# Nonreversible latent instanton sampler

- ID: 540
- Canonical URL: https://synthcore.org/idea/540/nonreversible-latent-instanton-sampler
- API JSON: https://synthcore.org/api/idea/540.json
- API Markdown: https://synthcore.org/api/idea/540.md
- Verification status: mech_ok_no_baseline
- Source: [arXiv:2607.05194](https://arxiv.org/abs/2607.05194)
- Category: sampling
- Solves: sample-efficiency, stability
- ML areas: diffusion-sampling, world-model, training
- Math tags: stochastic-processes, dynamical-systems, optimization, geometry
- Ratings: usefulness 6/10; difficulty 6/10; novelty 5/10

## Idea description

Represent a rare transition in a neural latent space by a controlled path whose drift is optimized directly, instead of obtaining it by reversing the relaxation dynamics. Learn a state-dependent mobility or diffusion matrix so that the sampler allocates noise and control effort according to the local stochastic geometry. This should improve generation of low-probability transitions in nonequilibrium world models and reduce the number of failed trajectories.

## Mathematical statement

The paper's reduced stochastic dynamics is $\dot{R}=\mathcal{M}(R)[-\partial_R U(R)]+\sqrt{2T\mathcal{M}(R)}\,\eta(t)$, where $R$ is a scalar reaction coordinate, $U(R)$ is an effective free energy or quasipotential, $\mathcal{M}(R)>0$ is the state-dependent mobility, $T$ is the small-noise scale, and $\eta(t)$ is unit Gaussian white noise with $\langle\eta(t)\eta(t')\rangle=\delta(t-t')$. The transferable point is that the noise amplitude and deterministic drift are coupled through mobility, while nonequilibrium optimal paths need not be time reversals of relaxation. For a learned latent SDE $dz=b_\theta(z,t)dt+\sigma_\theta(z,t)dW_t$, use the Freidlin–Wentzell discretization $S[z]=\frac{1}{2}\sum_{k=0}^{K-1}\Delta t\,(\dot z_k-b_\theta(z_k,t_k))^\top a_\theta(z_k,t_k)^{-1}(\dot z_k-b_\theta(z_k,t_k))$, where $a_\theta=\sigma_\theta\sigma_\theta^\top$, $\dot z_k=(z_{k+1}-z_k)/\Delta t$, and $S$ measures the small-noise action of a path. Optimize a separate reverse transition drift $r_\phi(z,t)$ against this action and endpoint constraints; do not impose $r_\phi=-b_\theta$ or assume reverse paths equal relaxation paths.

## Key formulas

- $$\dot{R}=\mathcal{M}(R)\left[-\partial_{R}U(R)\right]+\sqrt{2T\mathcal{M}(R)}\,\eta(t), \qquad \langle\eta(t)\eta(t')\rangle=\delta(t-t').$$
- $$S[z]=\frac{1}{2}\sum_{k=0}^{K-1}\Delta t\,\bigl(\dot z_k-b_\theta(z_k,t_k)\bigr)^\top a_\theta(z_k,t_k)^{-1}\bigl(\dot z_k-b_\theta(z_k,t_k)\bigr), \qquad a_\theta(z,t)=\sigma_\theta(z,t)\sigma_\theta(z,t)^\top.$$
- $$\mathcal{L}(\theta,\phi)=S[z_\phi]+\lambda_0\|z_\phi(0)-z_A\|^2+\lambda_1\|z_\phi(K)-z_B\|^2+\beta\sum_{k=0}^{K-1}\|z_{\phi,k+1}-z_{\phi,k}-\Delta t\,r_\phi(z_{\phi,k},t_k)\|^2.$$

