✗ Failed on benchmark 2026

Residual-to-State Update Throttle

Implementation & benchmark of arXiv:2608.09030 — Closing the loop in learning with missing data

Usefulness7/10

Difficulty4/10

Novelty7/10

Dynamics Stability Generalization Optimizer Training dynamics Scheduler Control theory Dynamical systems Optimization Linear algebra

Source paper: Closing the loop in learning with missing data arXiv:2608.09030 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Add a closed-loop scalar gain that throttles a neural-network update when the observed loss residual is inconsistent with the available masked-gradient geometry. This converts the paper's ISS-style residual-to-parameter boundedness idea into a trust-region optimizer that permits aggressive updates during recurrent excitation but freezes weakly observed or contradictory directions.

Formulas

$$u_t=B_t^{-1}g_t,\qquad q_t=\operatorname{tr}(G_t),\qquad G_t=J_t^\top J_t,$$

$$\delta_t=\frac{\|r_t\|_2}{\sqrt{q_t}+\varepsilon},\qquad a_t=\min\left\{1,\frac{\kappa}{\delta_t+\varepsilon}\right\},$$

$$\theta_{t+1}=\theta_t-\eta a_tB_t^{-1}g_t,$$

$$\|e_t\|\lesssim C\rho^{\lfloor t/W\rfloor}\|e_0\|+\frac{C}{1-\rho}\sup_i\|d_i\|,$$

Mathematical statement

Let $g_t$ be the observed minibatch gradient, $u_t=B_t^{-1}g_t$ the observability-preconditioned direction, and $r_t$ the vector of observed prediction residuals. Define the excitation energy $q_t=\operatorname{tr}(G_t)$, where $G_t=J_t^\top J_t$. Form the mismatch ratio $\delta_t=\|r_t\|_2/(\sqrt{q_t}+\varepsilon)$, with $\varepsilon>0$ preventing division by zero. Choose a bounded gain $a_t\in[0,1]$ using $a_t=\min\{1,\kappa/(\delta_t+\varepsilon)\}$, where $\kappa>0$ is a design threshold, and update $\theta_{t+1}=\theta_t-\eta a_tu_t$. Bounded mismatch then acts as a bounded forcing term in an ISS-like parameter-error inequality.

Implementation notes

Place the throttle between gradient computation and the optimizer update, after applying the masked observability preconditioner or a standard diagonal preconditioner. For each minibatch, compute predictions using observed inputs, retain the residual vector $r=y-f_\theta(Mx)$, estimate $q=\operatorname{tr}(J^\top J)$ with Hutchinson probes $q\approx m^{-1}\sum_s\|Jz_s\|^2$, and calculate $\delta=\|r\|/(\sqrt q+\varepsilon)$. Set $a=\min(1,\kappa/(\delta+\varepsilon))$, multiply the preconditioned gradient by $a$, optionally clip for numerical safety, and update parameters. The paper motivates the residual-to-state and ISS construction; constants $C$, $\rho$, and $W$ should be estimated from trajectories rather than assumed known. Start with synthetic linear regression using a two-layer MLP whose true parameter is known. Impose alternating masks that hide each feature for 10-50 steps, and compare SGD, Adam, and the throttled method at learning rates near the divergence threshold. Log $a_t$, parameter error, and residual. The quantitative prediction is a sharp reduction in parameter-error spikes whenever $\delta_t>\kappa$, with error settling to a plateau proportional to the empirical maximum mismatch. Under recurrent excitation, the error envelope should decay geometrically between mask transitions rather than grow without bound.

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 9/10

Built a deterministic residual-to-state throttle for masked linear regression, with exact gradient scaling and scalar stability checks. The predicted gain boundary a<1 occurred at residual norm κ√q with 0.06% relative sweep error; in the throttled regime, residual_norm·a/√q matched κ to about 4e-9. Plain scalar GD diverged for η>2 while throttled updates remained bounded in all tested cases. The mini-experiment showed the idea improved final error over SGD at η=0.24 (0.209 vs 0.353) and prevented instability, but was worse at η=0.08 because its mean gain was only 0.066, so this is a mechanism win and promising stability signal, not a universal optimization win.

Agent confidence: 9/10

Baseline: Masked SGD: final parameter error 0.247, 0.311, 0.353 at learning rates 0.08, 0.16, 0.24; scalar η=2.5 final magnitude 4.07e17 and η=3.0 final magnitude 1.27e30.
Idea: Throttle: final parameter error 0.977, 0.397, 0.209 at learning rates 0.08, 0.16, 0.24; scalar η=2.5 final magnitude 0.464 and η=3.0 final magnitude 0.400. Mean gains were 0.066-0.082 in the masked regression runs.

Limitations: The main mini-experiment used exact q for a linear model rather than Hutchinson Jacobian estimates or a two-layer MLP. It did not test large neural networks, wall-clock/FLOP overhead, realistic observability preconditioners B, noisy/nonlinear residual geometry, or longer recurrent-excitation trajectories. Adam was included only as a small comparison and was not hyperparameter-tuned.

How to run: python3 sweeps.py && /home/maxwelhelp/main/bin/python3 experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.09030) Sep 2, 2026

Ahead of the baseline (−2.7%) but not statistically significant (p=0.5844, wins 6 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0004
Idea mean: 0.0004
Effect (Δ): -0 (−2.7%; negative = idea better)
Wins: 6 / 8 paired seeds
p-value: 0.5844 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±13.9%
Mechanism: Not confirmed ✗
Practical effect: Inconclusive
Baseline tuning: swept over 3 configs

Limitations:

Only the built-in dynamics track was tested; no tabular, vision, sequence, custom-track, Hutchinson probe-count ablation, alternative preconditioner, Adam hyperparameter sweep beyond learning rate and weight decay=0, or longer/higher-learning-rate stability stress test was run.

How to run: python3 bench_runner.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

Built the Residual-to-State Update Throttle as a local optimizer intervention on the structurally matched actuated-pendulum dynamics track, using the shared rnn_small architecture and paired 8-seed evaluation. The scalar math check passed: ordinary eta=3 scalar GD grew to 2^60 while the throttled state remained bounded, with transition relative error about 0.1%. On the real neural benchmark, the idea had slightly lower MSE but no significant win (paired delta -1.08949e-05, permutation p=0.58435), and the trained-model behavioral signature was not confirmed because delta was usually far below kappa.

Baseline: Adam rnn_small dynamics, lr=0.01 selected by sweep: mean test MSE 0.0004087210, std 0.0000730923, per-seed [0.00035143, 0.00042901, 0.00044985, 0.00050326, 0.00025691, 0.00039922, 0.00039931, 0.00048078]
Idea: Residual-throttled Adam rnn_small, lr=0.01 and kappa=0.5: mean test MSE 0.0003978261, std 0.0001023163, per-seed [0.00029170, 0.00055108, 0.00044954, 0.00048635, 0.00025691, 0.00028041, 0.00039917, 0.00046745]; paired delta=-0.0000108949, 6/8 wins, p=0.58435; verdict=no significant win
Math check: Predicted throttle transition 0.926013 versus observed 0.926939, relative error 0.001; plain scalar final absolute state 1.1529e18 versus throttled 0.4000006
Mechanism signature: Held-out trained-model probe: residual_norm=0.0976714, sqrt_q=14.04515, delta=0.0069541, gain=1.0, kappa=0.5; observed throttled fraction=0.02381; confirmed=false

Benchmark result