✗ Failed on benchmark 2026

Second-order fusion prior for point-set diffusion

Implementation & benchmark of arXiv:2608.23742 — Second-order Fusion Asymptotics for Sine\b{eta} Correlation Functions

Usefulness6/10

Difficulty5/10

Novelty6/10

Sampling Accuracy Stability Sample efficiency Diffusion sampling Diffusion Regularization Graph nn Probability Stochastic processes Statistics Asymptotic analysis

Source paper: Second-order Fusion Asymptotics for Sine\b{eta} Correlation Functions arXiv:2608.23742 ⓘ · analyzed Aug 29, 2026

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 Not peer-reviewed.

Idea description

Add the paper's local Sine_beta fusion law as an analytic score prior for diffusion models that generate unordered point configurations. The model is trained to match both the usual diffusion score and an explicit short-range repulsion score, including the second-order correction that describes finite-scale fused configurations.

Formulas

$$V(a)=\sum_{1\leq i<j\leq m}(a_i-a_j)^2$$

$$\rho^{(m)}_{\beta}(\varepsilon a_1,\ldots,\varepsilon a_m)=C^{(m)}_{\beta}|\varepsilon|^{\beta\binom{m}{2}}\prod_{i<j}|a_i-a_j|^{\beta}\left[1-\frac{\beta^{2}V(a)}{8(m\beta-1)(2m+1)}\varepsilon^{2}+o(\varepsilon^{2})\right],\qquad m\beta>1$$

$$\log p_{m,\beta}(x)=\beta\sum_{i<j}\log|x_i-x_j|-\kappa_{m,\beta}\sum_{i<j}(x_i-x_j)^2,\qquad \kappa_{m,\beta}=\frac{\beta^2}{8(m\beta-1)(2m+1)}$$

$$\frac{\partial\log p_{m,\beta}}{\partial x_k}=\beta\sum_{j\neq k}\frac{1}{x_k-x_j}-2\kappa_{m,\beta}\sum_{j\neq k}(x_k-x_j)$$

Mathematical statement

For m in the positive integers, beta > 0, and distinct local coordinates a_1,...,a_m, the paper defines V(a) = sum_{i<j}(a_i-a_j)^2 and proves, when m beta > 1, rho_beta^(m)(epsilon a_1,...,epsilon a_m) = C_beta^(m) |epsilon|^(beta binom(m,2)) product_{i<j}|a_i-a_j|^beta [1 - beta^2 V(a) epsilon^2 / (8(m beta-1)(2m+1)) + o(epsilon^2)]. Let kappa_(m,beta) = beta^2/[8(m beta-1)(2m+1)]. In physical local coordinates x_i, use the truncated log-density log p(x) = beta sum_{i<j} log|x_i-x_j| - kappa_(m,beta) sum_{i<j}(x_i-x_j)^2. Its coordinate score is s_k(x) = beta sum_{j != k} 1/(x_k-x_j) - 2 kappa_(m,beta) sum_{j != k}(x_k-x_j). The first term is Vandermonde repulsion and the second is the paper's second-order fusion correction. For numerical stability replace 1/d by d/(d^2+delta^2), where delta is a small collision-regularization constant.

