△ Mechanism confirmed, baseline not beaten 2026

Path-Space Boundary Screening Regularizer

Implementation & benchmark of arXiv:2607.13391 — The nonequilibrium statistical mechanics of Markov interacting particles

Usefulness8/10

Difficulty5/10

Novelty6/10

Regularization Stability Generalization Sample efficiency Rnn Ssm World model Regularization Probability Information theory Stochastic processes

Source paper: The nonequilibrium statistical mechanics of Markov interacting particles arXiv:2607.13391 ⓘ · analyzed Aug 30, 2026

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 Not peer-reviewed.

Idea description

Train a sequential model with an explicit boundary state B so that exterior history Y and interior history X become conditionally independent given the entire boundary history, not merely given the current boundary value. Penalize estimated conditional mutual information from conditional sequence likelihoods; this should remove hidden temporal feedback and improve modular long-horizon prediction.

Formulas

$$P(dy_{0:T},dx_{0:T}\mid b_{0:T})=P(dy_{0:T}\mid b_{0:T})P(dx_{0:T}\mid b_{0:T})$$

$$I(Y_{0:T};X_{0:T}\mid B_{0:T})=\mathbb{E}\left[\log\frac{p(y_{0:T},x_{0:T}\mid b_{0:T})}{p(y_{0:T}\mid b_{0:T})p(x_{0:T}\mid b_{0:T})}\right]$$

$$\mathcal{L}(\theta)=\mathcal{L}_{\mathrm{pred}}(\theta)+\lambda\widehat{I}_{\theta}(Y_{0:T};X_{0:T}\mid B_{0:T})$$

Mathematical statement

Let P be the joint law of exterior, boundary, and interior trajectories (Y_0:T, B_0:T, X_0:T). The desired path-space boundary property is P(dy_0:T, dx_0:T | b_0:T) = P(dy_0:T | b_0:T) P(dx_0:T | b_0:T). Its violation is conditional mutual information I(Y_0:T; X_0:T | B_0:T) = E[log p(y_0:T, x_0:T | b_0:T) - log p(y_0:T | b_0:T) - log p(x_0:T | b_0:T)], equivalently a conditional KL divergence. Optimize prediction loss plus lambda times this quantity. The key property is that the penalty detects delayed feedback, hidden innovations, and memory variables even when instantaneous I(Y_t; X_t | B_t) is small.

Implementation notes

Use a recurrent state-space model or transformer with three designated streams. The boundary stream B_t receives the information intended to pass between exterior stream Y_t and interior stream X_t. Implement three conditional likelihood heads: a joint head for p_theta(Y_0:T, X_0:T | B_0:T), an exterior head for p_theta(Y_0:T | B_0:T), and an interior head for p_theta(X_0:T | B_0:T). The marginal heads must not access the opposite stream. For each teacher-forced trajectory, compute log_joint, log_y, and log_x, then set cmi_hat = mean(log_joint - log_y - log_x) and loss = prediction_loss + lambda times cmi_hat. Use a slowly updated copy of the marginal heads or stop-gradient through their targets if optimization is unstable. The paper supplies the factorization and KL/CMI identity; the sequence likelihoods are estimated by the neural models. As a first cheap experiment, create a synthetic delayed-channel system where Y_t and X_t are independent given B_t but X_{t-1} influences Y_t. Compare a GRU with instantaneous CMI regularization against path-CMI regularization. Sweep sequence length T and lambda. The prediction is that instantaneous CMI remains near zero while path CMI is positive, and path CMI decreases monotonically with lambda. On a genuinely screened system, estimated path CMI should reach the finite-sample noise floor; on the delayed system it should remain positive without the regularizer. Long-horizon error should correlate with residual path CMI.

Verification

Mechanism confirmed, baseline not beaten

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

Stage 2 · Practical benchmark result: No measurable effect

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

Stage 1 — Mechanism check agent confidence 9/10

Built a reproducible analytic MVP of path-space boundary screening for a delayed Gaussian channel, with exact and empirical CMI checks plus regularization and long-horizon-risk sweeps. The mechanism manifested clearly: instantaneous CMI was approximately 1.7e-6 versus exact 0, path CMI was 9.983 versus exact 9.976, and path CMI grew linearly with horizon with observed slope 1.24706 matching prediction 1.24706. The half-suppression transition followed the predicted 1/T scaling within about 0.5%; however, suppressing the genuine delayed dependency worsened forecast MSE, so this demonstrates the screening mechanism rather than a predictive win.

Agent confidence: 9/10

Baseline: lambda=0: path CMI=9.9765, forecast MSE=0.0900, coefficient=1.0000
Idea: lambda=0.5: path CMI=4.6250, forecast MSE=0.4005, coefficient=0.4427; lambda=1.0: path CMI=0.6787, forecast MSE=0.8486, coefficient=0.1290. Predicted/observed half-effect lambda for T=2,4,8,16,32: 1.889/1.884, 0.944/0.945, 0.472/0.474, 0.236/0.236, 0.118/0.118.

Limitations: This is an analytically controlled scalar toy and not a neural recurrent model with separately parameterized joint, exterior, and interior likelihood heads. It does not test optimization stability, finite-sample neural CMI estimation, genuinely screened systems, FLOPs, or a learned boundary stream; the forecast comparison intentionally reveals the cost of penalizing real delayed feedback.

How to run: python3 path_screening_experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.13391) Sep 2, 2026

Ahead of the baseline (−0.0%) but not statistically significant (p=0.5201, wins 4 / 8 paired seeds).

Benchmark: Sequence modeling
Model: transformer_tiny
Paired seeds: 8
Baseline mean: 0.572
Idea mean: 0.5718
Effect (Δ): -0.0002 (−0.0%; negative = idea better)
Wins: 4 / 8 paired seeds
p-value: 0.5201 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±0.1%
Mechanism: Confirmed ✓
Practical effect: No effect
Baseline tuning: swept over 3 configs

Limitations:

The built-in sequence track provides scalar forecast windows rather than explicit Y/B/X streams, so the intervention uses a two-half boundary proxy and nonlinear cross-half interaction penalty rather than separately parameterized joint, exterior, and interior conditional likelihood heads. No vision, tabular, dynamics, or custom track was tested; runtime/speed and longer-horizon rollout metrics were not evaluated.

How to run: python3 path_screen_bench.py

Verdict computed by deterministic test code from paired-seed statistics — not by the language model.

Stage 2 — Benchmark Aug 30, 2026 latest Failed ✗

Agent confidence: 9/10

Implemented the path-space screening intervention as a local training-loss modification on the structurally matched sequence forecast track, using the shared transformer architecture and paired-seed protocol. The Gaussian sanity check confirmed near-zero instantaneous MI (9.33e-7) and positive path MI (9.969 versus exact 9.976), while the trained-model interaction signature showed only a 0.23% reduction. Test MSE was effectively unchanged: delta=-0.000195 with permutation p=0.52005, so the benchmark did not establish a significant win.

Baseline: Best lr=0.0015; 8-seed test MSE mean=0.57203097, std=0.09826160
Idea: Best lr=0.0015, lambda=0.001; 8-seed test MSE mean=0.57183603, std=0.09824222; paired delta=-0.00019494, p=0.52005
Mechanism signature: Trained-model mean absolute cross-half interaction: baseline=0.05617786, idea=0.05604979, reduction=0.22798%, confirmed=true

Benchmark result