✓✓ Beats tuned baseline 2026

Time-Shell Long-Horizon Decoder

Implementation & benchmark of arXiv:2607.19208 — Dynamical correlation functions of extensive charges after global quantum quenches

Usefulness7/10

Difficulty5/10

Novelty8/10

Architecture Accuracy Memory Scalability Ssm Rnn Transformer World model Statistical mechanics Stochastic processes Dynamical systems Renormalization

Source paper: Dynamical correlation functions of extensive charges after global quantum quenches arXiv:2607.19208 ⓘ · analyzed Aug 30, 2026

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 Not peer-reviewed.

Idea description

Replace dense pairwise interactions between all forecast horizons with nested time-shell summaries. For sorted horizons, the readout at shell j receives a cumulative embedding of all coefficients or queries assigned to later horizons, reproducing the paper's dependence on products such as $\Pi_j=\prod_{l>j}e^{\alpha_l}=e^{\sum_{l>j}\alpha_l}$. This gives an $O(Kd)$ multi-horizon interaction instead of an $O(K^2d)$ temporal attention block and should work best for weak-memory, approximately ballistic sequence dynamics.

Formulas

$$\Pi_j=\prod_{l=j+1}^{K}e^{\alpha_l}=e^{\sum_{l=j+1}^{K}\alpha_l},\qquad B_j=\sum_{l=j}^{K}\beta_l.$$

$$\mathcal F^{(K)}(\{t_j,\beta_j\})\approx\sum_{j=1}^{K}f_\theta\!\left(h_j,B_j\right),\qquad h_{j+1}=F_\theta\!\left(h_j,x_{j:j+1}\right).$$

$$\left\langle\mathcal T Q(t_K)\cdots Q(t_1)\right\rangle_c\approx G_K(t_{\min}),\qquad t_{\min}=\min_j t_j.$$

$$\operatorname{cost}_{\mathrm{shell}}(K,d)=O(Kd),\qquad \operatorname{cost}_{\mathrm{pairwise}}(K,d)=O(K^2d).$$

Mathematical statement

Let $t_1\leq t_2\leq\cdots\leq t_K$ be forecast times, $\beta_j$ the query or measurement coefficient at time $t_j$, and $B_j=\sum_{l=j}^{K}\beta_l$ the cumulative shell coefficient. The paper's space-time construction implies that the multi-time cumulant-generating function can be approximated by a sum of single-time contributions evaluated on nested cumulative coefficients, $\mathcal F^{(K)}(\{t_j,\beta_j\})\approx\sum_{j=1}^{K}f(t_j,B_j)$, up to boundary and finite-size corrections. The equivalent transfer quantity is $\Pi_j=\prod_{l=j+1}^{K}e^{\alpha_l}=e^{\sum_{l=j+1}^{K}\alpha_l}$, showing that shell j depends only on the sum of later insertions, not their individual arrangement. The paper also predicts that ordered connected correlations depend only on the smallest time, $\langle\mathcal T Q(t_K)\cdots Q(t_1)\rangle_c\approx G_K(t_1)$, when temporal correlations are sufficiently weak. In the neural version, $f_\theta$ is a learned shell readout, $h_j$ is the hidden state at horizon $t_j$, and $\phi_\theta(h_j,B_j)$ is the learned single-shell contribution.

