△ Mechanism confirmed, baseline not beaten 2026

Braid-Monodromy Set State

Implementation & benchmark of arXiv:2607.09601 — Homological Topological Quantum Field Theories

Usefulness6/10

Difficulty5/10

Novelty7/10

Dynamics Accuracy Sample efficiency Stability Graph nn World model Rnn Attention Topology Representation theory Algebra Geometry

Source paper: Homological Topological Quantum Field Theories arXiv:2607.09601 ⓘ · analyzed Aug 30, 2026

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 Not peer-reviewed.

Idea description

Replace a standard permutation-invariant object pool with a latent state on an unordered configuration together with a fiber vector transported along the observed object trajectories. The instantaneous state remains invariant to reordering, but loops and exchanges of objects act through learned monodromy matrices, allowing the network to represent path-dependent interactions without assigning arbitrary permanent object indices.

Formulas

$$\mathrm{Conf}_{n}(X):=(X^{\times n}\setminus\Delta)/S_{n},\qquad \Delta:=\{(x_{1},\dots,x_{n})\in X^{\times n}\mid \exists i\neq j,\ x_i=x_j\}.$$

$$(\gamma\cdot v,\gamma\cdot u)=(v,u),\qquad \forall\gamma\in\pi_{1}(X,x_{0}),\ \forall u,v\in L.$$

$$\rho:\pi_{1}(\mathrm{Conf}_{n}(X),x_{0})\to GL(L),\qquad \rho(\gamma)^*G\rho(\gamma)=G.$$

$$h_{t+1}=\rho(\gamma_t)h_t+\eta_t,\qquad \rho(\gamma_t)=\exp(\Delta tA_t),\qquad A_t^*G+GA_t=0.$$

Mathematical statement

The paper defines the unordered configuration space of n distinct points by $\mathrm{Conf}_{n}(X):=(X^{\times n}\setminus\Delta)/S_{n}$, where X is the object-state manifold, $\Delta=\{(x_{1},\ldots,x_{n})\mid \exists i\neq j:x_i=x_j\}$ is the collision set, and $S_n$ permutes labels. A local system $\mathcal L$ is represented at a base configuration $x_0$ by a vector space L and a monodromy representation $\rho:\pi_1(\mathrm{Conf}_n(X),x_0)\to GL(L)$. The extracted text states the equivariant pairing condition $(\gamma\cdot v,\gamma\cdot u)=(v,u)$ for all loops $\gamma$ and vectors u,v; with pairing matrix G this becomes $\rho(\gamma)^*G\rho(\gamma)=G$. The adaptation uses a learned local generator $A_t$ and transports the fiber state according to $h_{t+1}=\rho(\gamma_t)h_t+\eta_t$, with $\rho(\gamma_t)=\exp(\Delta tA_t)$. For G equal to the identity, impose $A_t^T+A_t=0$, so transport is norm-preserving. Here $h_t\in\mathbb R^k$ is the fiber state, $\eta_t$ is a learned local innovation, and $\gamma_t$ is the observed path segment in configuration space.

Implementation notes

(1) Integrate the module after an object detector or object-token encoder and before temporal self-attention. At time t, inputs are n object embeddings $x_{t,i}\in\mathbb R^d$, positions $p_{t,i}$, velocities, and the previous fiber state $h_t\in\mathbb R^k$. Treat the object collection as unordered: all pair features and pooling operations must be symmetric sums, not dependence on an arbitrary token index. Start with $G=I_k$.

(2) Pseudocode: compute $z_t=\mathrm{SetEncoder}(\{[x_{t,i},p_{t,i}]\}_{i=1}^n)$. For every pair i<j, compute $q_{ij}=\mathrm{MLP}([x_{t,i},x_{t,j},p_{t,i}-p_{t,j},v_{t,i}-v_{t,j}])$. Predict $B_{ij}=\mathrm{reshape}(\mathrm{MLP}(q_{ij}))$ and set the skew generator $A_{ij}=B_{ij}-B_{ij}^{T}$. Compute bounded interaction weights $w_{ij,t}=\sigma(\mathrm{MLP}(q_{ij}))$ and $A_t=\sum_{i<j}w_{ij,t}A_{ij}$. Update $h_{t+1}=\exp(\Delta tA_t)h_t+\mathrm{MLP}([z_t,h_t])$. Feed $[z_t,h_{t+1}]$ to the downstream predictor. Use a differentiable matrix exponential for small k, or a second-order Cayley approximation $\rho_t=(I-\frac{\Delta t}{2}A_t)^{-1}(I+\frac{\Delta t}{2}A_t)$.