## Implementation notes

Integrate this at the latent transition module of a world model or at the reverse sampler of a diffusion-like generator. First train or freeze an encoder-decoder $(E,D)$ and a baseline latent drift $b_\theta(z,t)$ on ordinary trajectories. Add a positive diagonal mobility network $M_\psi(z,t)=\operatorname{softplus}(h_\psi(z,t))+\epsilon$ and set $a_\psi(z,t)=2T M_\psi(z,t)$; this is the vector analogue of the paper's scalar noise amplitude $2T\mathcal{M}(R)$. For each source-target pair $(z_A,z_B)$, initialize a $K$-step path by linear interpolation plus noise. Compute $\dot z_k=(z_{k+1}-z_k)/\Delta t$, evaluate the action $S$ above, and update either the path variables directly for 20--100 inner gradient steps or predict the path with a small path network $z_{\phi,k}=g_\phi(z_A,z_B,k)$. Use endpoint penalties with $\lambda_0,\lambda_1$ and optionally the reverse-drift consistency penalty in the third formula. Estimate $a^{-1}$ analytically because the mobility is diagonal; estimate the latent reaction coordinate only if desired, using the norm or a learned one-dimensional projection. The paper supplies the mobility-weighted stochastic form and the non-time-reversal principle; the discretized action and neural parameterization are the ML adaptation. Begin on a 2D double-well or nonequilibrium rotational Langevin system, comparing standard reverse-dynamics sampling against this action-optimized sampler at equal neural evaluations. Track successful endpoint reach rate, mean action, transition probability, and trajectory diversity. A useful signal is at least 2x higher successful transition rate or substantially lower action at equal rollout budget, without worse reconstruction likelihood on ordinary trajectories.

## Verification

- Status: mech_ok_no_baseline
- Mechanism evidence: yes
- Mechanism confirmed: yes
- Practical verdict: harms
- Verdict: Built a 2D nonequilibrium double-well MVP with state-dependent diagonal mobility, mobility-weighted Freidlin–Wentzell action, direct endpoint-constrained path optimization, and a reverse-relaxation baseline. The stochastic moment check matched the prescribed SDE, and action optimization reduced action from 14.27 for the straight path to 1.01 while perturbations increased it to 25.89. In 1,200 stochastic trials, the optimized-path controller reached the target in 98.75% of trials versus 0.17% for reverse relaxation, but the baseline often diverged, so this is a promising toy signal rather than a fair neural-model benchmark.

### Mechanism check

- Verdict: Built a 2D nonequilibrium double-well MVP with state-dependent diagonal mobility, mobility-weighted Freidlin–Wentzell action, direct endpoint-constrained path optimization, and a reverse-relaxation baseline. The stochastic moment check matched the prescribed SDE, and action optimization reduced action from 14.27 for the straight path to 1.01 while perturbations increased it to 25.89. In 1,200 stochastic trials, the optimized-path controller reached the target in 98.75% of trials versus 0.17% for reverse relaxation, but the baseline often diverged, so this is a promising toy signal rather than a fair neural-model benchmark.
- Confidence: 6/10
- Limitations: No encoder-decoder, learned neural drift, or learned mobility network was trained. The baseline reverse dynamics became numerically unstable and receives no analogous stabilizing path controller, so the success-rate gap may overstate the practical advantage. Only one 2D toy system, one seed, one discretization, and one target pair were tested; no FLOP-matched neural experiment or trajectory-diversity analysis was performed.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.0081
- Paired wins: 0/8
- Benchmark verdict: idea worse (significant)

## Artifacts

- [bench_experiment.py](https://synthcore.org/code/1380/bench_experiment.py)
- [bench_report.json](https://synthcore.org/code/1380/bench_report.json)
- [experiment.py](https://synthcore.org/code/1380/experiment.py)
- [report.md](https://synthcore.org/code/1380/report.md)
- [report_bench_2026-09-04T175122.md](https://synthcore.org/code/1380/report_bench_2026-09-04T175122.md)
- [results.json](https://synthcore.org/code/1380/results.json)
- [Download all files as ZIP](https://synthcore.org/download/1380)

## Disclaimer

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