Unverified 2026

Invariant-Lattice Adapter

Implementation & benchmark of arXiv:2608.21238 — $G_2$-Manifolds from 4d $\mathcal{N}=1$ Quivers

Usefulness5/10

Difficulty4/10

Novelty8/10

Regularization Memory Generalization Accuracy Fine tuning Transformer Memory Regularization Algebra Geometry Representation theory Linear algebra

Source paper: $G_2$-Manifolds from 4d $\mathcal{N}=1$ Quivers arXiv:2608.21238 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Restrict a fine-tuning adapter or output head to the subspace invariant under a prescribed monodromy, analogous to the paper's unbroken flavor lattice. This removes update directions intentionally changed by the domain-loop transformation, producing a parameter-efficient adapter with an explicit algebraic constraint.

Formulas

$$\mathcal{M}[4\alpha_1]:\quad \alpha_1\mapsto\alpha_1-4E,\qquad \alpha_3\mapsto\alpha_3+2E,\qquad \alpha_i\mapsto\alpha_i\quad(i=2,4,5,6,7,8)$$

$$\Lambda_{n=4}=\langle\alpha_1+\alpha_2+2\alpha_3+2\alpha_4+\alpha_5,\alpha_2,\alpha_4,\alpha_5,\alpha_6,\alpha_7,\alpha_8\rangle_{\mathbb Z}=\alpha_1^\perp$$

$$P=V(V^TGV)^{-1}V^TG,\qquad P^2=P$$

$$\Delta W=AP\qquad\text{or}\qquad\Delta W\leftarrow\Delta W P$$

Mathematical statement

For the n=4 construction, the paper states that the monodromy acts on simple roots by $\alpha_1\mapsto\alpha_1-4E$, $\alpha_3\mapsto\alpha_3+2E$, and $\alpha_i\mapsto\alpha_i$ for $i=2,4,5,6,7,8$. The invariant root sublattice is $\Lambda_{n=4}=\alpha_1^\perp\subset\Lambda_{E_8}$, explicitly generated by $v_1=\alpha_1+\alpha_2+2\alpha_3+2\alpha_4+\alpha_5$ and $v_2=\alpha_2,v_3=\alpha_4,v_4=\alpha_5,v_5=\alpha_6,v_6=\alpha_7,v_7=\alpha_8$; the paper identifies this lattice with $-E_7$. Let $V=[v_1,\ldots,v_7]$ be the d-by-7 matrix of these generators in feature coordinates and let G be the feature-space inner-product matrix. The projector onto the invariant subspace is $P=V(V^TGV)^{-1}V^TG$, satisfying $P^2=P$. A LoRA update can be constrained as $\Delta W=AP$, or an unconstrained update can be projected after every optimizer step.

Implementation notes

Apply this to LoRA fine-tuning in a Transformer attention projection, preferably the query, key, or value matrix, where parameter and update reductions are easy to inspect. Select d feature coordinates and construct seven basis vectors from the paper's invariant generators. For an MVP, embed the symbolic roots α_i as standard coordinate vectors in R^8, giving v1=(1,1,2,2,1,0,0,0), v2=(0,1,0,0,0,0,0,0), v3=(0,0,0,1,0,0,0,0), v4=(0,0,0,0,1,0,0,0), v5=(0,0,0,0,0,1,0,0), v6=(0,0,0,0,0,0,1,0), and v7=(0,0,0,0,0,0,0,1). Use G=I initially, orthonormalize V numerically, and compute P=V(V^TV)^(-1)V^T once. Replace ordinary LoRA ΔW=BA with ΔW=BA P, or parameterize B'=BP so every update is invariant. During training, optionally add λ||ΔW−TΔW T^(-1)||_F^2 when an explicit monodromy T is available, but the hard projection is the primary MVP. The paper supplies the invariant generators and subspace dimension; empirical choices are adapter rank and insertion point. Compare full LoRA, ordinary LoRA at equal parameter count, and projected LoRA on a small ViT or language model over several classification or instruction-tuning tasks. Record trainable parameters, peak memory, validation accuracy, domain-shift accuracy, and cross-task forgetting. Success means matching ordinary LoRA with fewer parameters or reducing forgetting at equal parameter count.

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.