# Target-Law Neural Stopping

- ID: 2543
- Canonical URL: https://synthcore.org/idea/2543/target-law-neural-stopping
- API JSON: https://synthcore.org/api/idea/2543.json
- API Markdown: https://synthcore.org/api/idea/2543.md
- Verification status: failed_benchmark
- Source: [arXiv:2608.18783](https://arxiv.org/abs/2608.18783)
- Category: sampling
- Solves: sample-efficiency, stability, speedup
- ML areas: diffusion-sampling, rl, inference-speedup
- Math tags: stochastic-processes, probability, measure-theory, optimization
- Ratings: usefulness 7/10; difficulty 5/10; novelty 7/10

## Idea description

Learn a path-dependent stopping policy for a stochastic neural trajectory so that the state at stopping time matches a prescribed target distribution, instead of optimizing only a scalar terminal reward. This can turn a fixed-length diffusion sampler or iterative latent refinement process into an adaptive sampler that stops early when its sample distribution is already sufficiently close to the target.

## Mathematical statement

The paper formulates a distribution builder through a stopping time \(\tau\) for a diffusion \(X_t\), with the requirement that the stopped state has target law \(\mathcal{L}(X_\tau)=\nu\), where \(\nu\) is the desired probability distribution. The underlying process is \(\mathrm{d}X_t=\mu(X_t)\,\mathrm{d}t+\sigma(X_t)\,\mathrm{d}W_t\), where \(\mu\) is the drift, \(\sigma\) is the diffusion coefficient, and \(W_t\) is Brownian motion. The Skorokhod-embedding viewpoint says that the central object is not a fixed terminal time but a stopping rule realizing \(\nu\). For a neural implementation, represent the stopping rule by a conditional hazard \(h_\theta(x_t,t)\ge 0\); in a discretization with step \(\Delta t\), the probability of stopping at step \(k\) is \(q_k=1-\exp[-h_\theta(x_k,k\Delta t)\Delta t]\), and the survival probability is \(S_k=\prod_{j<k}(1-q_j)\). The induced terminal law is the mixture \(\widehat\nu_\theta=\sum_k S_k q_k\,\mathcal{L}(X_k)+S_{K+1}\mathcal{L}(X_{K+1})\). Train \(\theta\) by minimizing a differentiable discrepancy between \(\widehat\nu_\theta\) and \(\nu\), plus a small expected-computation penalty. The mathematical guarantee being approximated is law matching at a stopping time; the neural method estimates the law empirically using minibatch samples and uses a soft hazard rather than an exact Skorokhod barrier.

## Key formulas

- $$\mathrm{d}X_t=\mu(X_t)\,\mathrm{d}t+\sigma(X_t)\,\mathrm{d}W_t,$$
- $$\mathcal{L}(X_\tau)=\nu,$$
- $$q_k=1-\exp\big[-h_\theta(X_k,k\Delta t)\Delta t\big],\qquad S_k=\prod_{j=0}^{k-1}(1-q_j),$$
- $$\widehat{\nu}_\theta=\sum_{k=0}^{K}S_kq_k\,\mathcal{L}(X_k)+S_{K+1}\mathcal{L}(X_{K+1}),\qquad \mathcal{J}(\theta)=D\!\left(\widehat{\nu}_\theta,\nu\right)+\lambda\sum_{k=0}^{K}S_k\Delta t.$$

## Implementation notes

(1) Integration point: modify a diffusion sampler, stochastic latent refinement loop, or iterative RL environment. At each sampler step, expose the current state tensor \(x_k\), timestep embedding \(k/K\), and optionally the model score or denoiser features to a small stopping head \(h_\theta(x_k,k)\). The base sampler continues to produce \(x_{k+1}\); the stopping head controls when the current sample is emitted. Do not use a hard sampled stop during training: retain the full survival-weighted mixture so gradients reach all steps. (2) Pseudocode: initialize \(S=1\), \(\widehat{L}=0\); for \(k=0,\ldots,K\), compute \(x_k\), set \(q=1-\exp(-\operatorname{softplus}(h_\theta(x_k,k))\Delta t)\), add \(S q\,\phi(x_k)\) to a batch statistic, then update \(S\leftarrow S(1-q)\); at the final step add \(S\phi(x_K)\). Here \(\phi\) can be random Fourier features, histogram bins, classifier features, or a critic trained to distinguish generated stopped samples from target samples. Minimize the discrepancy between the accumulated feature mean and target feature mean, plus \(\lambda\sum_k S\Delta t\). At inference, sample a Bernoulli stop using \(q_k\), or stop when the cumulative hazard exceeds a sampled exponential clock. (3) Computed from the paper's mathematics: the stopping-time law constraint, survival mixture, and expected stopping-time penalty. Estimated empirically: \(D\), target moments/features, and the state distributions at each step. For a diffusion model, choose \(\nu\) as the training-data distribution in image-feature space or as the distribution of full-budget samples; for RL, choose a desired return/state distribution. (4) First experiment: use a 2D Ornstein-Uhlenbeck or Brownian sampler with a known Gaussian target, then CIFAR-10 DDIM/ancestral diffusion with a 10-step, 20-step, and 50-step budget. Compare fixed-step sampling and an entropy-regularized learned early-exit head at equal FID and equal sample count. Success is lower mean function evaluations at matched FID or lower target-law MMD at matched compute; also verify calibration by plotting empirical \(\mathbb{P}(\tau\le k)\) and the stopped-state discrepancy across checkpoints.

## Verification

- Status: failed_benchmark
- Mechanism evidence: yes
- Mechanism confirmed: no
- Practical verdict: harms
- Verdict: Built a differentiable survival-mixture stopping-policy MVP for Brownian trajectories and verified its core equations numerically. Survival slopes matched predicted values to below 5e-16, mixture mass error was 2.2e-16, and discretization error halved when the time step was halved. The learned policy reduced toy target-law feature MSE from 0.3263 to 0.000548 while reducing expected compute from 3.0 to 1.147 time units.

### Mechanism check

- Verdict: Built a differentiable survival-mixture stopping-policy MVP for Brownian trajectories and verified its core equations numerically. Survival slopes matched predicted values to below 5e-16, mixture mass error was 2.2e-16, and discretization error halved when the time step was halved. The learned policy reduced toy target-law feature MSE from 0.3263 to 0.000548 while reducing expected compute from 3.0 to 1.147 time units.
- Confidence: 8/10
- Limitations: Only a small CPU toy Brownian experiment was tested. The fitted hazard is time-dependent rather than state/path-dependent, no real diffusion or neural sampler was used, the discrepancy is handcrafted feature MSE rather than MMD or FID, and no multiple-seed statistical study was performed.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.01555
- Paired wins: 1/8
- Benchmark verdict: idea worse (significant)

## Artifacts

- [bench_report.json](https://synthcore.org/code/976/bench_report.json)
- [bench_target_law.py](https://synthcore.org/code/976/bench_target_law.py)
- [experiment.py](https://synthcore.org/code/976/experiment.py)
- [report.md](https://synthcore.org/code/976/report.md)
- [report_bench_2026-09-01T171847.md](https://synthcore.org/code/976/report_bench_2026-09-01T171847.md)
- [results.json](https://synthcore.org/code/976/results.json)
- [Download all files as ZIP](https://synthcore.org/download/976)

## Disclaimer

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