# Dual-Ensemble Latent Transition Model

- ID: 1149
- Canonical URL: https://synthcore.org/idea/1149/dual-ensemble-latent-transition-model
- API JSON: https://synthcore.org/api/idea/1149.json
- API Markdown: https://synthcore.org/api/idea/1149.md
- Verification status: failed_benchmark
- Source: [arXiv:2607.19452](https://arxiv.org/abs/2607.19452)
- Category: dynamics
- Solves: accuracy, stability, sample-efficiency
- ML areas: world-model, ssm, rnn, training-dynamics
- Math tags: stochastic-processes, dynamical-systems, statistics, linear-algebra, control-theory
- Ratings: usefulness 7/10; difficulty 6/10; novelty 8/10

## Idea description

Train a latent recurrent or state-space model with separate equilibrium and source-sink transition matrices instead of forcing one transition matrix to explain all latent dynamics. Use the equilibrium matrix for stationary occupancy and reversible statistics, and use a recycling matrix for directed hitting times, committors, and source-to-target flow; this should remove fixed-lag coarse-graining bias in latent world models.

## Mathematical statement

Let $z_t$ be a discrete or softly clustered latent state with states $i,j\in\{1,\ldots,K\}$. The equilibrium transition matrix is $P_{ij}=\Pr(z_{t+\tau}=j\mid z_t=i)$ under the equilibrium data distribution, while the source-sink recycling matrix is $Q_{ij}=\Pr(z_{t+\tau}=j\mid z_t=i)$ under trajectories initialized in source set $A$ and reset to the source after first entering sink set $B$. Let $\pi$ satisfy $\pi^\top P=\pi^\top$ and let $\tilde\pi$ satisfy $\tilde\pi^\top Q=\tilde\pi^\top$. For membership weights $C_{i\alpha}\geq0$ assigning microscopic state $\alpha$ to coarse state $i$, the unbiased coarse equilibrium operator is $P^c_{ij}=\frac{\sum_{\alpha,\beta}C_{i\alpha}\pi_\alpha P_{\alpha\beta}C_{j\beta}}{\sum_\alpha C_{i\alpha}\pi_\alpha}$, and the source-sink operator is $Q^c_{ij}=\frac{\sum_{\alpha,\beta}C_{i\alpha}\tilde\pi_\alpha Q_{\alpha\beta}C_{j\beta}}{\sum_\alpha C_{i\alpha}\tilde\pi_\alpha}$. The transferable property is ensemble matching: $P^c$ estimates equilibrium observables, whereas $Q^c$ estimates recycling or nonequilibrium observables.

## Key formulas

- $$P_{ij}=\Pr(z_{t+\tau}=j\mid z_t=i),\qquad Q_{ij}=\Pr(z_{t+\tau}=j\mid z_t=i,\ \text{source-sink recycling}),$$
- $$P^c_{ij}=\frac{\sum_{\alpha,\beta}C_{i\alpha}\pi_\alpha P_{\alpha\beta}C_{j\beta}}{\sum_\alpha C_{i\alpha}\pi_\alpha},\qquad Q^c_{ij}=\frac{\sum_{\alpha,\beta}C_{i\alpha}\tilde\pi_\alpha Q_{\alpha\beta}C_{j\beta}}{\sum_\alpha C_{i\alpha}\tilde\pi_\alpha},$$
- $$\pi^\top P=\pi^\top,\qquad \tilde\pi^\top Q=\tilde\pi^\top,$$
- $$\mathcal{L}=\mathcal{L}_{\mathrm{eq}}(P_\theta)+\lambda_{\mathrm{ss}}\mathcal{L}_{\mathrm{ss}}(Q_\theta)+\lambda_{\mathrm{match}}\left\|\hat\pi_\theta-\mathrm{stopgrad}(\hat\pi_{\mathrm{data}})\right\|_2^2,$$