Implementation notes

Integrate the prior at the score-prediction stage of a diffusion model whose output is a one-dimensional or low-dimensional unordered point set X with shape n by d. During each training step, randomly select local subsets of m points, center each subset by its mean, and apply a local scale normalization. Use only m and beta satisfying m beta > 1, and apply the prior only when the normalized subset diameter is below a threshold r_max. Sample diffusion time t and Gaussian noise z, form X_t = alpha_t X + sigma_t z, and evaluate the baseline score network s_theta(X_t,t). For every selected subset S, form pair differences d_kj = S_k - S_j. Compute rep_k = beta times the sum over j not equal to k of d_kj/(d_kj squared plus delta squared), and corr_k = -2 kappa_(m,beta) times the sum over j not equal to k of d_kj. Set q_k = rep_k + corr_k and add lambda(t) times the mean squared error between the network's local score and q to the ordinary diffusion score-matching loss. Use lambda(t) that decays at high noise levels, and clip q by a fixed norm to control near-collision explosions. The coefficient kappa is taken directly from the theorem; only local scale, lambda(t), delta, and r_max are tuned empirically. Pseudocode: S = random_local_subset(X,m); S = center_and_scale(S); d = S[:,None] - S[None,:]; rep = beta * sum_j(d/(d*d + delta*delta)); corr = -2*kappa * sum_j(d); q = clip(rep + corr); loss = diffusion_loss + lambda_t * mean((score_theta(S_t,t)-q)**2). Start with 16 to 64 points sampled from a known Sine_beta simulator. Compare the baseline, repulsion-only prior, and full second-order prior using pair-correlation error, nearest-neighbor spacing error, collision rate, and diffusion steps needed to reach a fixed Wasserstein error. A positive result is at least a 20 percent reduction in short-range correlation error or collision rate without increasing model size or 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: 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 reproducible analytic prior and toy point-set Langevin experiment. The beta=2 determinantal Sine-kernel expansion matched the theorem’s kappa to relative errors of 5.4e-10 for m=2 and 3.0e-9 for m=3, while the implemented coordinate score matched finite differences within 1.45e-10. The prior reduced normalized short-gap collisions from 9.11% baseline to 0.27% with the full correction, but the second-order correction itself did not improve collision rate over repulsion-only, so its incremental benefit remains unproven.

Agent confidence: 8/10

Baseline: collision rate 0.0911; gap standard deviation 0.9454; 5th-percentile normalized gap 0.0531
Idea: full second-order prior: collision rate 0.00271; gap standard deviation 0.4516; 5th-percentile normalized gap 0.3308; theorem kappa relative errors 5.4e-10 (m=2) and 3.0e-9 (m=3)

Limitations: This is not a trained diffusion network or Sine-beta simulator; it is a small Langevin proxy with global rather than randomly selected local subsets, no Wasserstein or pair-correlation evaluation, and no systematic tuning or multi-seed confidence intervals. The correction was evaluated at fixed beta=2 and its incremental effect versus repulsion-only was not a demonstrated win.

How to run: python3 run_experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.23742) Sep 3, 2026

Significantly WORSE than the tuned baseline (n/a, p=0.0081, wins 0 / 8 paired seeds).

Benchmark: Unordered_pointset_denoising
Model: mlp_tiny
Paired seeds: 8
Baseline mean: 0
Idea mean: 0.085
Effect (Δ): 0.0132 (n/a; negative = idea better)
Wins: 0 / 8 paired seeds
p-value: 0.0081 (permutation test, 20 000 shuffles)
Smallest detectable effect: n/a
Mechanism: Not confirmed ✗
Practical effect: Hurts
Baseline tuning: swept over 0 configs

Limitations:

The custom benchmark uses 6 one-dimensional points, a compact synthetic denoising distribution, and an MLP rather than a full multi-dimensional diffusion sampler or Sine_beta simulator. It evaluates fixed-noise denoising MSE rather than Wasserstein distance or multi-step generation quality. No larger architectures or point counts were tested.

How to run: python3 run_bench.py

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

Stage 2 — Benchmark Sep 3, 2026 latest Failed ✗

Agent confidence: 9/10

Built a structurally matched unordered point-set denoising track and an end-to-end MLP training intervention implementing the Vandermonde repulsion plus second-order fusion correction. Across 8 paired seeds, the best idea configuration had test MSE 0.0849502 versus baseline 0.0717791, paired delta +0.0131711 and permutation p=0.0081, hence a significant loss. The trained-model mechanism signature was also not confirmed: predicted short-gap collision rate increased from 0.0582812 to 0.0653125.

Baseline: test MSE mean 0.0717791, std 0.0038503; normalized predicted collision rate 0.0582812; best lr=0.003, weight_decay=0.0001
Idea: test MSE mean 0.0849502, std 0.0036924; normalized predicted collision rate 0.0653125; best lr=0.003, weight_decay=0.0001, lambda=0.1, beta=2, delta=0.08
Mechanism signature: predicted collision delta +0.00703125; confirmed=false
Paired delta: idea - baseline = +0.0131711 MSE; 0/8 idea wins; permutation p=0.0081

Benchmark result