△ Mechanism confirmed, baseline not beaten 2026

Excitation-Gated Latent Frame Calibration

Implementation & benchmark of arXiv:2608.09464 — Trajectory-Induced Self-Calibration for Hidden-Target Localization Through an Unknown-Pose Range-Bearing Relay

Usefulness8/10

Difficulty6/10

Novelty7/10

Dynamics Stability Accuracy Generalization Training dynamics Rnn Ssm World model Regularization Control theory Dynamical systems Linear algebra Geometry

Source paper: Trajectory-Induced Self-Calibration for Hidden-Target Localization Through an Unknown-Pose Range-Bearing Relay arXiv:2608.09464 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Add an explicit unknown-frame variable to a recurrent world model or multimodal sensor-fusion network, and train it only on temporal windows whose latent motion provides enough excitation to identify that frame. The model should use a two-view or multi-view consistency loss and an adaptive gate based on the smallest singular value of the window Jacobian, preventing optimization from confidently fitting geometrically ambiguous trajectories.

Formulas

$$q_k=x+R(\psi)\ell_k^v,\qquad p=x+R(\psi)\ell_k^t,$$

$$q_2-q_1=R(\psi)(\ell_2^v-\ell_1^v),\qquad \ell_2^v\neq\ell_1^v,$$

$$J_W=\frac{\partial}{\partial\theta}\begin{bmatrix}W_1^{1/2}r_1(\theta)\\ \vdots\\ W_T^{1/2}r_T(\theta)\\ W_1^{1/2}s_1(\theta)\\ \vdots\\ W_T^{1/2}s_T(\theta)\end{bmatrix},\qquad \kappa_W=\frac{\sigma_{\max}(J_W)}{\sigma_{\min}(J_W)},$$

$$\mathcal L=\sum_{k\in W}\rho\!\left(\|W_k^{1/2}r_k(\theta)\|^2\right)+\rho\!\left(\|W_k^{1/2}s_k(\theta)\|^2\right)+\lambda\,\max(0,\tau-\sigma_{\min}(J_W))^2,$$

Mathematical statement

For observations indexed by k, let q_k be a known or predicted vehicle state, p be a shared latent target, x be an unknown relay translation, R(psi) be a two-dimensional rotation with unknown yaw psi, and ell_k^v, ell_k^t be relay-frame vehicle and target vectors. The measurement equations are q_k=x+R(psi)ell_k^v and p=x+R(psi)ell_k^t. With only one k, translation and yaw remain continuously ambiguous; with two distinct observations k=1,2, the difference q_2-q_1=R(psi)(ell_2^v-ell_1^v) identifies psi when ell_2^v is not equal to ell_1^v. In a neural model, let h_k be its predicted latent pose, y_k its relay-frame observation, and theta=(p,x,psi) latent calibration parameters. Define residual vector r_k(theta)=h_k-x-R(psi)ell_k^v and s_k(theta)=p-x-R(psi)ell_k^t. Let J_W be the Jacobian of all whitened residuals in a window W with respect to theta; the excitation score is sigma_min(J_W), or equivalently the condition number kappa(J_W)=sigma_max(J_W)/sigma_min(J_W). Training updates are downweighted or delayed when sigma_min(J_W) is below a threshold.

Implementation notes

1. Integration point: use a small differentiable calibration head on top of an RNN, SSM, or transformer world model. The head predicts a shared latent target p, frame translation x, yaw psi, and per-step relay-frame vectors; the backbone predicts h_k and the calibration residuals are evaluated over sliding windows of length T. 2. Pseudocode: for each window, predict h[1:T], p, x, psi; compute R from psi; form r[k]=h[k]-x-R@ell_v[k] and s[k]=p-x-R@ell_t[k]; compute robust residual loss; obtain J by autodiff or a randomized Jacobian-vector approximation; compute sigma_min; multiply the consistency loss by g=min(1,sigma_min/tau) and add the hinge penalty shown above; update network and calibration parameters jointly. 3. Computed from the paper: the residual geometry, shared-target constraint, and excitation/conditioning criterion. Estimated empirically: the Jacobian uses the current neural model, tau is selected from a calibration set, and residual covariance W is estimated from validation noise. 4. First cheap experiment: train a GRU latent dynamics model on synthetic 2-D trajectories with random unknown rotations and translations and a shared hidden landmark. Compare ordinary consistency training, random-window training, and excitation-gated training. Sweep trajectory spread and record calibration error versus sigma_min and kappa. The quantitative prediction is a sharp failure boundary: windows with kappa above roughly 100 should show substantially higher frame error and occasional divergent training, while windows with kappa below 100 should converge reliably; increasing T or trajectory spread should monotonically increase sigma_min and reduce target error.

