✗ Failed on benchmark 2026

Holonomy-designed recurrent memory

Implementation & benchmark of arXiv:2608.30231 — "More Is Different'' in Neural Circuits: Algebraic Emergence of Effective Theories in Canonical Recurrent Motifs of Biological Neuronal Networks

Usefulness7/10

Difficulty5/10

Novelty8/10

Architecture Stability Accuracy Generalization Rnn Ssm Architecture Training dynamics Algebra Combinatorics Dynamical systems

Source paper: "More Is Different'' in Neural Circuits: Algebraic Emergence of Effective Theories in Canonical Recurrent Motifs of Biological Neuronal Networks arXiv:2608.30231 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Build a recurrent module from two finite-state factors: a normalization state and a winner-selection state. Choose or learn their coupling so that the joint transition system contains a certified composite two-cycle, giving the network a small robust memory state, while every fixed-input generator still collapses most states toward attractors. The module can be embedded in a continuous RNN using soft state assignments during training and straight-through discretization for algebraic auditing.

Formulas

$$3=(0,1,1)\mapsto 1=(0,0,1)\mapsto 4=(1,0,0)\mapsto 5=(1,0,1),$$

$$M=\langle f_u:u\in U\rangle=\{f_{u_k}\circ\cdots\circ f_{u_1}:k\geq 0,\ u_i\in U\},$$

$$C=\{(D{:}0,W{:}4),(D{:}1,W{:}5)\},\qquad g|_C:(D{:}0,W{:}4)\leftrightarrow(D{:}1,W{:}5),\qquad \langle g|_C\rangle\cong\mathbb{Z}_2,$$

$$\mathcal{L}_{\mathrm{cycle}}=\|F_{u_{1:L}}(q_0)-q_1\|_2^2+\|F_{v_{1:K}}(q_1)-q_0\|_2^2,\qquad F_{u_{1:L}}=f_{u_L}\circ\cdots\circ f_{u_1}.$$

Mathematical statement

The paper represents each input-conditioned recurrent update as a transformation of a finite state set, and the generated transition monoid is M = <f_u : u in U>, consisting of all finite compositions of generators f_u. For the WTA example, repeated application of one generator is aperiodic: 3=(0,1,1) maps to 1=(0,0,1), then to 4=(1,0,0), then to 5=(1,0,1), so no cycle longer than one is produced by that frozen drive. After coupling normalization and competition, the joint state set contains C = {(D:0,W:4),(D:1,W:5)}. The paper reports that an element of the generated monoid exchanges these two states, so its restriction to C is the nontrivial permutation group Z_2. Here D is the discrete normalization state, W is the discrete WTA state, and C is an image set of the joint transition system. The important property is composite action: both coordinates change under the same monoid element, so the cycle is not inherited from either factor separately. In an ML adaptation, let q_t=(d_t,w_t) be a quantized recurrent state, let f_u(q) be the transition under input symbol or input bin u, and require a target word u_{1:L} to satisfy f_{u_L} compose ... compose f_{u_1}(q_0)=q_1 and the reverse word to map q_1 back to q_0, while a contraction penalty sends non-target states toward fixed points.

Implementation notes

Integrate this into a small GRU-like recurrent block or an SSM state update, not into the output head. Split the hidden state into two low-dimensional factors: d_t for normalization/gain and w_t for winner identity. Compute continuous updates h'_t = RNN(h_{t-1},x_t), then map them to soft discrete states p_t = softmax(A h'_t / tau), with p_t containing logits over the finite joint states q=(d,w). Implement each input bin u as a transition matrix T_u over joint states, or parameterize T_u through differentiable DN and WTA updates; use p_t = p_{t-1} T_u during the algebraic training branch. Pseudocode: initialize p0; for each symbol u, p <- p @ T[u]; retain the resulting p for the cycle loss; decode h_t from p or use the continuous hidden state as the task representation. Select a target pair q0,q1 and two short input words a,b, then add the displayed cycle loss so T_b T_a maps q0 to q1 and the reverse word maps q1 to q0. Also enumerate all states and generators after training, construct M by breadth-first closure under composition, and reject candidates with unwanted long cycles or excessive nonabsorbing states. The paper's exact algebraic contribution is the finite transformation-monoid closure and the Z2 restriction test; transition probabilities, temperature, and task loss are engineering choices. Estimate robustness by perturbing logits and measuring whether the decoded state remains in C. First experiment: train a 2-layer vanilla GRU and this 16- or 32-state algebraic cell on a delayed-copy task with distractor symbols. Compare accuracy, hidden-state corruption under noise, and number of training examples needed. Success means higher delayed-copy accuracy and a visibly stable two-state orbit, without increasing parameter count by more than 2x.

