Merge Decomposition Overview

• Merge Decomposition is a collection of techniques that use intermediate representations to isolate and analyze merge operations across various domains.
• It enables canonical factorization in finite semigroup theory and structured decompositions in graph, clustering, and neural model settings.
• The approach facilitates efficient computation of reconstruction, similarity, and geodesic metrics, enhancing both theoretical analysis and practical applications.

Merge decomposition denotes a family of decomposition techniques in which a merge operation, or the effects of merging, are represented through an intermediate structure that isolates compatible components before reconstruction or analysis. In the cited literature, the term is used for an algebraic construction on finite semigroups (Gool et al., 2017), for graph construction sequences measured by merge-width (Dreier et al., 12 Jul 2026), for branch decompositions of merge trees in topological data analysis (Pont et al., 2021), for split-merge comparison of clusterings (Xiang et al., 2012), for decomposition-based neural model merging and merge-recipe recovery from sampled fingerprints (Chaichana et al., 29 May 2025, Adil et al., 12 Jul 2026), and for decompositions of Merge dynamics in generative linguistics and LLM programs (Marcolli et al., 21 Dec 2025, Saha et al., 2023).

1. Scope and recurrent structure

The cited usages are domain-specific, but they share a common operational schema: an object is first decomposed into components on which merge interactions become structurally explicit, then merging or comparison is performed at that level, and finally a global object or verdict is reconstructed. This suggests a recurrent pattern in which decomposition is not merely descriptive; it is the mechanism that makes merging well-defined, analyzable, or computationally tractable.

Domain	Decomposed object	Function of decomposition
Finite semigroup theory	Homomorphisms and semigroups generated by two subsemigroups	Factorization through a merge semigroup and a two-sided semidirect product
Structural graph theory	Construction sequences with parts and unresolved pairs	Quantify radius-$r$ merge-width
Merge trees	Branch decomposition trees	Compute distances, geodesics, barycenters
Clustering comparison	Bipartite overlap graph and meet of two partitions	Define split-merge similarity
Neural model merging and lineage	Stacked weight deltas or sampled fingerprints	Align merge coordinates or recover mixture weights
Linguistics and LLM programs	Workspaces or multi-faceted tasks	Separate merge subdynamics or branch-solve-merge subtasks

In the algebraic and graph-theoretic settings, merge decomposition is a structural theorem or parameterization. In the merge-tree, clustering, and neural settings, it is a computational representation. In the linguistic and LLM settings, it is a process decomposition: Internal Merge is separated from External Merge and Sideward Merge, or a task is branched into subproblems and then merged back into a final output (Gool et al., 2017, Dreier et al., 12 Jul 2026, Pont et al., 2021, Xiang et al., 2012, Chaichana et al., 29 May 2025, Adil et al., 12 Jul 2026, Marcolli et al., 21 Dec 2025, Saha et al., 2023).

2. Algebraic merge decomposition in finite semigroup theory

In finite semigroup theory, merge decomposition is introduced as a new algebraic technique for factoring homomorphisms from free semigroups and for decomposing finite semigroups generated by two subsemigroups (Gool et al., 2017). The setup fixes disjoint subalphabets $A_1,A_2$ with $A=A_1\cup A_2$, homomorphisms $\psi_1:A_1^+\to T_1$ and $\psi_2:A_2^+\to T_2$, and a homomorphism $\chi:(T_1\times T_2)^+\to T_0$. A word $w\in A^+$ is uniquely factored as $w=v_2uv_1$ with $v_2\in A_2^*$, $u\in(A_1^+A_2^+)^*$, and $A_1,A_2$0, giving the three-coordinate map

$A_1,A_2$1

where $A_1,A_2$2.

The central construction is the merge semigroup

$A_1,A_2$3

a triple product that can be viewed as a two-sided semidirect product. The merge morphism $A_1,A_2$4 is defined on generators so that there exists a function $A_1,A_2$5 with $A_1,A_2$6 (Gool et al., 2017). This factorization is the key algebraic content of Proposition 3.1 in the paper.