Verification

Mechanism confirmed, baseline not beaten

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

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 8/10

Built a reproducible 2-D latent-frame calibration MVP with autodiff Jacobians, singular-value excitation scoring, and gated versus ordinary consistency optimization. The mechanism manifested clearly: zero motion gave rank 4 versus rank 5 with motion, the small-motion log-log slope was 0.994 (predicted 1), and sigma_min increased monotonically from 0.501 at T=2 to 0.665 at T=8. However, the toy gated optimizer had effectively identical parameter error to baseline at every excitation level, so this is evidence for the identifiability mechanism rather than a demonstrated training improvement.

Agent confidence: 8/10

Baseline: Mean final theta error for d={0.03,0.1,0.3,1.0}: {2.29565, 2.23181, 1.83331, 0.18794}.
Idea: Mean final theta error for d={0.03,0.1,0.3,1.0}: {2.29565, 2.23181, 1.83331, 0.18794}; no practical improvement over baseline in this isolated optimization test. Core checks: rank d=0=4, rank d=1=5; sigma_min/d median=0.641; log-log slope=0.994; T sweep sigma_min={0.5014,0.5054,0.5340,0.6005,0.6650}.

Limitations: This is a small analytic calibration experiment, not a GRU/world-model integration. The gating comparison optimizes one calibration vector rather than a learned recurrent network and does not test generalization, real sensor noise models, randomized Jacobian approximations, robust losses, or FLOP/speed tradeoffs. The gate is detached from the Jacobian in the toy optimizer, so the result does not establish end-to-end gated training benefit.

How to run: python3 experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.09464) Sep 2, 2026

No significant difference vs. the tuned baseline (+9.4%, p=0.2782, wins 4 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0007
Idea mean: 0.0008
Effect (Δ): 0.0001 (+9.4%; negative = idea better)
Wins: 4 / 8 paired seeds
p-value: 0.2782 (permutation test, 20 000 shuffles)
Smallest detectable effect: n/a
Mechanism: Confirmed ✓
Practical effect: Inconclusive
Baseline tuning: swept over 3 configs

Limitations:

Only the built-in dynamics pendulum track was tested. No multimodal sensor-fusion or learned explicit (p,x,psi) calibration head was evaluated; the gate used a precomputed analytic Jacobian score from observed trajectory states. Only 15 epochs, 400 training samples, one threshold, and one gate floor were tested.

How to run: python3 stage2_bench.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

Implemented excitation-gated recurrent training on the structurally matched dynamics track using the shared rnn_small architecture and a Jacobian smallest-singular-value gate. The mechanism signature was confirmed: low-excitation windows had higher trained-model residual MSE (0.001299 vs 0.000594), but the best gated model was worse than the tuned baseline (MSE 0.000751 vs 0.000687), with paired delta +0.0000644 and permutation p=0.2782. Therefore there was no significant task-metric win.

Baseline: Tuned lr=0.01; full 8-seed mean MSE 0.000686979, std 0.000196252.
Idea: Best lr=0.01; full 8-seed mean MSE 0.000751409, std 0.000237290; paired delta +0.0000644303, 4/8 wins, p=0.2782.
Mechanism signature: q25 sigma=0.0773184, q75 sigma=0.1157134; low-excitation train MSE=0.00129865, high-excitation train MSE=0.000593553; confirmed=true

Benchmark result