(3) The paper supplies the configuration-space quotient, path composition, and invariant-pairing constraint. The neural network estimates the local connection $A_{ij}$ from data. To expose explicit braid behavior, detect near exchanges by tracking the sign changes of pairwise relative angles and compose the corresponding local transports in temporal order. The mathematical constraint is exact for skew generators; the learned pair weights, connection, and innovation are empirical components.

(4) First cheap experiment: create 2D simulated particles with targets determined by whether two particles exchange positions clockwise or counterclockwise before a collision. Randomly permute object-token order independently at every frame. Compare DeepSets pooling, a standard object transformer, and this monodromy layer at matched parameter counts on 5-particle sequences. Test unseen initial permutations and trajectories twice as long as training trajectories. Success means higher exchange-direction accuracy, no performance degradation under token permutation, and lower long-horizon rollout error. Monitor $\|A_t^T+A_t\|_F$ and $|\|h_t\|_2-\|h_0\|_2|$; the constrained model should exhibit less latent norm drift and more stable long rollouts.

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 7/10

Built a compact Braid-Monodromy set-state module using skew-symmetric generators and Cayley transport, plus a DeepSets-GRU baseline on independently permuted 5-object trajectories. The Cayley update preserved fiber norms to 2.4e-7 error versus 0.0535 for an Euler control, and permutation changes altered logits by only 3.7e-8. In the toy exchange-direction task, the monodromy model reached 100% accuracy at both 12 and 24 frames, versus 79.1% and 68.4% for the baseline; this is a promising signal, though not yet a definitive general result.

Agent confidence: 7/10

Baseline: DeepSets-GRU, 6,354 parameters: accuracy_T12=0.7906, accuracy_T24=0.6844
Idea: Braid monodromy, 9,371 parameters: accuracy_T12=1.0000, accuracy_T24=1.0000; max skew residual=0, max Cayley norm error=2.38e-7, max permutation logit difference=3.73e-8

Limitations: Only one random seed and one synthetic trajectory family were tested. The models were not matched exactly for parameter count or FLOPs, no explicit rollout prediction metric was evaluated, and the CUDA attempt failed due to a shared-environment cuDNN allocation error so the completed run used CPU.

How to run: python3 experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.09601) Sep 4, 2026

Significantly WORSE than the tuned baseline (+208.1%, p=0.0081, wins 0 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0005
Idea mean: 0.0017
Effect (Δ): 0.0011 (+208.1%; negative = idea better)
Wins: 0 / 8 paired seeds
p-value: 0.0081 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±72.1%
Mechanism: Confirmed ✓
Practical effect: Hurts
Baseline tuning: swept over 3 configs

Limitations:

Only the built-in dynamics pendulum track was tested; the benchmark input is not an unordered multi-object configuration, so transfer of explicit braid exchange behavior was not directly evaluated. No vision, sequence, long-horizon rollout beyond the fixed dynamics target, or parameter/FLOP matching beyond the shared base-task protocol was tested.

How to run: python3 run_bench.py

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

Stage 2 — Benchmark Sep 4, 2026 latest Failed ✗

Agent confidence: 10/10

Implemented Braid-Monodromy transport with skew-symmetric generators and Cayley updates as a matched modification of the dynamics RNN system. The trained-model signature confirmed zero skew residual and 8.0e-8 mean transport-only norm drift. However, the idea had higher test MSE than the tuned baseline on all 8 seeds: 0.0016795 versus 0.0005452, paired delta +0.0011343, permutation p=0.0081; therefore the mechanism is numerically valid but gives no benchmark win.

Baseline: dynamics/rnn_small; 8-seed tuned sweep best lr=0.01; test MSE mean=0.0005451723, std=0.0002036184; per_seed=[0.0004930912,0.0004765945,0.0004673649,0.0010376222,0.0005116638,0.0002723346,0.0005571768,0.0005455301]
Idea: BraidDynamics; 8-seed idea sweep best lr=0.003; test MSE mean=0.0016794731, std=0.0005899470; per_seed=[0.0020148745,0.0009998990,0.0017773953,0.0030421584,0.0015461834,0.0013431071,0.0013974033,0.0013147637]
Paired delta: idea - baseline = +0.0011343008; idea wins=0/8; p=0.0081; verdict=idea worse (significant)

Benchmark result