✓✓ Beats tuned baseline 2026

Coverage-Controlled Adaptive Time Sampling

Implementation & benchmark of arXiv:2608.29559 — Conformal Prediction Regions for Continuous-Time Trajectories under Random Sampling

Usefulness7/10

Difficulty5/10

Novelty8/10

Dynamics Stability Speedup Accuracy World model Ssm Inference speedup Dynamical systems Control theory Numerical analysis

Source paper: Conformal Prediction Regions for Continuous-Time Trajectories under Random Sampling arXiv:2608.29559 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Use the conformal regularity inflation law as a controller for observation placement or neural-ODE solver refinement. Sample or evaluate the learned dynamics more densely only where the predicted continuous-time uncertainty exceeds a prescribed safety radius, rather than using a uniform time grid.

Formulas

$$r(t)=q_{1-\alpha}+\Gamma\delta(t),\qquad \Gamma=\widehat L+\widehat L_{\mathrm{pred}},\qquad \delta(t)=\min_k|t-t_k|.$$

$$\delta_{\max}=\frac{\varepsilon-q_{1-\alpha}}{\Gamma},\qquad r(t)\le\varepsilon\quad\text{whenever}\quad \delta(t)\le\delta_{\max}.$$

$$\max_k(t_{k+1}-t_k)\le\frac{2(\varepsilon-q_{1-\alpha})}{\Gamma}\quad\Longrightarrow\quad \sup_{t\in[0,T^*]}r(t)\le\varepsilon.$$

Mathematical statement

Let q_(1-alpha) be the conformal residual quantile and Gamma=L_hat+L_hat_pred the sum of an upper bound on true trajectory regularity and a bound on predictor regularity. The tube radius at time t is r(t)=q_(1-alpha)+Gamma delta(t), where delta(t) is the distance to the nearest anchor time. To enforce r(t)<=epsilon, every time must be within delta_max=(epsilon-q_(1-alpha))/Gamma of an anchor. For an interval with endpoint anchors, the worst nearest-anchor distance is half its length, so a sufficient maximum-gap condition is max_k(t_(k+1)-t_k)<=2(epsilon-q_(1-alpha))/Gamma.

Implementation notes

Use this as an inference-time controller for neural-ODE integration, latent-dynamics rollout, sensor acquisition, or teacher-forcing time selection. First train a base predictor normally and calibrate q, L_hat, and L_hat_pred as in the conformal tube method. Choose a safety threshold epsilon larger than q. Start with the available irregular observation times or with the endpoints of a rollout interval. For each interval [t_k,t_(k+1)], use the conservative radius q+Gamma(t_(k+1)-t_k)/2. If it exceeds epsilon, insert a midpoint, evaluate or integrate the model there, and repeat. A priority-queue implementation can always split the interval with the largest current bound. Pseudocode is: while max_interval_length > 2*(epsilon-q)/Gamma, split the longest interval; after each split, recompute local bounds if state-dependent slopes are used. The paper supplies the linear gap-inflation law and regularity-based guarantee; neural predictor slopes are estimated from adjacent solver evaluations. The first cheap experiment is a neural ODE pendulum or Lorenz predictor with fixed end-time accuracy, comparing uniform RK evaluation with adaptive splitting under randomly sampled observations. The predicted sharp signature is a refinement threshold h_c=2(epsilon-q)/Gamma: the maximum tube radius crosses epsilon when the largest gap crosses h_c, while adaptive sampling uses fewer evaluations than a uniform grid and preserves dense-grid coverage near the target level.

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 9/10

Built a deterministic coverage-controlled sampler with longest-gap midpoint refinement and numerical verification of the tube-radius law. The three predicted mechanisms manifested: radius slope Gamma/2 and intercept q matched to below 9e-16 relative error, the safety transition occurred around h_c=2(epsilon-q)/Gamma for every tested Gamma, and required uniform interval counts matched ceil(T/h_c) exactly. Adaptive refinement preserved the target radius in all mini-cases, but did not provide a speedup because retained irregular anchors caused 13–19 evaluations versus 5–14 for a freshly generated uniform grid.

Agent confidence: 9/10

Baseline: Uniform endpoint grid: predicted and observed interval counts matched exactly; mini-case evaluations were [14, 10, 6, 5], with radii [0.1769, 0.2111, 0.3000, 0.3500].
Idea: Adaptive longest-gap refinement: all 4 cases passed the epsilon coverage bound, with evaluations [19, 16, 14, 13] and radii [0.1745, 0.2111, 0.2412, 0.2641]. Mean evaluation reduction versus uniform was -0.973.

Limitations: This is a toy geometric verification, not a trained neural ODE, conformal calibration, sensor-acquisition study, or wall-clock benchmark. Gamma and q were treated as known constants; state-dependent regularity, model error, stochastic observations, and physical trajectory accuracy were not tested. The baseline was a freshly constructed uniform grid, whereas adaptive sampling retained the initial irregular anchors, so the experiment does not establish an inference-time speedup.

How to run: python3 coverage_sampling.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.29559) Sep 2, 2026

Beats the tuned baseline by 13.6% (p=0.0224, wins 7 / 8 paired seeds; metric: lower is better).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0021
Idea mean: 0.0018
Effect (Δ): -0.0003 (−13.6%; negative = idea better)
Wins: 7 / 8 paired seeds
p-value: 0.0224 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±12.4%
Mechanism: Not confirmed ✗
Practical effect: Helps
Baseline tuning: swept over 3 configs

Limitations:

Only the built-in actuated-pendulum dynamics track was tested. No wall-clock or function-evaluation reduction benchmark was performed, and q/Gamma were fixed proxies rather than conformally calibrated estimates; the mechanism signature therefore remains unconfirmed.

How to run: python3 adaptive_sampling_bench.py

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

Stage 2 — Benchmark Sep 2, 2026 latest Worked ✓

Agent confidence: 7/10

Implemented coverage-controlled adaptive temporal resampling on the structurally matched dynamics track using separate end-to-end trained rnn_small systems. Across 8 paired seeds, adaptive resampling reduced test MSE from 0.0021058194 to 0.0018193617, with delta_mean=-0.0002864578 and permutation p=0.0224, meeting the benchmark's significant-win criterion. The mechanism signature was not confirmed quantitatively because the fixed-gap prediction was not calibrated against trained-model residuals.

Baseline: rnn_small, lr=0.006, 12 epochs: mean test MSE 0.0021058194, std 0.0003894366
Idea: CoverageResampler + rnn_small, lr=0.006, 12 epochs: mean test MSE 0.0018193617, std 0.0003773601; 7/8 paired wins; delta=-0.0002864578; p=0.0224

Benchmark result