Verification

Failed on benchmark

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Not reproduced ✗

Stage 2 · Practical benchmark result: Inconclusive

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 a 16-state deterministic two-factor transition system whose individual generators have only fixed-point cycles, while their composition contains a certified joint Z2 two-cycle exchanging states (0,0) and (1,1); monoid closure had 10 transformations. The quantitative predictions manifested: two-leg reliability exactly matched (1-eps)^4 across eps=0..0.4, the composite orbit alternated for 8 steps, and all non-attractor states reset under A within at most 2 steps. Soft exchange fidelity decreased from 1.0 at tau=0 to 0.371 at tau=0.4, verifying the mechanism, but the algebraic cell did not beat the GRU on the fixed toy task.

Agent confidence: 8/10

Baseline: Small 8-unit GRU, 330 parameters, delayed-copy accuracy 1.000 on the fixed 600-example toy set after 80 epochs; CUDA fell back to CPU after a cuDNN allocation error.
Idea: 16-state holonomy cell with exact q0<->q1 swap, generator cycle lengths [1], composite cycle lengths [1,2], and monoid size 10. Delayed-copy accuracy was 1.000 at noise 0, 0.748 at 0.05, 0.560 at 0.10, 0.303 at 0.20, and 0.171 at 0.30. The noise prediction observed=(1-eps)^4 exactly at every tested point; orbit was [5,0,5,0,5,0,5,0].

Limitations: This MVP is a hand-designed finite-state algebraic cell rather than a learned continuous GRU-like module with soft assignments and straight-through discretization. The delayed-copy comparison was not matched training: the finite-state cell was specified analytically, while the GRU was trained on one small fixed dataset. No repeated-seed statistics, FLOP-normalized comparison, variable delays, learned transition robustness, or larger tasks were tested. The corruption model was uniform state replacement, and the temperature sweep used uniform leakage as a proxy for softmax temperature.

How to run: python3 holonomy_experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.30231) Sep 2, 2026

Ahead of the baseline (−16.6%) but not statistically significant (p=0.0727, wins 7 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0005
Idea mean: 0.0005
Effect (Δ): -0.0001 (−16.6%; negative = idea better)
Wins: 7 / 8 paired seeds
p-value: 0.0727 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±17.6%
Mechanism: Not confirmed ✗
Practical effect: Inconclusive
Baseline tuning: swept over 3 configs

Limitations:

The implementation uses a cycle-inspired factor-state regularizer rather than a full learned finite transition-monoid closure. No FLOP-normalized or variable-horizon comparison was run, and the favorable point estimate did not reach permutation-test significance.

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: 9/10

Built and evaluated the holonomy-regularized recurrent system on the registered dynamics track with the same 64-unit GRU backbone as the baseline. The idea achieved lower mean MSE (0.0004547013 vs 0.0005451723) and won 7/8 paired seeds, but the permutation p-value was 0.07265, so the required significant-win criterion was not met. The trained-model perturbation signature was confirmed at observed 1.0 versus predicted 1.0, but mechanism confirmation does not make the task-metric result a significant win.

Baseline: Best lr=0.01; mean test MSE 0.0005451723 over 8 seeds.
Idea: lr=0.01, cycle_weight=0.03; mean test MSE 0.0004547013 over 8 seeds; paired delta -0.0000904710; 7/8 wins; p=0.07265.

Benchmark result