✗ Failed on benchmark 2026

Wasserstein Speed-Limit Controller

Implementation & benchmark of arXiv:2608.27938 — The thermodynamic freedom of a thermodynamic computer

Usefulness7/10

Difficulty5/10

Novelty7/10

Dynamics Stability Speedup Sample efficiency Optimizer Training dynamics Scheduler Stochastic processes Optimal transport Dynamical systems Control theory

Source paper: The thermodynamic freedom of a thermodynamic computer arXiv:2608.27938 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Wrap stochastic optimization or iterative neural inference in a controller that measures how far the state distribution moves during each interval and compares this motion with the available noise-dependent entropy-production budget. The controller increases the learning rate or reduces inference steps only while the trajectory remains inside the predicted speed-limit region, preventing fast jumps that cause accuracy collapse.

Formulas

$$\dot{x}=\mu f_\theta(x,t)+\sqrt{2\mu k_{\rm B}T_t}\,\xi(t),\qquad D_t=\mu k_{\rm B}T_t$$

$$W_2^2(p_0,p_\tau)\leq D\,\tau\,\Sigma,\qquad \Sigma=\int_0^\tau\sigma_t\,dt$$

$$\sigma_t=\frac{1}{D}\int p_t(x)\left\|\frac{J_t(x)}{p_t(x)}\right\|^2dx,\qquad \tau_{\min}=\frac{W_2^2(p_0,p_\tau)}{D\Sigma}$$

$$\eta_{t+1}=\eta_t\,\operatorname{clip}\left[\left(\frac{\widehat W_{2,t}^2}{D_t\Delta t\,\widehat\Sigma_t}\right)^\alpha,\frac{\eta_{\min}}{\eta_t},\frac{\eta_{\max}}{\eta_t}\right]$$

Mathematical statement

Let p_t(x) be the distribution of network states x, J_t(x) the probability current, v_t(x)=J_t(x)/p_t(x) the probability-flow velocity, D=μ k_B T the diffusion coefficient, and σ_t the entropy-production rate divided by k_B. The continuity equation is ∂_t p_t+∇·(p_t v_t)=0. The Wasserstein action bound implies W_2(p_0,p_τ)^2≤D τ Σ, where Σ=∫_0^τ σ_t dt. Therefore any protocol transporting the distribution over Wasserstein distance W_2 in time τ must satisfy τ≥W_2^2/(DΣ). In a neural implementation, p_t is the empirical distribution of parameter replicas, hidden states, or stochastic predictions; W_2 and Σ are estimated from minibatch samples and consecutive updates.

Implementation notes

Integrate the controller into stochastic gradient descent, SGLD, or a diffusion sampler. For an optimizer experiment, maintain K parameter replicas and update each replica with θ_{t+1}^{(k)}=θ_t^{(k)}−η_t g_t^{(k)}+sqrt(2η_t T_t) z_t^{(k)}, where g_t^{(k)} is the minibatch gradient, T_t is an injected temperature, and z_t^{(k)} is standard Gaussian noise. Every m updates, collect replica states before and after the interval. Estimate W_2 with sliced Wasserstein distance using 32 random projections, and estimate the current velocity using v_k=(x_{t+1}^{(k)}−x_t^{(k)})/Δt. Estimate σ_t by the average squared velocity divided by D_t=μ k_B T_t, using an exponential moving average to reduce noise. Compute r=Ŵ_2^2/(D_t Δt Σ̂_t). If r approaches or exceeds one, reduce η or increase the allowed number of steps; if r is well below one and validation accuracy is stable, cautiously increase η. Known quantities are the injected noise and D_t; Wasserstein displacement, velocity, and entropy production are empirical estimates. The first cheap experiment should use a two-layer MLP on MNIST or Fashion-MNIST with 32 replicas, comparing Adam, fixed-step SGLD, and the controller at matched final accuracy. Plot r, loss, accuracy, and replica displacement. The falsifiable prediction is that stable high-accuracy intervals satisfy r≤1 up to estimation error, while deliberately increasing η creates persistent r>1 before divergence or accuracy collapse. The measured critical learning rate should predict the collapse boundary within 20%.

Verification

Failed on benchmark

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Not reproduced ✗

Stage 2 · Practical benchmark result: Significantly worse ✗

Methodology: 8 paired seeds, permutation test (p<0.05) against a learning-rate-tuned baseline. How verification works

Stage 1 — Mechanism check agent confidence 7/10

Built a replica-based SGLD Wasserstein controller and a toy translating-distribution verification. The bound prediction was confirmed: the dimension-corrected sliced-Wasserstein/action ratio had median 1.00, with worst relative error 28.2%; displacement scaling versus speed²·time² ranged from 0.72 to 1.25. In the quadratic stability sweep, the predicted fixed-step boundary eta=0.5 was observed: baseline diverged at eta=0.55 and 0.70, while the controller reduced eta and remained finite, although it sometimes hurt already-stable low-eta runs.

Agent confidence: 7/10

Baseline: Quadratic SGLD final loss: 1.02e39 at eta=0.55 and 3.42e123 at eta=0.70; stable final losses 0.00480, 0.00605, and 0.01204 at eta=0.10, 0.30, and 0.45.
Idea: Controller final loss: 0.00701 at eta=0.55 and 0.00573 at eta=0.70, with final eta 0.453 and 0.306 respectively; toy bound ratio median 1.000 and speed/time scaling range 0.718–1.249.

Limitations: No MNIST/MLP experiment, wall-clock speed, or matched-FLOP study was performed. The dimension-corrected sliced-Wasserstein quantity is an estimator/proxy rather than an exact full W2 distance, and the optimizer test used a quadratic objective with injected noise rather than a realistic neural-network training workload.

How to run: python3 wasserstein_controller.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.27938) Sep 2, 2026

Significantly WORSE than the tuned baseline (+36118.4%, p=0.0081, wins 0 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0014
Idea mean: 0.4975
Effect (Δ): 0.4961 (+36118.4%; negative = idea better)
Wins: 0 / 8 paired seeds
p-value: 0.0081 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±4007.8%
Mechanism: Not confirmed ✗
Practical effect: Hurts
Baseline tuning: swept over 6 configs

Limitations:

Only the required dynamics track was tested. The intervention used a compact 400-sample, 12-epoch benchmark and did not test tabular, vision, sequence, wall-clock speed, or larger neural networks. The empirical ratio is based on a sliced Wasserstein estimator over flattened parameters and injected-noise dynamics, not exact distributional Wasserstein distance.

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 2, 2026 latest Failed ✗

Agent confidence: 10/10

Implemented the Wasserstein speed-limit controller for the structurally matched dynamics track using the shared rnn_small architecture, with an 8-seed paired evaluation, baseline Adam sweep, shared learning-rate search space, and trained-model mechanism signature. The controller was significantly worse: test MSE 0.497507 versus baseline 0.001374, paired delta +0.496133 with permutation p=0.0081. The predicted speed-limit ratio was not confirmed: observed mean ratio was 12911.47 rather than <=1.

Baseline: Best Adam: lr=0.01, weight_decay=0.0001; 8-seed mean test MSE 0.0013736290, std 0.0004336284.
Idea: Best controller setting among shared rates [0.001, 0.003, 0.01]: lr=0.001, weight_decay=0; 8-seed mean test MSE 0.4975068, std 0.0614269; paired delta +0.4961332, p=0.0081, 0/8 wins.
Mechanism signature: Measured W2^2/(D*dt*Sigma): predicted <=1, observed mean 12911.4681 and mean maximum interval ratio 18444.1357; confirmed=false.

Benchmark result