.1497
Effect (Δ): -0.5627 (−1.8%; negative = idea better)
Wins: 5 / 8 paired seeds
p-value: 0.4106 (permutation test, 20 000 shuffles)
Smallest detectable effect: n/a
Mechanism: Not confirmed ✗
Practical effect: Inconclusive
Baseline tuning: swept over 3 configs

Limitations:

Only the custom flat uniform-square score task and mlp_tiny were tested. Curved supports, corners, reverse diffusion sampling, constraint violations, and larger architectures were not evaluated. The mechanism signature was not independently measured.

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 Failed ✗

Agent confidence: 9/10

Implemented and ran the full 8-seed paired benchmark with a custom uniform-square score-regression track and matched mlp_tiny architecture. The idea reduced mean MSE from 30.7124 to 30.1497, but the paired permutation p-value was 0.41055, so there was no significant win.

Baseline: Best sweep lr=0.01, epochs=12; full mean MSE 30.7124, std 7.7514.
Idea: Best of three settings; mean MSE 30.1497, std 7.2603; paired delta -0.5627, p=0.41055, 5/8 wins.

Benchmark result