△ Mechanism confirmed, baseline not beaten 2026

Excitation-Gated Neural Calibration

Implementation & benchmark of arXiv:2608.12528 — Excitation-Supervised Closed-Loop Self-Calibration and Target Seeking for an Unknown-Pose Range-Bearing Relay

Usefulness8/10

Difficulty6/10

Novelty7/10

Dynamics Stability Accuracy Sample efficiency World model Rl Optimizer Training dynamics Control theory Statistics Dynamical systems Linear algebra

Source paper: Excitation-Supervised Closed-Loop Self-Calibration and Target Seeking for an Unknown-Pose Range-Bearing Relay arXiv:2608.12528 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Add a calibration head to an online world model or sensor-fusion network that estimates an unknown nuisance transform, such as sensor-to-body rotation, feature-space alignment, or a latent affine offset. Maintain a recent trajectory excitation certificate and permit the policy or predictor to use the calibrated latent state only when the certificate exceeds an accuracy-derived threshold; otherwise inject an exploratory perturbation whose direction is chosen not to oppose the nominal task update. This replaces a fixed exploration-decay schedule with feedback from the information actually accumulated by the network.

Formulas

S_u(k) = sum_{j=k-L+1}^k ||u_j - u_bar_k||^2, where u_bar_k = (1/L) sum_{j=k-L+1}^k u_j.

var(psi_hat_k) <= sigma_eff^2 / S_v(k), where sigma_eff = max(sigma_r, r_max sigma_theta).

F_k = sum_{j=k-L+1}^k J_j^T Sigma^(-1) J_j, with J_j = partial h_phi(u_j, a) / partial a, and require lambda_min(F_k) >= gamma = epsilon_a^(-2).

g_perp = g - (g^T e)e, and Delta = g_perp + alpha e if lambda_min(F_k) < gamma, otherwise Delta = g.

Mathematical statement

For a recent window of L samples with excitation vectors u_j in R^d, define the centered spread S_u(k) as the sum over j from k-L+1 to k of the squared norm of u_j minus their window mean. In the paper's planar range-bearing problem, eliminating unknown translation gives yaw information at least S_v divided by sigma_eff squared, where sigma_eff equals the maximum of the range-noise standard deviation sigma_r and r_max times the bearing-noise standard deviation sigma_theta. Thus the yaw estimator variance is at most sigma_eff squared divided by S_v. For a neural calibration parameter a in R^m, use the empirical Gauss-Newton or Fisher matrix F_k equal to the sum of J_j transpose Sigma inverse J_j, where J_j is the calibration-head Jacobian and Sigma is the observation-noise covariance. Require the smallest eigenvalue of F_k to exceed gamma, with gamma equal to the inverse square of the desired calibration standard deviation epsilon_a. During uncertified operation, let g_k be the nominal policy or training direction and e_k be a normalized exploration direction. Remove the component of g_k parallel to e_k, then add exploration. Because a rolling window forgets old observations, certification must be checked at every step.

Implementation notes

1. Integration point: use an online recurrent world model or sensor-fusion MLP with model parameters phi, a calibration vector a representing an unknown sensor-frame rotation, translation, or latent affine map, and a task head predicting the next observation or control-relevant state. Retain the last L feature or input vectors and their calibration Jacobians. Before certification, either freeze the calibration estimate or provide the task head with a bounded fallback representation; after certification, use the calibrated representation normally. 2. Pseudocode: at every step, compute each recent calibration Jacobian J_j by autodiff; accumulate F = sum of J_j transpose Sigma inverse J_j; estimate the smallest eigenvalue of F with a symmetric eigensolver; set certified = (eig_min >= gamma). Compute the nominal action or parameter update g. If uncertified, sample or construct a normalized excitation direction e, replace g with g - dot(g,e)e + alpha e, execute or apply the update, and optimize the calibration loss. If certified, use the unrestricted nominal update. 3. Taken directly from the paper: the rolling-window spread certificate, the accuracy-based threshold, and repeated monitoring after old samples leave the window. Estimated empirically: neural Jacobians, Sigma, the effective noise level, and the excitation direction. Add damping F + lambda I if the Jacobian is nearly rank deficient. 4. First cheap experiment: train a small GRU sensor-fusion model on synthetic two-dimensional trajectories with a randomly rotated sensor frame, Gaussian range and bearing noise, and a limited exploration budget. Compare fixed exponentially decaying exploration with the supervisor using the same action budget and L = 20. Sweep epsilon_a and set gamma = epsilon_a^(-2). Log lambda_min(F), calibration error, task prediction error, and certification time. The primary prediction is inverse information scaling: doubling lambda_min(F) should approximately halve calibration variance, within noise. Certification should occur near lambda_min(F) = gamma. A second prediction is a sharp failure boundary for the fixed schedule when its decay time is shorter than the median time required to reach gamma, whereas supervised runs should retain at least 95 percent certification success under the same noise and budget.

