Home / Curated / Algorithms / KronA

KronA

Algorithms AIMech:KronA DRAFT

Aliases: Kronecker Adapter, KronA Parameter-Efficient Kronecker Tuning, Kronecker Product Adapter (KronA)

A parameter-efficient fine-tuning adapter that replaces low-rank bottleneck projections with a Kronecker product of two smaller factor matrices A in R^{a_1 x a_2} and B in R^{b_1 x b_2}, computing the transformed activation via reshaped matrix-matrix multiplications without explicitly materializing the full Kronecker product matrix.

Identity Notes & Concept Scope Distinct from standard LoRA (AIMech:LoRA, which parameterizes weight updates as a low-rank outer matrix product B A bounded by rank r), from Compacter (AIMech:Compacter, which uses parameterized hypercomplex multiplication sums of n Kronecker products with shared slow weights and rank-1 fast weights inside a nonlinear bottleneck adapter), and from LoKr (AIMech:LoKr, which further decomposes the second Kronecker factor into a low-rank product u_2 v_2^T inside the LyCORIS framework rather than KronA's two dense Kronecker factors A kron B deployed as sequential or parallel residual adapters with optional input-Output scaling). Formulated by Edalati et al. (arXiv:2212.10650).

Claims & Grounded Evidence

4 Claims

Mechanism #krona-kronecker-rank-multiplication-and-fast-matvec

KronA parameterizes a linear transformation of dimension (a_1 b_1) x (a_2 b_2) as the Kronecker product W_K = A kron B of two factor matrices A in R^{a_1 x a_2} and B in R^{b_1 x b_2} using only a_1 a_2 + b_1 b_2 trainable parameters while achieving multiplicative rank(A kron B) = rank(A) rank(B), and evaluates (A kron B) x for x in R^{a_2 b_2} via the reshape identity vec(B X A^T) where X = reshape(x, (b_2, a_2)) without materializing the (a_1 b_1) x (a_2 b_2) Kronecker matrix.

supports Source: KronA: Parameter Efficient Tuning with Kronecker Adapter · Locator: Section 3.1 Kronecker Product & Eq. 5, p. 3

Mechanism #krona-sequential-and-parallel-residual-variants

KronA defines both a sequential adapter module instilled after feed-forward layers and a parallel residual module KronA_B (and rescaled KronA_B^resB) placed alongside pretrained linear layers that updates the output as y = W_0 x + s (A kron B) x, allowing the Kronecker product s (A kron B) to be merged directly into W_0 at inference time with zero additional latency.

supports Source: KronA: Parameter Efficient Tuning with Kronecker Adapter · Locator: Section 3.2 Kronecker Adapter (KronA), Eq. 6-7 and Figure 1, p. 3-4

Application #krona-glue-language-model-adaptation

KronA is applied to parameter-efficient fine-tuning of pretrained T5-base encoders and decoders on the GLUE natural language understanding benchmark, matching or exceeding the accuracy of LoRA, Adapter, and Compacter while using 0.07% of total model parameters and achieving faster inference throughput than PHM-based Compacter modules.

supports Source: KronA: Parameter Efficient Tuning with Kronecker Adapter · Locator: Section 4 Experiments, Tables 2-3, p. 4-6

Limitation #krona-factor-dimension-divisibility-constraint

KronA requires the input and output dimensions d_in and d_out of the adapted module to factorize into integer pairs d_out = a_1 b_1 and d_in = a_2 b_2, constraining the available parameter budgets to discrete divisor pairs of the layer dimensions.

supports Source: KronA: Parameter Efficient Tuning with Kronecker Adapter · Locator: Section 3.2 & Section 4.2 Hyperparameter Sensitivity (Factor Sizes), Table 5, p. 3-4, 6

⚠️ Curated Limitations & Negative Results

Requires the adapted layer dimensions d_out and d_in to factorize into integer dimensions (a_1, b_1) and (a_2, b_2), restricting the granularity of trainable parameter budgets.
When deployed as an unmerged sequential adapter with a fixed skip-connection scale rather than a parallel residual adapter, inference incurs a small additional matrix-reshaping and multiplication overhead.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:ArtificialNeuralNetwork` Artificial Neural Network	related_match	KronA adapts pretrained neural network modules via Kronecker-product adapter transformations.

Evidence Sources & Primary Literature

3 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`r679-primary-pdf`	KronA: Parameter Efficient Tuning with Kronecker Adapter Ali Edalati, Marzieh Tahaei, Ivan Kobyzev, Vahid Partovi Nia, James J. Clark, Mehdi Rezagholizadeh	primary paper full_text	ID: https://dblp.org/rec/journals/corr/abs-2212-10650.xml sha256:32ed561896c0…
`r679-arxiv-abs`	KronA: Parameter Efficient Tuning with Kronecker Adapter (arXiv abstract) Ali Edalati, Marzieh Tahaei, Ivan Kobyzev, Vahid Partovi Nia, James J. Clark, Mehdi Rezagholizadeh	primary paper abstract_only	ID: https://arxiv.org/abs/2212.10650 sha256:374a9161b35e…
`r679-metadata-cr`	KronA: Parameter Efficient Tuning with Kronecker Adapter (Crossref metadata) Ali Edalati, Marzieh Tahaei, Ivan Kobyzev, Vahid Partovi Nia, James J. Clark, Mehdi Rezagholizadeh	primary paper abstract_only	ID: https://dblp.org/rec/journals/corr/abs-2212-10650.xml sha256:e8b3a758e3de…

Ontology & Grounding Topology

Open in interactive network graph →

Curation Provenance

Curator: Antigravity

Method: ai_assisted

Model: Gemini

Created: 2026-10-19T12:42:00Z

Updated: 2026-10-19T12:42:00Z

Source File: data/curated/KronA.yaml · Schema: curated_concept.yaml

Explore across X-Mech

Related computational & biological Mechs: Embedding Map Network Graph ProteinTraitsMech (ESM-2 pLM) DUFMech (Domain pLM) TraitMech CultureMech All 14 X-Mech Repositories