△ Mechanism confirmed, baseline not beaten 2026

Residual-Gated Lift Depth

Implementation & benchmark of arXiv:2607.17664 — Graph-Induced Tensor Liftings for Networked SEIR Models: Dimensional Reduction and Residual Analysis

Usefulness7/10

Difficulty4/10

Novelty8/10

Dynamics Scalability Stability Accuracy Graph nn Training dynamics Scheduler World model Dynamical systems Graph theory Control theory Numerical analysis Linear algebra

Source paper: Graph-Induced Tensor Liftings for Networked SEIR Models: Dimensional Reduction and Residual Analysis arXiv:2607.17664 ⓘ · analyzed Aug 30, 2026

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 Not peer-reviewed.

Idea description

Use the paper's localized truncation residual as an online certificate for whether the current polynomial lift is expressive enough. Start with a low-degree edge lift and activate additional degree blocks or a learned closure only when the residual exceeds a calibrated threshold, avoiding the cost and instability of always using a large polynomial dictionary.

Formulas

$$\dot z_d=K_dz_d+r_d,\qquad r_d=\operatorname{col}(0,\ldots,0,D\Phi_d(x)f_Q(x)).$$

$$\widehat r_t=\frac{z_{t+1}-z_t}{\Delta t}-K_dz_t,\qquad \rho_t=\frac{\|\widehat r_t\|_2}{\|K_dz_t\|_2+\varepsilon}.$$

$$\|r_d(x)\|_2\le c_dqM^{d+1},\qquad \|x\|_2\le M,\quad \|f_Q(x)\|_2\le q\|x\|_2^2.$$

$$d_{t+1}=\begin{cases}d_t+1,&\rho_t>\tau_{\mathrm{up}}\text{ for }m\text{ consecutive steps},\\d_t-1,&\rho_t<\tau_{\mathrm{down}}\text{ for }m\text{ steps},\\d_t,&\text{otherwise}.\end{cases}$$

Mathematical statement

Let $z_d$ be the order-$d$ graph-induced lift and let $K_d$ be its retained linear operator. The exact lifted derivative is $\dot z_d=K_dz_d+r_d$, with $r_d$ supported only in the final block and equal to $\mathcal L_{f_Q}\Phi_d$. During training or inference, estimate the residual using finite differences, $\widehat r_t=(z_{t+1}-z_t)/\Delta t-K_dz_t$, or by comparing degree-$d$ and degree-$(d+1)$ models. A normalized residual indicator is $\rho_t=\|\widehat r_t\|_2/(\|K_dz_t\|_2+\varepsilon)$. Under bounded state norm $\|x\|\le M$ and quadratic-field norm $\|f_Q(x)\|\le qM^2$, homogeneity gives the scaling bound $\|r_d(x)\|\le c_dqM^{d+1}$, where $c_d$ depends on the selected tensor dictionary and derivative multiplicities. This predicts that closure error grows as degree $d+1$ in state amplitude and is amplified by graph weight accumulation through $q$, motivating a threshold controller for lift depth.

Implementation notes

1. Integration point: place the controller around a graph-lifted state-space model, latent ODE, or recurrent world model. Maintain degree-2 features initially, and pre-build sparse degree-3 and degree-4 feature maps so activation does not require recompiling the network. 2. Pseudocode: at each window, compute z_d; predict dz_pred=K_d@z_d+closure(z_d); from the next observed state compute r_hat=(z_next-z)/dt-K_d@z; calculate rho=norm(r_hat)/(norm(K_d@z)+eps); update an exponential moving average; if it exceeds tau_up, activate the next degree or increase closure width; if it remains below tau_down, prune the highest degree. Train with L=L_prediction+lambda_r*mean(rho^2) and optionally stop gradients through the discrete gate. 3. $K_d$ and the residual support pattern come from the lifting construction; $c_d$, $q$, measurement-noise variance, and thresholds are estimated empirically. Use hysteresis and a minimum dwell time to prevent degree chattering. 4. First cheap experiment: train on short trajectories from a 16-node SEIR graph and a randomly generated quadratic graph system, comparing fixed degree-2, fixed degree-3, and residual-gated models. Use identical decoders and optimizer settings. Test one-step error, 50-step rollout error, active feature count, and residual ratio. Calibrate $\tau_{\mathrm{up}}$ to the 95th percentile of normalized training residuals. The falsifiable signature is a sharp increase in rollout error when $\rho$ crosses the threshold, followed by recovery after degree activation; across state rescalings by factor $a$, the unnormalized residual should scale approximately as $a^{d+1}$, and across graph weight rescalings by factor $b$, it should scale approximately as $b$ in the weak-nonlinearity regime.