A prototypical application takes a finite semigroup $A_1,A_2$7 generated by subsemigroups $A_1,A_2$8, sets $A_1,A_2$9, and shows that $A=A_1\cup A_2$0 divides the corresponding merge triple product. This yields short proofs of the two-sided Krohn–Rhodes decomposition theorem and Henckell’s aperiodic pointlike theorem (Gool et al., 2017). The variety-theoretic control is explicit: if $A=A_1\cup A_2$1 and $A=A_1\cup A_2$2, with $A=A_1\cup A_2$3 generated by monoids and containing $A=A_1\cup A_2$4, then the merge semigroup belongs to $A=A_1\cup A_2$5. In this setting, merge decomposition is not an heuristic decomposition of a merge process; it is a canonical factorization device with direct consequences for division, product varieties, and pointlike sets.

3. Graph merge decompositions and radius-1 merge-width

In structural graph theory, merge decomposition appears through the merge-width framework of Dreier and Toruńczyk, formulated as construction sequences on a fixed vertex set (Dreier et al., 12 Jul 2026). A construction sequence maintains a partition $A=A_1\cup A_2$6 of $A=A_1\cup A_2$7 and a partition of $A=A_1\cup A_2$8 into edges $A=A_1\cup A_2$9, non-edges $\psi_1:A_1^+\to T_1$0, and unresolved pairs $\psi_1:A_1^+\to T_1$1. Each step performs exactly one of three operations: merge two parts, resolve positively all unresolved pairs between two parts, or resolve negatively all unresolved pairs between two parts. The sequence ends with one part and no unresolved pairs, thereby constructing the final graph.

The radius-$\psi_1:A_1^+\to T_1$2 width of such a sequence is defined through the resolved graph $\psi_1:A_1^+\to T_1$3, where both resolved edges and resolved non-edges count as adjacency and unresolved pairs are ignored. At every step and for every vertex $\psi_1:A_1^+\to T_1$4, one counts how many current parts are reachable from $\psi_1:A_1^+\to T_1$5 by a path of length at most $\psi_1:A_1^+\to T_1$6 in $\psi_1:A_1^+\to T_1$7. The radius-$\psi_1:A_1^+\to T_1$8 merge-width of a graph is the minimum such width over all construction sequences (Dreier et al., 12 Jul 2026).

The paper proves two structural consequences of monadic dependence. First, every monadically dependent class has almost linear neighborhood complexity: for every graph $\psi_1:A_1^+\to T_1$9 in the class and every set $\psi_2:A_2^+\to T_2$0, the family $\psi_2:A_2^+\to T_2$1 has size $\psi_2:A_2^+\to T_2$2. Second, every $\psi_2:A_2^+\to T_2$3-vertex graph in a monadically dependent class has radius-1 merge-width $\psi_2:A_2^+\to T_2$4 (Dreier et al., 12 Jul 2026). The algorithmic theorem is explicit: given an $\psi_2:A_2^+\to T_2$5-vertex graph $\psi_2:A_2^+\to T_2$6 such that

$\psi_2:A_2^+\to T_2$7

for every nonempty $\psi_2:A_2^+\to T_2$8, there is an $\psi_2:A_2^+\to T_2$9-time algorithm computing a construction sequence of radius-1 merge-width $\chi:(T_1\times T_2)^+\to T_0$0 (Dreier et al., 12 Jul 2026).

The proof strategy is algorithmic and uses “fractional twins” together with multiplicative weight updates. The paper characterizes the result as the first decomposition-based structural description of monadically dependent graph classes, settling the radius-1 case of the conjectured connection between monadic dependence and almost bounded merge-width (Dreier et al., 12 Jul 2026).

4. Branch decompositions of merge trees

For merge trees, decomposition is the computational backbone of a unified framework for distances, geodesics, and barycenters (Pont et al., 2021). Given a piecewise linear scalar field on a PL manifold, the join tree or split tree is a merge tree. In the persistence-driven branch decomposition layout, each persistent branch corresponds one-to-one to an extremum–saddle persistence pair, and the branch decomposition tree $\chi:(T_1\times T_2)^+\to T_0$1 is the dual graph whose nodes are branches and whose arcs encode adjacency between branches in the merge tree (Pont et al., 2021).