Implementation notes

1. Integration point: use a recurrent state-space model or causal Transformer to produce hidden states h_j at K requested forecast horizons. Add a shell decoder after shared state computation. Sort horizons, associate each horizon with a coefficient beta_j, and compute cumulative shell coefficients B_j with a reverse cumulative sum. Predict a scalar contribution, distribution parameter, or horizon-specific residual from phi_theta([h_j,B_j]); avoid pairwise horizon attention in this decoder. 2. Pseudocode: encode h=SSM(x); order horizons t[1:K]; B[K+1]=0; for j=K down to 1: B[j]=B[j+1]+beta[j]; shell[j]=phi(h[j],B[j]); total_prediction=sum(shell); optimize task_loss(total_prediction,y). Optionally add collapse_loss, which samples horizon tuples with equal minimum horizon and penalizes differences between their connected residual statistics. 3. Computed from the paper: the reverse cumulative-shell structure and the minimum-time dependence hypothesis. Estimated empirically: the neural shell map phi, finite-context corrections, and the regime in which temporal correlations are weak. 4. First cheap experiment: train a vanilla temporal-attention decoder, an S4-style state-space model, and the shell decoder on a 1D advection equation or linear cellular-automaton dataset with controllable propagation speed and noise. Use K=32 forecast horizons and measure validation loss, wall-clock time, memory, and connected third-order residual correlations. The quantitative prediction is linear scaling in K for the shell decoder versus quadratic scaling for pairwise attention. For ballistic advection, tuples with the same t_min should have connected-correlation curves agreeing within 10-20% once context length exceeds propagation time. For a diffusive or strongly mixing-control dataset, this collapse should fail and dense attention should regain an advantage, providing a falsification test.

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 8/10

Built a reproducible shell decoder toy experiment and verified three quantitative predictions: cumulative shell sums reproduce later-insertion products to about 1e-14 error, fitted complexity exponents are 1.00 for shell and 2.00 for pairwise, and normalized shell amplification approaches 0.5 while dense amplification grows as K. On the K=32 toy task, shell and pairwise MSE were essentially identical in both ballistic and mixing regimes, while the shell fit used 160 versus 1120 parameters and took 0.00485 versus 0.0183 seconds in the ballistic case. Thus the claimed scalability/stability mechanism manifested, but no accuracy advantage was observed.

Agent confidence: 8/10

Baseline: Pairwise decoder: ballistic MSE 0.0009228337, mixing MSE 0.9994222, 1120 parameters, fit time 0.0183 s / 0.0354 s.
Idea: Time-shell decoder: ballistic MSE 0.0009228338, mixing MSE 0.9994222, 160 parameters, fit time 0.00485 s / 0.00741 s; product identity max error <=1.1e-14; complexity exponents 1.00 vs 2.00.

Limitations: The decoder comparison is a small NumPy linear toy, not a trained recurrent/SSM or Transformer model. It does not measure GPU memory, wall-clock autoregressive inference, connected third-order residual collapse, variable horizon ordering, or a genuine advection PDE; the pairwise baseline is a feature-rich linear comparator rather than full temporal attention.

How to run: python3 experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.19208) Sep 2, 2026

Beats the tuned baseline by 7.3% (p=0.0081, wins 8 / 8 paired seeds; metric: lower is better).

Benchmark: Multihorizon_sequence
Model: custom_gru_decoder
Paired seeds: 8
Baseline mean: 0.2238
Idea mean: 0.2075
Effect (Δ): -0.0163 (−7.3%; negative = idea better)
Wins: 8 / 8 paired seeds
p-value: 0.0081 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±8.3%
Mechanism: Not confirmed ✗
Practical effect: Helps
Baseline tuning: swept over 6 configs

Limitations:

The built-in sequence track is single-horizon, so the structurally matched custom multihorizon_sequence track was promoted and registered before rerunning. The dense comparator uses an all-horizon coefficient vector rather than an explicit quadratic attention matrix, so O(K^2d) wall-clock scaling was not independently measured. No variable-horizon, memory, or connected third-order-correlation tests were run; the trained-model mechanism signature failed quantitatively.

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 Aug 31, 2026 latest Worked ✓

Agent confidence: 7/10

The registered multihorizon_sequence track was benchmarked with matched end-to-end GRU systems and a tuned baseline sweep. Time-shell decoding achieved lower MSE than the dense all-horizon baseline on all 8 paired seeds, with delta_mean=-0.0163224 and permutation p=0.0081, so the benchmark verdict is idea better (significant). The trained-model mechanism signature did not confirm the proposed early-over-late sensitivity prediction.

Baseline: Mean test MSE 0.2238331437, best cfg lr=0.006 scale=1.0, 8-seed full evaluation.
Idea: Mean test MSE 0.2075107191, best cfg lr=0.006 scale=1.0, 8-seed evaluation; 7.3% lower MSE than baseline.

Benchmark result