✗ Failed on benchmark 2026

Backward-Bifurcation Competitive Memory

Implementation & benchmark of arXiv:2608.05914 — Backward bifurcations in spatial replicator models:when invasion criteria fail to predict coexistence

Usefulness7/10

Difficulty6/10

Novelty8/10

Dynamics Stability Accuracy Generalization Rnn Ssm Training dynamics Mlp Dynamical systems Bifurcations Spectral theory Pde Control theory

Source paper: Backward bifurcations in spatial replicator models:when invasion criteria fail to predict coexistence arXiv:2608.05914 ⓘ · analyzed Aug 31, 2026

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 Not peer-reviewed.

Idea description

Replace an ordinary contracting recurrent state with two spatially coupled competing latent populations whose nonlinear interaction admits a stable finite-amplitude coexistence state even when the infinitesimal invasion eigenvalue is negative. This creates hysteretic, robust memory: a representation survives small perturbations and weak evidence, but can be switched by a sufficiently large input pulse.

Formulas

$$\mathcal{L}_{\mu}\phi=D\Delta\phi-v(x)\cdot\nabla\phi+a(x;\mu)\phi,\qquad \lambda(\mu)=s(\mathcal{L}_{\mu})$$

$$\dot A=A\left[\lambda(\mu)+cA-kA^{2}\right],\qquad A_{\pm}=\frac{c\pm\sqrt{c^{2}+4k\lambda}}{2k}$$

$$\lambda_{\mathrm{fold}}=-\frac{c^{2}}{4k},\qquad \lambda_{\mathrm{fold}}<\lambda<0\ \Longrightarrow\ A_{+}>0\text{ stable},\ A_{-}>0\text{ unstable}$$

$$p_{t+1}=\operatorname{clip}_{[0,1]}\left[p_t+\Delta t\left(D\Delta_h p_t-v\nabla_h p_t+p_t\left(\lambda_t+c p_t-kp_t^{2}\right)+u_t\right)\right]$$

Mathematical statement

Let p(x,t) ∈ [0,1] be the activation frequency of latent population 1 on a spatial domain Ω, with diffusion coefficient D > 0, advection velocity v(x), and control parameter μ. Its linearized invasion operator around p = 0 is L_μ φ = DΔφ - v·∇φ + a(x;μ)φ, and the spatial invasion rate is λ(μ) = s(L_μ), the principal eigenvalue or spectral bound. Near the critical point λ = 0, projection onto the principal eigenfunction gives the amplitude equation dA/dt = A[λ(μ) + cA - kA²], where A ≥ 0 is the amplitude, c > 0 is the quadratic cooperative coefficient, and k > 0 is the saturating cubic coefficient. A stable positive equilibrium exists for λ < 0 whenever c² + 4kλ > 0: A_± = [c ± √(c² + 4kλ)]/(2k). The saddle-node fold is λ_fold = -c²/(4k); A_+ is stable and A_- is unstable. Thus λ < 0 does not imply extinction or zero latent activity. The neural implementation uses this same normal form at every feature location, with λ estimated from the Jacobian of the discretized spatial operator.

Implementation notes

1. Exact integration point: add this as a recurrent latent block before the readout of a GRU, S4-style state-space model, or sequence MLP. Discretize the feature or token axis into N sites; p[i] is the active population at site i, λ_t[i] is produced from the current input by a learned linear projection, and u_t[i] is an external input pulse. Use a periodic or reflecting finite-difference Laplacian and optional learned advection. 2. Pseudocode: initialize p; for each input x_t compute λ=Linear(x_t), u=InputProjection(x_t), lap=laplacian(p), adv=advection(p); set dp=D*lap-adv+p*(λ+c*p-k*p*p)+u; update p=clip(p+dt*dp,0,1); pool p and concatenate it with x_t for the task head. Constrain c,k,D to be positive using softplus. 3. Compute the principal eigenvalue λ_hat of the linearized N×N operator DΔ_h-v∇_h+diag(λ) by 10-20 power iterations or Arnoldi; estimate c and k by fitting measured equilibria after projecting onto the leading eigenvector. The paper supplies the eigenvalue-controlled invasion mechanism and fold normal form; the neural coefficients are empirical estimates. 4. First cheap experiment: train this block and a GRU with matched parameter count on a synthetic sequence dataset containing two persistent spatially localized binary patterns and a delayed-switch task. Before training, sweep constant λ and verify a saddle-node: fit equilibria versus λ and locate λ_fold≈-c²/(4k), with measured fold within 20%. In λ_fold<λ<0, the block should retain a nonzero state after input removal, whereas a linearized control decays. Test hysteresis by increasing and decreasing the input pulse; the predicted signature is two stable response branches separated by an unstable threshold and switching thresholds differing by at least 10%.

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 8/10

Built and ran a reproducible backward-bifurcation normal-form and finite-difference spatial verification in verify_bifurcation.py. The predicted fold lambda_fold=-c²/(4k) was recovered across four (c,k) settings with maximum grid error 0.00019; for lambda<0 but above the fold, the upper finite-amplitude equilibrium persisted while infinitesimal states decayed. Hysteresis was strongly observed, with switching thresholds 0.00250 upward versus 0.31424 downward (125.7x ratio), so the proposed mechanism manifested in the toy system.

Agent confidence: 8/10

Baseline: Contracting linear state: final amplitude 0.001983 after 60 time units at lambda=-0.1.
Idea: Backward-bifurcation state: final amplitude 0.887298, matching the analytic A+=0.887298; measured equilibrium errors were below 3.1e-3 across the tested scaling sweep. Principal spatial eigenvalue was -0.1 to numerical precision.

Limitations: This is a mechanism-level toy verification, not a trained GRU/sequence-task comparison. The spatial test used periodic diffusion and a constant lambda; the localized-seed state was not used to claim uniform convergence. Hysteresis was measured in the scalar normal form with pulses, and finite-time integration plus clipping were not stress-tested across neural training settings.

How to run: python3 verify_bifurcation.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.05914) Sep 2, 2026

Significantly WORSE than the tuned baseline (+202.1%, p=0.0081, wins 0 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.001
Idea mean: 0.0032
Effect (Δ): 0.0021 (+202.1%; negative = idea better)
Wins: 0 / 8 paired seeds
p-value: 0.0081 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±126.1%
Mechanism: Not confirmed ✗
Practical effect: Hurts
Baseline tuning: swept over 3 configs

Limitations:

Only the registered built-in actuated-pendulum dynamics track was tested; custom persistent-memory tasks, longer horizons, alternative widths, advection, and broader tuning were not evaluated.

How to run: python3 bench_backward_bifurcation.py

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

Stage 2 — Benchmark Sep 1, 2026 latest Failed ✗

Agent confidence: 10/10

The backward-bifurcation recurrent state lost significantly to the matched standard GRU on the registered dynamics track. Baseline mean test MSE was 0.0010460769990459085 versus 0.0031606621923856437 for the idea, with paired delta_mean 0.0021145851933397353 and permutation p_value 0.0081. The trained-model mechanism signature was not confirmed because the pulse decreased the measured idea response.

Baseline: mean test MSE=0.0010460769990459085; best lr=0.006
Delta mean: 0.0021145851933397353
Idea: mean test MSE=0.0031606621923856437; best lr=0.006
Idea wins: 0
N pairs: 8
P value: 0.0081

Benchmark result