The paper restricts matching to rooted partial isomorphisms between subtrees of branch decomposition trees. This produces the branch-restricted Wasserstein distance $\chi:(T_1\times T_2)^+\to T_0$2, defined on branches with Euclidean costs in the birth/death plane. The metric is strictly equivalent to the $\chi:(T_1\times T_2)^+\to T_0$3-Wasserstein distance between extremum persistence diagrams, but it is restricted to the space of rooted partial isomorphisms between branch decomposition trees (Pont et al., 2021). In general,

$\chi:(T_1\times T_2)^+\to T_0$4

with equality when branch structure is suppressed by maximal saddle merging.

Computation proceeds by dynamic programming on subtrees and forests. Subtree deletion costs are computed bottom-up; subtree matching is allowed only at identical depth; local forest assignments are solved by the Hungarian or Auction algorithm. The implementation uses a task-based engine with sparse matrices for subtree and forest distances and shared-memory parallelism (Pont et al., 2021). This decomposition makes linear interpolation geodesics possible: matched branches interpolate linearly in the birth/death plane, deleted branches move to the diagonal, created branches move from the diagonal, and a local normalization/inversion step preserves nested intervals and hence merge-tree validity. The same framework supports barycenters through an assignment/update loop that minimizes the Fréchet functional

$\chi:(T_1\times T_2)^+\to T_0$5

The reported empirical result is computationally specific: barycenter computations are obtained in the orders of minutes for the largest examples, and the resulting barycenter merge trees visually summarize the features of interest found in the ensemble (Pont et al., 2021). In this literature, merge decomposition is therefore a branch-level reduction of the merge tree that simultaneously constrains the metric space and simplifies geodesic and barycentric computations.

5. Split-merge decomposition for comparing clusterings

In clustering evaluation, merge decomposition appears in the split-merge framework for comparing two hard partitions of the same dataset (Xiang et al., 2012). The relation between a true clustering $\chi:(T_1\times T_2)^+\to T_0$6 and a predicted clustering $\chi:(T_1\times T_2)^+\to T_0$7 is modeled as a bipartite graph with nodes $\chi:(T_1\times T_2)^+\to T_0$8 and edges $\chi:(T_1\times T_2)^+\to T_0$9 whenever $w\in A^+$0. The weakly connected components of this graph partition the join of $w\in A^+$1 and $w\in A^+$2 into localized regions of interaction. Many measures can be written as a component-based decomposition, and the paper advocates the join-weighted form

$w\in A^+$3

Each component is then refined by split and merge subcomponents. For a cluster $w\in A^+$4, the split graph $w\in A^+$5 encodes how $w\in A^+$6 is split across the induced clustering $w\in A^+$7. For a cluster $w\in A^+$8, the merge graph $w\in A^+$9 encodes how the induced clustering $w=v_2uv_1$0 must merge to form $w=v_2uv_1$1. The meet

$w=v_2uv_1$2

indexes subcomponent pairs, and the global split-merge measure is

$w=v_2uv_1$3

where $w=v_2uv_1$4 and $w=v_2uv_1$5 are subcomponent scores (Xiang et al., 2012).

The framework is designed to satisfy conditional normalization. Given the true clustering $w=v_2uv_1$6, one has $w=v_2uv_1$7 iff $w=v_2uv_1$8, $w=v_2uv_1$9 iff $v_2\in A_2^*$0 is a worst clustering in $v_2\in A_2^*$1, and $v_2\in A_2^*$2 otherwise (Xiang et al., 2012). An entropy-based instance defines

$v_2\in A_2^*$3

leading to the entropy-based split-merge measure

$v_2\in A_2^*$4

The paper emphasizes that the framework can make use of data point information, such as feature vectors and pairwise distances, by substituting alternative subcomponent scores (Xiang et al., 2012). On a coreference resolution dataset, only the split-merge entropy measure decreases strictly from $v_2\in A_2^*$5 to $v_2\in A_2^*$6 along a constructed path from the true clustering to a worst clustering. Here, merge decomposition is the meet-weighted localization of merge effects inside each overlap region of two partitions.