Verification

Mechanism confirmed, baseline not beaten

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

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 readable scalar quadratic lift MVP with exact truncated operators, residual certificate, hysteretic degree-2-to-3 gating, scaling checks, finite-difference verification, and fixed-degree rollout baselines. The mechanism manifested: residual norms scaled as a^(d+1) with observed exponents 3.000, 4.000, and 5.000 for degrees 2–4, scaled linearly with q with slope 1.000, and vanished at q=0; finite-difference relative error was below 3e-7. In the toy rollout, gating achieved RMSE 0.006175 versus 0.009882 for fixed degree 2 and 0.006287 for fixed degree 3 while averaging degree 2.98.

Agent confidence: 8/10

Baseline: Fixed d2: rollout RMSE 0.009882, final absolute error 0.008872, mean active degree 2.0. Fixed d3: rollout RMSE 0.006287, final absolute error 0.004609, mean active degree 3.0.
Idea: Residual-gated: rollout RMSE 0.006175, final absolute error 0.004466, mean active degree/features 2.98. Calibrated tau_up=0.20165; initial rho=0.21752 and activation occurred after the two-step dwell condition. Amplitude exponents observed/predicted: d2 3.000/3, d3 4.000/4, d4 5.000/5. q exponent observed/predicted: 1.000/1. Zero-q residual norm: 0.

Limitations: Only a one-dimensional quadratic dynamical system was tested, not the paper's 16-node SEIR graph or a learned neural closure. The certificate used noiseless Euler next states, thresholds were calibrated in the same toy regime, and no wall-clock, FLOP, or noisy-measurement robustness study was performed.

How to run: python3 residual_gated_lift.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.17664) Sep 2, 2026

Significantly WORSE than the tuned baseline (+38.7%, p=0.0324, wins 2 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0016
Idea mean: 0.0022
Effect (Δ): 0.0006 (+38.7%; negative = idea better)
Wins: 2 / 8 paired seeds
p-value: 0.0324 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±32.6%
Mechanism: Confirmed ✓
Practical effect: Hurts
Baseline tuning: swept over 6 configs

Limitations:

The intervention used a normalized cubic feature proxy rather than finite-difference observed derivatives or a learned closure; no noisy-observation robustness, wall-clock/FLOP accounting, or longer rollout benchmark was tested. The baseline and idea used 12 epochs on the standard 400/160 dynamics split.

How to run: python3 bench_residual_gated.py

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

Stage 2 — Benchmark Aug 31, 2026 latest Failed ✗

Agent confidence: 9/10

Implemented residual-gated polynomial lift features around a matched GRU on the structurally appropriate controlled-pendulum dynamics track. The trained-model cubic proxy scaled by 3.375 under 1.5x input amplitude, matching 1.5^3, but the gate was active on 99.5% of test examples. The idea was significantly worse: delta_mean=+0.0006127432 MSE, permutation p=0.0324, so worked=false.

Baseline: Fixed degree-2 lifted GRU, best lr=0.006: mean test MSE 0.001581461, std 0.000439250, 8 paired seeds.
Idea: Residual-gated cubic lifted GRU, best lr=0.006 threshold=0.4: mean test MSE 0.002194204, std 0.000755785, 2/8 paired wins; delta_mean=+0.0006127432, p=0.0324.
Mechanism signature: map[confirmed:true mean_active_fraction:0.9953125 mean_rho:0.8967330754 observed_mean_proxy_ratio_at_1p5:3.375 predicted_ratio_at_1p5:3.375 prediction:cubic residual proxy scales as amplitude^3 track_match:controlled-pendulum dynamics; standard test MSE on independently trained systems]

Benchmark result