## Implementation notes

1. Integration point: insert the module between an encoder and a latent transition model. Encode each observation as soft memberships $c_t=\mathrm{softmax}(f_\theta(x_t))\in\mathbb{R}^K$. Use one transition-logit head $g_P(h_t)$ for ordinary trajectories and a second head $g_Q(h_t,a_t,b_t)$ for trajectories carrying source and sink labels. Apply a row-wise softmax to obtain $P_\theta$ and $Q_\theta$. 2. Pseudocode: collect equilibrium pairs $(c_t,c_{t+\tau})$ and accumulate $N^P_{ij}\mathrel{+}=c_{t,i}c_{t+\tau,j}$; set $P_{ij}=N^P_{ij}/(\sum_jN^P_{ij}+\epsilon)$ and optimize transition cross-entropy. Generate source-sink rollouts; whenever a rollout first enters $B$, record its transition and reset its latent state to a source sample. Accumulate $N^Q$ identically and fit $Q$. At evaluation, obtain stationary vectors by power iteration $v_{n+1}=v_nP$ or $v_{n+1}=v_nQ$, selecting the operator according to the observable. 3. Taken from the paper: two ensemble-specific transition matrices, stationary weighted averaging, and recycling semantics. Estimated empirically: memberships, counts, stationary vectors, and bootstrap uncertainty. 4. First cheap experiment: construct the paper's eight-state Markov chain, merge adjacent microscopic states into four coarse states, and compare a single-matrix latent model with this dual model at lag $\tau=1$. Repeat on a small grid-world or MuJoCo dataset with source and target regions. Measure stationary occupancy, mean first-passage time, committors, and directed flux. The falsifiable prediction is that the dual model's coarse stationary vector converges to the aggregated microscopic vector $(18,12,6,1)/37\approx(0.486,0.324,0.162,0.027)$, whereas the single-matrix model retains a fixed-lag bias. Increasing data should reduce dual-model error approximately as $N^{-1/2}$ without requiring a longer lag.

## Verification

- Status: failed_benchmark
- Mechanism evidence: yes
- Mechanism confirmed: no
- Practical verdict: inconclusive
- Verdict: Built a dual-operator finite-state MVP with equilibrium matrix P and source-sink recycling matrix Q, plus stationary-weighted coarse-graining and MFPT evaluation. Matched operators recovered their respective coarse stationary distributions across lags 1–32 to about 1e-15 error, and the recycling MFPT matched the direct first-passage value to 3.8e-11 while the equilibrium-only estimate had errors around 797–806 steps. Finite-sample dual occupancy error decreased with N, although the fitted slope was -0.368 rather than the predicted -0.5, so that scaling claim is only partially supported.

### Mechanism check

- Verdict: Built a dual-operator finite-state MVP with equilibrium matrix P and source-sink recycling matrix Q, plus stationary-weighted coarse-graining and MFPT evaluation. Matched operators recovered their respective coarse stationary distributions across lags 1–32 to about 1e-15 error, and the recycling MFPT matched the direct first-passage value to 3.8e-11 while the equilibrium-only estimate had errors around 797–806 steps. Finite-sample dual occupancy error decreased with N, although the fitted slope was -0.368 rather than the predicted -0.5, so that scaling claim is only partially supported.
- Confidence: 8/10
- Limitations: Synthetic eight-state Markov chain only; no learned encoder, soft memberships, neural training, optimization comparison, wall-clock benchmark, grid-world/MuJoCo data, or joint P/Q learning. The recycling operator is specified from exact dynamics rather than learned from raw source-sink rollouts, and the finite-sample scaling sweep is limited by rare-sink effects and four sample sizes.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.7975
- Paired wins: 4/8
- Benchmark verdict: no significant win

## Artifacts

- [REPORT.md](https://synthcore.org/code/458/REPORT.md)
- [bench_dual.py](https://synthcore.org/code/458/bench_dual.py)
- [bench_report.json](https://synthcore.org/code/458/bench_report.json)
- [dual_ensemble_experiment.py](https://synthcore.org/code/458/dual_ensemble_experiment.py)
- [report.md](https://synthcore.org/code/458/report.md)
- [report_bench_2026-08-31T124819.md](https://synthcore.org/code/458/report_bench_2026-08-31T124819.md)
- [Download all files as ZIP](https://synthcore.org/download/458)

## Disclaimer

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