6. Neural model merging and fingerprint-based merge recovery

In neural model merging, merge decomposition is formulated as the principle that model merging should occur in a coordinated feature space, not the raw parameter space (Chaichana et al., 29 May 2025). “Decom-Renorm-Merge” constructs that space by applying a joint SVD to per-layer weight deltas

$v_2\in A_2^*$7

partitioning $v_2\in A_2^*$8, and reconstructing each task update as $v_2\in A_2^*$9 (Chaichana et al., 29 May 2025). The paper argues that direct entry-wise merging fails because of neuron permutation or rotation, feature drift and contextualization, and polysemantic neurons. DRM-H uses horizontal stacking and a shared column basis $u\in(A_1^+A_2^+)^*$0; DRM-V uses vertical stacking and a shared row basis.

The distinguishing operation is renormalization. After partitioning $u\in(A_1^+A_2^+)^*$1, the rows of each $u\in(A_1^+A_2^+)^*$2 are no longer unit norm, so DRM rescales rows to unit length and transfers the norms into task-specific singular scales $u\in(A_1^+A_2^+)^*$3. The paper identifies renormalization as the crucial component for creating a robust and even joint space for merging (Chaichana et al., 29 May 2025). The ablation is explicit: removing renormalization degrades performance by $u\in(A_1^+A_2^+)^*$4 on ViT-B/32, $u\in(A_1^+A_2^+)^*$5 on T5-Base, $u\in(A_1^+A_2^+)^*$6 on DeBERTa-Base, and $u\in(A_1^+A_2^+)^*$7 on Llama3.1-8B LoRA. After renormalization, DRM applies pruning, sign election, and disjoint averaging in the aligned space, and no additional finetuning is required (Chaichana et al., 29 May 2025).

Empirically, DRM outperforms several state-of-the-art merging techniques across encoder, encoder-decoder, and decoder-only settings (Chaichana et al., 29 May 2025). Without validation tuning, DRM-H improves over the strongest baseline by $u\in(A_1^+A_2^+)^*$8 on ViT-B/32 and $u\in(A_1^+A_2^+)^*$9 on ViT-L/14; on DeBERTa-Base, DRM-H improves by $A_1,A_2$00 without tuning and $A_1,A_2$01 with tuning; on Llama3.1-8B with LoRA, DRM-H gives $A_1,A_2$02 without tuning, while DRM-V gives $A_1,A_2$03 with tuning (Chaichana et al., 29 May 2025). The authors summarize this as merging on the “right space”.

A second neural usage appears in model lineage analysis. modelDNA exploits the fact that mainstream weight-merging methods in mergekit are (near-)linear per tensor and that fingerprint sample positions are deterministic functions of tensor identity (Adil et al., 12 Jul 2026). As a result, a merged model’s fingerprint is the same linear combination of its parents’ fingerprints at the sampled positions, enabling mixture recovery from fingerprints alone by the constrained least-squares problem

$A_1,A_2$04

The KKT solution is given in closed form, with optional Tikhonov regularization and optional non-negativity constraints (Adil et al., 12 Jul 2026).

This decomposition is evaluated against merges with published mergekit configurations as ground truth. The reported results are exact: the method recovers a slerp merge’s layer-interpolation curves at $A_1,A_2$05 and a dare_ties merge’s mixture weights to within $A_1,A_2$06 of the published values, without downloading any weights beyond the fingerprints (Adil et al., 12 Jul 2026). On a benchmark of $A_1,A_2$07 real Hub models with org-documented parentage, judged against $A_1,A_2$08 candidate bases, the system achieves AUROC $A_1,A_2$09, zero false positives at its reporting threshold, and $A_1,A_2$10 correct top-1 parent attribution (Adil et al., 12 Jul 2026). In this setting, merge decomposition is not the construction of a merged model; it is the inverse problem of reading the merge recipe from sampled, element-aligned weight fingerprints.

