△ Mechanism confirmed, baseline not beaten 2026

Topological Response Basis Layer

Implementation & benchmark of arXiv:2607.12096 — Topological building blocks of nonequilibrium response

Usefulness7/10

Difficulty6/10

Novelty8/10

Architecture Accuracy Stability Generalization Mlp World model Training dynamics Stochastic processes Graph theory Linear algebra Dynamical systems

Source paper: Topological building blocks of nonequilibrium response arXiv:2607.12096 ⓘ · analyzed Aug 30, 2026

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 Not peer-reviewed.

Idea description

Insert a small continuous-time Markov latent module between a neural encoder and decoder, with input-dependent transition rates and a fixed library of graph topologies such as directed cycles, reversible chains, and branching motifs. The output is an observable of the stationary distribution, while a learned convex mixture over topology-specific response curves constrains the network to represent responses as combinations of interpretable nonequilibrium mechanisms.

Formulas

d p(t)/dt = W(x) p(t), with W_ij(x) >= 0 for i != j and W_ii(x) = -sum_{j != i} W_ji(x)

W(x) pi(x) = 0, 1^T pi(x) = 1, R(x) = g^T pi(x)

W(x) pi'(x) = -W'(x) pi(x), 1^T pi'(x) = 0, pi'(x) = -W#(x) W'(x) pi(x)

R(x) = sum_{k=1}^K alpha_k R_k(x), alpha_k >= 0, sum_{k=1}^K alpha_k = 1

Mathematical statement

For a strongly connected directed graph with states i = 1,...,N, let W(x) be its generator at scalar input x, with off-diagonal rates W_ij(x) >= 0 for i != j and diagonal entries W_ii(x) = -sum_{j != i} W_ji(x). The stationary distribution pi(x) satisfies W(x) pi(x) = 0 and 1^T pi(x) = 1. For observable values g_i, the response is R(x) = g^T pi(x). Differentiation gives W(x) pi'(x) = -W'(x) pi(x) and 1^T pi'(x) = 0; using the group inverse W#, pi'(x) = -W#(x) W'(x) pi(x). Construct topology-specific responses R_k(x) and impose R(x) approximately equal to sum_k alpha_k R_k(x), with alpha_k >= 0 and sum_k alpha_k = 1. The paper's structural conjecture is that relevant responses are generated by convex combinations of topologically optimal extremal models.

Implementation notes

Use this as a differentiable latent layer in an MLP for scalar-response prediction or in a world model for systems with discrete modes. First encode the input into a scalar or low-dimensional condition x, then produce positive edge rates with W_ij(x) = softplus(a_ij^T h(x) + b_ij) + epsilon for edges belonging to one of K fixed strongly connected graph motifs. For each motif, solve the linear stationary system W_k(x) pi_k(x) = 0 with normalization, compute R_k(x) = g_k^T pi_k(x), and combine outputs using alpha = softmax(u). An optional residual head can be trained with penalty lambda times the squared difference between its output and the convex mixture. Pseudocode is: h = encoder(x); for each k compute rates, construct W_k, solve pi_k, compute R_k; y = sum_k softmax(u)_k R_k; loss = task_loss(y,target) + lambda times convex_hull_penalty. Stationary distributions are computed exactly for small motifs and differentiated by implicit linear solves. Start with a two-layer MLP on synthetic sigmoidal and nonmonotonic response curves, comparing a standard MLP and mixture-of-experts baseline. The quantitative prediction is that held-out curves should lie near the nonnegative simplex hull of motif responses, and that a single extremal directed-cycle motif should have larger maximum slope than random convex mixtures. Failure is indicated if more than 20 percent of held-out curves require negative coefficients or have normalized hull residual above 0.05.

Verification

Mechanism confirmed, baseline not beaten

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

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 9/10

Built and numerically verified a differentiable Topological Response Basis Layer using fixed strongly connected motifs, positive input-dependent rates, exact stationary linear solves, and convex softmax mixing. The mechanism manifested: implicit derivative residuals stayed below 1.2e-15 across rate-sensitivity values, centered finite-difference error scaled as h^1.999, and small-sensitivity response slopes were exactly linear. On a synthetic convex-mixture response, the recovered mixture was [0.1499997, 0.5499991, 0.3000012] and beat the small standard MLP in held-out MSE, although this is only a toy interpolation experiment.

Agent confidence: 9/10

Baseline: Small sklearn MLP, 2x24 tanh, held-out MSE 7.7555e-06
Idea: Topological convex mixture: held-out MSE 1.6647e-16; recovered alpha [0.1499997, 0.5499991, 0.3000012]; implicit derivative max residual 1.1990e-15; finite-difference log-log slope 1.9993 versus predicted 2; small-beta linear relative deviation 0.0; outside-hull constrained residual MSE 3.9290e-04

Limitations: Only scalar synthetic responses and three hand-designed motifs were tested. No large-scale dataset, learned encoder-decoder training comparison, robustness to poorly conditioned generators, FLOP/latency benchmark, or world-model experiment was performed.

How to run: python3 experiment.py && /home/maxwelhelp/main/bin/python3 topological_layer.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.12096) Sep 2, 2026

Significantly WORSE than the tuned baseline (+3723.2%, p=0.0081, wins 0 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0003
Idea mean: 0.0123
Effect (Δ): 0.012 (+3723.2%; negative = idea better)
Wins: 0 / 8 paired seeds
p-value: 0.0081 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±1000.5%
Mechanism: Confirmed ✓
Practical effect: Hurts
Baseline tuning: swept over 6 configs

Limitations:

Only the built-in dynamics track was tested. The topological head used fixed three motifs and a relatively short 16-epoch budget; no ablation of motif count, observable parameterization, rate-head capacity, residual decoder, or longer training was performed. The reported mechanism signature verified trained-model convex-weight normalization and output response behavior, but did not independently reconstruct every trained generator residual in the signature path.

How to run: python3 stage2_topological_bench.py

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

Stage 2 — Benchmark Aug 31, 2026 latest Failed ✗

Agent confidence: 10/10

Implemented the Topological Response Basis Layer as an end-to-end replacement head on the shared GRU-64 dynamics model and ran the paired-seed protocol with baseline and idea hyperparameter sweeps. The baseline achieved mean test MSE 0.0003221 versus 0.0123137 for the idea; the paired delta was +0.0119916 with permutation p=0.0081, indicating a significant loss. The trained model did exhibit normalized convex mixture weights, but this mechanism signal did not translate into benchmark performance.

Alpha sum error: 1.1920928955078125e-07
Baseline: dynamics/rnn_small, best lr=0.006, weight_decay=0, 16 epochs; 8-seed mean MSE 0.00032208267020905623, std 0.00013150113298489614
Idea: Topological stationary-response mixture head on GRU-64, best lr=0.006, weight_decay=0, 16 epochs; 8-seed mean MSE 0.012313722632825375, std 0.0036399413142409096
Paired delta: idea - baseline = +0.011991639962616318 MSE
Permutation p: 0.0081
Stationary residual: 8.940696716308594e-08

Benchmark result