Unverified 2026

Annihilating Density-Transport Feature Mixer

Usefulness6/10

Difficulty5/10

Novelty8/10

Dynamics Stability Generalization Accuracy Cnn Transformer Rnn Regularization Statistical mechanics Dynamical systems Pde Stochastic processes

Source paper: Effective Hardcore Exclusion Without Exclusion: Two-Species Pair Annihilation Revisited arXiv:2608.30630 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Replace an ordinary local convolution or token-mixing block by two nonnegative feature populations A and B that diffuse and drift along the spatial or token axis, with transport rates increasing quadratically with local population and with an optional directional bias. Add a local A plus B to empty reaction so mutually conflicting feature mass is removed rather than merely averaged. The module should produce adaptive competition, and its isolated relaxation should exhibit a measurable t raised to minus 1/3 density-decay regime when quadratic asymmetric transport is active, versus t raised to minus 1/4 when transport is symmetric or effectively three-body.

Formulas

R_A±(x) = D_A × (1 ± d_A / 2) × x + Q_A × (1 ± q_A / 2) × x × (x − 1); R_B±(x) = D_B × (1 ± d_B / 2) × x + Q_B × (1 ± q_B / 2) × x × (x − 1).

a_i(t + 1) = a_i(t) + Δt × [R_A+(a_(i−1)(t)) + R_A−(a_(i+1)(t)) − R_A+(a_i(t)) − R_A−(a_i(t)) − sigma × a_i(t) × b_i(t)].

b_i(t + 1) = b_i(t) + Δt × [R_B+(b_(i−1)(t)) + R_B−(b_(i+1)(t)) − R_B+(b_i(t)) − R_B−(b_i(t)) − sigma × a_i(t) × b_i(t)].

rho(t) = (1 / L) × sum over i of [a_i(t) + b_i(t)] approximately proportional to t^(−1/3) for pair-induced asymmetric transport, and approximately proportional to t^(−1/4) for symmetric or effectively three-body transport.

Mathematical statement

For lattice location i, let a_i and b_i be nonnegative feature populations. The hopping rates are R_A plus or minus of x equals D_A times (1 plus or minus d_A divided by 2) times x plus Q_A times (1 plus or minus q_A divided by 2) times x times (x minus 1), with an analogous expression for species B. Here D is the linear hopping rate, Q is the pair-induced hopping rate, and d and q are directional biases. A reaction removes one A and one B at rate sigma times a_i times b_i. In a differentiable neural implementation, replace x times (x minus 1) by x squared and use an Euler update containing incoming transport, outgoing transport, and the annihilation term. The predicted one-dimensional density is rho of t proportional to t raised to minus 1/3 when at least one species has nonzero pair-induced asymmetric transport, whereas symmetric or three-body-dominated transport gives rho of t proportional to t raised to minus 1/4. The neural mechanism is therefore a controllable change in the relaxation exponent through Q, q, and sigma.

Implementation notes

First, integrate the mechanism as a residual mixer before an MLP in a one-dimensional CNN, token-mixing transformer, or state-space model. Project the input tensor h into two nonnegative populations using a = softplus(W_A h) and b = softplus(W_B h), with sequence position or spatial position as the lattice coordinate. Apply one or more reaction-diffusion steps along that coordinate, then project the concatenated populations or their difference a minus b back to the model width and add the result residually to h. Pseudocode is: compute rpA = D_A × (1 + d_A / 2) × a + Q_A × (1 + q_A / 2) × a squared; compute rmA = D_A × (1 − d_A / 2) × a + Q_A × (1 − q_A / 2) × a squared; compute divA = shift_left(rpA) + shift_right(rmA) − rpA − rmA; update a = softplus(a + dt × (divA − sigma × a × b)). Repeat the same operations for b and use circular or reflective boundaries. The paper supplies the transport and reaction mechanism and the expected decay exponents; the implementation must estimate the exponent from a log-log regression of rho against time and must empirically tune dt for nonnegative bounded states. The first cheap experiment should use a length-512 synthetic one-dimensional denoising task, comparing a standard depthwise convolution, a symmetric mixer with Q equal to zero, and an asymmetric pair-transport mixer with Q positive and q equal to 0.5. Before training, initialize spatially separated A and B populations and measure rho over 1,000 updates. The predicted signature is a fitted slope near minus 1/3 for quadratic asymmetric transport and near minus 1/4 after removing directional asymmetry or using a cubic transport surrogate. The exponent difference should be approximately 0.08 and should persist across lattice sizes. Then measure validation error and surviving feature mass to determine whether annihilation improves denoising until excessive sigma causes underfitting.

Verification

This idea has not been verified yet.

Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.

Artifacts

Artifacts unavailable.