7. Linguistic Merge dynamics and branch-solve-merge programs

In a Hopf algebra Markov chain model of syntactic structure formation, merge decomposition separates the dynamics of Internal Merge from those of External Merge and Sideward Merge (Marcolli et al., 21 Dec 2025). The state space consists of binary rooted forests with labelled leaves, interpreted as workspaces. A partition map $A_1,A_2$11 records the leaf counts of the components of a forest, producing a decomposition of the state space into fibers $A_1,A_2$12 over partitions of $A_1,A_2$13 (Marcolli et al., 21 Dec 2025). Internal Merge preserves the partition, while External Merge and minimal Sideward Merge change it.

The paper proves that, for any partition with at least one block of size at least $A_1,A_2$14, each connected component of the Internal Merge graph is strongly connected and aperiodic, with equal in- and out-degrees

$A_1,A_2$15

and hence the restricted Hopf algebra Markov chain has uniform stationary distribution on each component (Marcolli et al., 21 Dec 2025). By contrast, the unweighted dynamics of Internal Merge, External Merge, and minimal Sideward Merge is ergodic on the whole workspace graph, but Sideward Merge prevents convergence to connected trees. For $A_1,A_2$16, the stationary mass on partitions is highest on $A_1,A_2$17 and lowest on $A_1,A_2$18, with values approximately $A_1,A_2$19, $A_1,A_2$20, $A_1,A_2$21, and $A_1,A_2$22 for $A_1,A_2$23, $A_1,A_2$24, $A_1,A_2$25, and $A_1,A_2$26, respectively (Marcolli et al., 21 Dec 2025). Cost functions based on Minimal Search, Minimal Yield, and Complexity Loss do not suffice to force convergence to connected trees, whereas augmenting the cost with the Shannon entropy of the source partition yields a unique critical circuit at $A_1,A_2$27 and makes the stationary distribution concentrate, at leading order, on connected trees (Marcolli et al., 21 Dec 2025).

A programmatic analogue appears in "Branch-Solve-Merge" for LLMs (Saha et al., 2023). There, merge decomposition is operationalized by a controller that decomposes a task into parallel sub-tasks, solves them independently, and fuses the partial outputs. For response evaluation, the branch module generates up to five criteria conditioned only on the question, the solve module assigns criterion-specific scores to two responses, and the merge module aggregates by a deterministic sum-of-scores together with swap-based consistency to reduce order effects (Saha et al., 2023). For constrained story generation, the branch module splits a concept set into two groups and proposes a topic, the solve module drafts two sub-stories, and the merge module synthesizes a final story preserving all concepts.

The empirical claims are explicit. Across evaluation tasks, BSM improves human-LLM agreement by up to $A_1,A_2$28 absolute and reduces position and length biases by up to $A_1,A_2$29 absolute, and in constrained generation it improves constraint satisfaction by $A_1,A_2$30 absolute (Saha et al., 2023). On the writing domain for LLaMA-2-70B-chat, agreement rises from $A_1,A_2$31 to $A_1,A_2$32, position bias falls from $A_1,A_2$33 to $A_1,A_2$34, and length bias falls from $A_1,A_2$35 to $A_1,A_2$36 (Saha et al., 2023). This is a process-level notion of merge decomposition: the merge is deferred until task-specific evidence has been generated on decomposed branches.

Across these literatures, merge decomposition is not a single formal object but a recurring research strategy: merge is made reliable by changing representation first. In semigroup theory this yields factorization through a merge semigroup; in graph theory it yields low-width construction sequences; in merge trees and clustering it yields component-wise metrics and similarities; in neural systems it yields aligned merge spaces or recoverable merge recipes; and in linguistic or LLM settings it yields explicit separation between subdynamics or subtasks before synthesis (Gool et al., 2017, Dreier et al., 12 Jul 2026, Pont et al., 2021, Xiang et al., 2012, Chaichana et al., 29 May 2025, Adil et al., 12 Jul 2026, Marcolli et al., 21 Dec 2025, Saha et al., 2023).