Verification

Mechanism confirmed, baseline not beaten

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

Stage 2 · Practical benchmark result: No measurable effect

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 readable NumPy MVP implementing the centered-spread certificate, translation-eliminated Fisher information, inverse-variance calibration prediction, threshold gating, and rolling-window recertification. The mechanism manifested: empirical yaw variance followed sigma^2/S within 5%, the observed certification transition matched the predicted spread threshold, and certification was lost exactly when excited samples exited the rolling window. Gating retained certification for 90% of steps versus 22.5% for fast fixed decay in the controlled toy comparison, although both methods certified initially.

Agent confidence: 9/10

Baseline: Fast fixed decay: certification rate by step 40 = 1.00, median first certification = 3 steps, mean certified fraction = 0.225.
Idea: Feedback-gated excitation: certification rate by step 40 = 1.00, median first certification = 2 steps, mean certified fraction = 0.900; empirical/predicted variance ratios = 0.951-1.049.
Math checks: Predicted gamma = 100 and required spread = 0.64; observed transition between spread 0.166 (uncertified) and 0.663 (certified). Rolling-window loss occurred at step 39, matching predicted 2L-1 = 39.

Limitations: This is a linearized NumPy mechanism test, not a full GRU, neural calibration head, range-bearing simulator, or policy-learning benchmark. The acquisition comparison is controlled and deterministic across trials, so it does not establish a broad RL performance win or robustness under varied dynamics and noise.

How to run: python3 excitation_gated_calibration.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.12528) Sep 2, 2026

Ahead of the baseline (−0.1%) but not statistically significant (p=0.9392, wins 4 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0014
Idea mean: 0.0014
Effect (Δ): -0 (−0.1%; negative = idea better)
Wins: 4 / 8 paired seeds
p-value: 0.9392 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±2.0%
Mechanism: Confirmed ✓
Practical effect: No effect
Baseline tuning: swept over 9 configs

Limitations:

Only the built-in dynamics track was tested. The implementation uses action excitation as a proxy rather than a learned calibration head with autodiff Jacobians, and the mechanism signature uses an empirical-Fisher inverse proxy rather than direct calibration-error variance. No sensor-frame rotation, online deployment, broad noise sweep, or alternative dynamics task was tested.

How to run: python3 bench_excitation_gated.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

Excitation-gated training was evaluated on the structurally matched built-in dynamics track with the shared rnn_small architecture and standard test MSE. The idea slightly reduced mean MSE, but the paired improvement was not significant (delta_mean=-7.147682481445372e-07, p=0.93915), so there is no demonstrated benchmark win. The mechanism proxy was numerically consistent, but it is only an empirical-Fisher inverse proxy rather than direct learned calibration-variance evidence.

Baseline: Best fixed-decay configuration lr=0.01, decay=20.0; full 8-seed mean test MSE=0.0014148428163025528.
Idea: Best gated configuration lr=0.01, decay=20.0; full 8-seed mean test MSE=0.0014141280480544083; paired delta=-7.147682481445372e-07; 4/8 paired wins; permutation p=0.93915; verdict=no measurable effect.

Benchmark result