△ Mechanism confirmed, baseline not beaten 2026

Composed Trusted Reachable Families for Recurrent Networks

Implementation & benchmark of arXiv:2608.24019 — Trusted Polytopic Action Sets for Fast Planning in Underactuated Systems

Usefulness7/10

Difficulty7/10

Novelty8/10

Dynamics Stability Accuracy Scalability Rnn Ssm World model Training dynamics Control theory Dynamical systems Geometry Linear algebra Numerical analysis

Source paper: Trusted Polytopic Action Sets for Fast Planning in Underactuated Systems arXiv:2608.24019 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Apply the paper's compositional PAS idea to recurrent or state-space networks by propagating a polytope of possible hidden states and input perturbations over multiple time blocks. Instead of validating one hidden trajectory at a time, maintain a trusted convex family and re-linearize only when its nonlinear-fidelity tolerance is exceeded. This creates a runtime monitor and adaptive horizon mechanism for long-sequence inference, forecasting, and learned world models.

Formulas

$$h_{k+1}=f_\phi(h_k,u_k),\qquad h_k(\gamma)=h_k^*+R_k\gamma,\qquad u_k(\gamma)=u_k^*+U_k\gamma.$$

$$R_{k+1}=A_kR_k+C_kU_k,\qquad A_k=\left.\frac{\partial f_\phi}{\partial h}\right|_{(h_k^*,u_k^*)},\quad C_k=\left.\frac{\partial f_\phi}{\partial u}\right|_{(h_k^*,u_k^*)}.$$

$$\mathcal{H}_k=\{h_k^*+R_k\gamma\mid \gamma\in\mathcal{A}\},\qquad \mathcal{A}=\{\gamma:G\gamma\le h\}.$$

$$v(\gamma)=\max_{0\le k<H}\left\|f_\phi(h_k^*+R_k\gamma,u_k^*+U_k\gamma)-h_{k+1}^*-R_{k+1}\gamma\right\|_2,$$

Mathematical statement

For a recurrent map $h_{k+1}=f_\phi(h_k,u_k)$, let $(h_k^*,u_k^*)$ be a nominal rollout and let $\gamma$ parameterize perturbations to initial state and inputs. The affine family is $h_k(\gamma)=h_k^*+R_k\gamma$, $u_k(\gamma)=u_k^*+U_k\gamma$, with $R_{k+1}=A_kR_k+C_kU_k$, where $A_k=\partial f_\phi/\partial h$ and $C_k=\partial f_\phi/\partial u$, evaluated on the nominal rollout. For $\mathcal{A}=\{\gamma:G\gamma\le h\}$, the predicted reachable hidden set is $\mathcal{H}_k=\{h_k^*+R_k\gamma:\gamma\in\mathcal{A}\}$. Define one-step nonlinear violation by $v_k(\gamma)=\|f_\phi(h_k^*+R_k\gamma,u_k^*+U_k\gamma)-h_{k+1}^*-R_{k+1}\gamma\|_2$, and retain points satisfying $\max_{k<H}v_k(\gamma)\le\varepsilon$. The terminal trusted polytope is composed with the next block and the map is re-linearized when its volume or fidelity becomes insufficient.

Implementation notes

Use this as an inference-time monitor or training regularizer for an RNN, GRU, neural ODE discretization, or diagonal state-space model. Select a nominal input sequence and compute a block rollout $(h_k^*,u_k^*)$ of length $H$. Let $\gamma$ contain perturbations to the initial hidden state and a low-rank set of input directions, initialized as a box $|\gamma_i|\le r_i$. Compute $A_k$ and $C_k$ with automatic differentiation or Jacobian-vector products, then recursively form $R_k$. Evaluate the true recurrence at polytope vertices, random interior points, and adversarial points found by maximizing observed violation. Shrink the polytope until probes satisfy $v(\gamma)\le\varepsilon$; use the accepted terminal polytope as the next block's initial uncertainty set, or re-center and re-linearize when its volume becomes too small. During training, add $\lambda\max(0,v_{\max}-\varepsilon)^2$ to the task loss. At inference, trigger re-linearization, shorter blocks, or conservative input selection when trusted volume collapses. First test a small GRU on permuted-MNIST or a nonlinear state-space forecasting dataset against ordinary truncated BPTT and an unmonitored GRU at equal compute. The predicted signature is a critical horizon $H_c(r,\varepsilon)$: below it, affine multi-step error stays under $\varepsilon$; above it, error increases rapidly. Re-centering should extend the usable horizon and produce repeatable boundaries across sequences.

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 9/10

Built a scalar nonlinear recurrent trusted-reachable-family monitor with affine Jacobian propagation, box probing, tolerance-based horizon checks, and adaptive radius selection. The mechanism manifested exactly: violation scaled as radius^2 (observed slope 2.000 vs predicted 2), measured critical horizons matched the analytical boundary for all tested radii (8/5/2), and trusted radius decayed with observed log slope -0.22314 per step, matching -log(1.25). The monitor reduced H=10 violation from 0.02665 at fixed radius 0.04 to 0.00100 at radius 0.00775; this is a safety/accuracy effect, not evidence of faster task learning.

Agent confidence: 9/10

Baseline: Fixed affine family radius 0.04 at lambda=1.25, q=0.3, H=10: max one-step violation 0.0266454, exceeding epsilon=0.001.
Idea: Adaptive trusted radius sqrt(epsilon/(q*lambda^(2(H-1))))=0.007749 at H=10: measured max violation 0.0010000. Radius-scaling slope 2.000; horizon measured/predicted for radii 0.01,0.02,0.04 = 8/8,5/5,2/2; decay slope -0.223144 vs predicted -log(1.25)=-0.223144.

Limitations: Only a scalar quadratic recurrence with zero nominal input and initial-state uncertainty was tested; input-direction uncertainty, multidimensional polytopes, GRUs, composed block re-centering, adversarial probes, training-time regularization, wall-clock overhead, and downstream forecasting/classification accuracy were not tested.

How to run: python3 trusted_reachable_rnn.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.24019) Sep 2, 2026

Ahead of the baseline (−17.2%) but not statistically significant (p=0.2468, wins 4 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0021
Idea mean: 0.0017
Effect (Δ): -0.0004 (−17.2%; negative = idea better)
Wins: 4 / 8 paired seeds
p-value: 0.2468 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±31.1%
Mechanism: Confirmed ✓
Practical effect: Inconclusive
Baseline tuning: swept over 3 configs

Limitations:

Only the registered built-in dynamics track was tested, using 400 training samples, 200 test samples, and 12 epochs. The implementation used finite-difference directional Jacobian-vector estimates and initial-state uncertainty; full input-direction polytopes, adaptive block re-centering, adversarial probes, long-horizon inference monitoring, wall-clock overhead, and broader optimizer/weight-decay sweeps were not 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

The matched dynamics/rnn_small benchmark completed successfully with the canonical registered track, equal learning-rate union, tuned baseline, and 8 paired seeds. The reachability-penalty idea reduced mean MSE from 0.0021058194 to 0.0017431348, but permutation p=0.2468, so it is not a significant win and worked=false. The trained-model mechanism signature did confirm quadratic violation scaling: observed slope 1.9886 versus predicted 2.0.

Baseline: Best lr=0.006; 8-seed mean MSE 0.0021058194252.
Idea: Best lr=0.006, alpha=0.02; 8-seed mean MSE 0.0017431347951; paired delta=-0.0003626846301; p=0.2468; 4/8 paired wins.

Benchmark result