# Action-calibrated cycle-hopping RNN

- ID: 2555
- Canonical URL: https://synthcore.org/idea/2555/action-calibrated-cycle-hopping-rnn
- API JSON: https://synthcore.org/api/idea/2555.json
- API Markdown: https://synthcore.org/api/idea/2555.md
- Verification status: failed_benchmark
- Source: [arXiv:2608.19060](https://arxiv.org/abs/2608.19060)
- Category: dynamics
- Solves: accuracy, stability, sample-efficiency
- ML areas: rnn, ssm, world-model, training-dynamics
- Math tags: dynamical-systems, stochastic-processes, statistical-mechanics, control-theory, bifurcations
- Ratings: usefulness 7/10; difficulty 6/10; novelty 8/10

## Idea description

Build a continuous-time RNN or neural state-space model whose latent dynamics possess two stable periodic attractors representing persistent sequence modes, then inject weak calibrated noise to induce rare transitions between them. Instead of treating mode switching as an arbitrary classifier event, estimate the minimum transition action and tune the noise level or an explicit control input so that the observed switching rate matches the desired rate. This should improve long-horizon multimodal prediction by preserving stable within-mode oscillations while allowing statistically correct mode changes.

## Mathematical statement

Let z(t) in R^d be the latent state, f_theta(z,t) the learned drift, B the noise matrix, D>0 the noise intensity, and W_t standard Brownian motion. The latent model is dz=f_theta(z,t)dt+sqrt(2D)B dW_t. For a candidate path phi(t) joining neighborhoods of two stable limit cycles C_1 and C_2 over duration T, define the Freidlin-Wentzell action S[phi]=(1/4) integral from 0 to T of (dot(phi)-f_theta(phi,t))^T (BB^T)^(-1)(dot(phi)-f_theta(phi,t)) dt, assuming BB^T is positive definite. The activation action Delta S_12 is the infimum of S over all such paths, with initial and final phases on the two cycles optimized. In the weak-noise regime, the transition rate satisfies k_12(D) approximately A_12 exp(-Delta S_12/D), where A_12 is a prefactor. Thus log k_12 is linear in 1/D with slope -Delta S_12, rather than being determined by a static energy barrier. Stability of each learned cycle is monitored from transverse Floquet multipliers of the deterministic discretized flow; every multiplier other than the neutral phase multiplier should have magnitude below one.

## Key formulas

- $$dz=f_{\theta}(z,t)\,dt+\sqrt{2D}\,B\,dW_t$$
- $$S[\phi]=\frac{1}{4}\int_0^T\big(\dot{\phi}(t)-f_{\theta}(\phi(t),t)\big)^T(BB^T)^{-1}\big(\dot{\phi}(t)-f_{\theta}(\phi(t),t)\big)\,dt$$
- $$\Delta S_{12}=\inf_{T,\phi:\,C_1\to C_2}S[\phi],\qquad k_{12}(D)\simeq A_{12}\exp\left(-\frac{\Delta S_{12}}{D}\right)$$
- $$\log k_{12}(D)\simeq \log A_{12}-\frac{\Delta S_{12}}{D}$$

## Implementation notes

1. Integration point: use a continuous-time latent RNN, Neural ODE, or diagonal state-space model with Euler updates z[n+1]=z[n]+dt f_theta(z[n],x[n])+sqrt(2D dt) B epsilon[n], where epsilon[n] is standard Gaussian noise. Decode observations with y[n]=g_theta(z[n]). First create two stable latent cycles using two learned periodic prototype trajectories or initialize the drift from two weakly coupled oscillators; train on sequences containing two recurring modes and occasional mode transitions. Add a regularizer penalizing positive transverse Floquet growth and optionally add an action-calibration loss when transition statistics are labeled. 2. Pseudocode: simulate a deterministic rollout to identify cycles C1 and C2; inject noise; count first-passage transitions between phase-invariant distance thresholds around C1 and C2; estimate k=N_transitions/total_time for each D. Estimate Delta S with a discretized minimum-action path. Initialize a path between closest phase points on C1 and C2, then optimize its coordinates using Adam on the discrete action sum dt/4 times the squared Mahalanobis norm of (phi[n+1]-phi[n])/dt-f_theta(phi[n]), reparameterizing the path by arc length after each optimization block. 3. Computed from the mechanism: the action functional, exponential rate law, and phase-invariant endpoint condition. Estimated empirically: cycle locations, Floquet multipliers, prefactor A, and finite-noise rates. 4. First cheap experiment: train a two-dimensional Neural ODE or GRU on a synthetic mixture of two noisy periodic signals, with each mode generated by a different stable oscillator and transitions occurring at a known hazard. Compare a standard GRU, a deterministic two-mode model, and the action-calibrated model at horizons ten times the training window. Sweep D over at least five values and fit log k versus 1/D. The falsifiable prediction is a clear linear Arrhenius regime whose fitted slope agrees with the minimum-action estimate within 20%; transverse Floquet multipliers should remain below one, and long-horizon within-mode phase error should grow more slowly than in an unconstrained noisy RNN. When D becomes large enough that typical noise displacement per cycle exceeds basin width, the linear rare-event law should fail and switching should become frequent.

## Verification

- Status: failed_benchmark
- Mechanism evidence: yes
- Mechanism confirmed: no
- Practical verdict: no_effect
- Verdict: Built and numerically verified a two-cycle stochastic continuous-time latent system with action estimation, discrete minimum-action optimization, Floquet stability checks, and transition-rate sweeps. The action estimate agreed with the finite-endpoint integral within 0.07%, transverse Floquet multipliers were 0.0432 for both cycles, and the rare-noise Arrhenius slope was -0.01481 versus the predicted -0.015625 (5.2% error, R²=0.999). At high noise the fitted slope shifted to -0.0547, confirming breakdown of the rare-event law; this validates the mechanism in the toy setting, but not neural sequence prediction.

### Mechanism check

- Verdict: Built and numerically verified a two-cycle stochastic continuous-time latent system with action estimation, discrete minimum-action optimization, Floquet stability checks, and transition-rate sweeps. The action estimate agreed with the finite-endpoint integral within 0.07%, transverse Floquet multipliers were 0.0432 for both cycles, and the rare-noise Arrhenius slope was -0.01481 versus the predicted -0.015625 (5.2% error, R²=0.999). At high noise the fitted slope shifted to -0.0547, confirming breakdown of the rare-event law; this validates the mechanism in the toy setting, but not neural sequence prediction.
- Confidence: 8/10
- Limitations: This is a low-dimensional radial toy system rather than a trained GRU, Neural ODE, or world model. It does not test long-horizon multimodal prediction, sample efficiency, learned action calibration, prefactor estimation, anisotropic noise, or comparison against a standard RNN. The transition estimator uses finite-step Euler simulation and a separatrix threshold, and the action optimization uses fixed duration and endpoints.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.82325
- Paired wins: 4/8
- Benchmark verdict: no significant win

## Artifacts

- [bench_report.json](https://synthcore.org/code/986/bench_report.json)
- [bench_stage2.py](https://synthcore.org/code/986/bench_stage2.py)
- [experiment.py](https://synthcore.org/code/986/experiment.py)
- [report.md](https://synthcore.org/code/986/report.md)
- [report_bench_2026-09-01T174400.md](https://synthcore.org/code/986/report_bench_2026-09-01T174400.md)
- [results.json](https://synthcore.org/code/986/results.json)
- [Download all files as ZIP](https://synthcore.org/download/986)

## Disclaimer

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