Task Arithmetic: Editing & Merging Models

Updated 13 March 2026

Task arithmetic is a formalism that represents model adaptations as task vectors—the parameter differences between a fine-tuned model and a base model.
It employs linear operations like addition, subtraction, and analogy to merge, transfer, and edit neural network behaviors without full retraining.
Empirical results demonstrate improvements in speech translation, continual learning, and unlearning, with theoretical guarantees supporting reduced interference.

Task arithmetic is a formalism for manipulating neural network parameters by defining and composing “task vectors” in parameter space. It enables transfer, merging, and editing of model behaviors through linear operations, with direct applications in speech translation, continual learning, explainability transfer, unlearning, and federated averaging. This article presents an integrated technical overview of task arithmetic, focusing on its foundational principles, mathematical formulations, and major application domains developed in the research literature.

1. Core Concepts: Task Vectors and Elementary Operations

Task arithmetic centers on the notion of a task vector—the parameter difference between a model fine-tuned on a specific task and a common base model. Given a (possibly multimodal) pre-trained model with weights $\theta_0 \in \mathbb{R}^d$ , and a fine-tuned variant for task $t$ with weights $\theta_t$ , the task vector is defined as

$\tau_t = \theta_t - \theta_0.$

This vector encodes the full trajectory from pre-trained initialization to task specialization in parameter space.

Elementary Vector Operations:

Task arithmetic uses linear combinations of task vectors to steer models:

Merging (Addition): Combine capabilities by summing weighted task vectors:

$\theta_{\text{merge}} = \theta_0 + \sum_{i=1}^N \lambda_i \tau_{t_i},$

where $\lambda_i$ are interpolation coefficients controlling the contribution of each task (Ilharco et al., 2022, Cheng et al., 2024).

Negation/Subtractive Editing: Remove or suppress a task by subtracting its vector:

$\theta' = \theta_0 - \alpha \tau_t,$

with $\alpha$ a scalar (Ilharco et al., 2022, Cai et al., 29 Jan 2026).

Task Analogies: Synthesize unseen tasks by analogy:

$\tau_D \approx \tau_C + (\tau_B - \tau_A), \quad \text{for} \quad A:B::C:D,$

enabling zero-shot construction of parameters for task $D$ (Ilharco et al., 2022, Cheng et al., 2024).

These operations are always with respect to a shared parameter origin (typically the pre-trained model), and assume task vectors are computed in a common parameterization/alignment (Ilharco et al., 2022, Yoshikawa et al., 6 Jul 2025).

2. Mathematical Foundations and Theoretical Guarantees

Task arithmetic is rigorously connected to the geometry of SGD-based fine-tuning and multitask learning:

Under one epoch of gradient descent for each task with a fixed step size, the task vector $\tau_t^{(1)}$ for task $t$ exactly captures the negative gradient:

$\tau_t^{(1)} = -\eta \nabla \bar{L}_t (\theta_0),$

where $\bar{L}_t$ is the task loss and $\eta$ the learning rate (Zhou et al., 22 Aug 2025).

Summing these vectors and adding to $\theta_0$ is mathematically equivalent to a joint one-step update on the sum of the task losses (i.e., multitask gradient descent):

$\theta_{\text{TA}}^{(1)} = \theta_0 + \sum_{t \in T} \tau_t^{(1)} = \theta_0 - \eta \sum_{t\in T} \nabla \bar{L}_t(\theta_0) = \theta^{(1)}_{\text{MT}},$

formalizing task arithmetic as distributed approximate multitask learning (Zhou et al., 22 Aug 2025).

For multi-epoch fine-tuning, the equivalence holds up to a second-order curvature term $O(\eta^2)$ , which is bounded for feed-forward networks and is empirically negligible for typical tunings (demonstrated by high cosine similarity between early and later gradients, and the dominance of the first-epoch direction in the finetuning trajectory) (Zhou et al., 22 Aug 2025).
Generalization Bounds:

For task addition, subject to mild smoothness and low alignment ( $\epsilon$ ) between task vectors, the generalization gap is bounded as:

$\mathcal{L}_i(\theta_{\mathrm{Add}}) - \mathcal{L}_i(\theta_i) \le L_i C (1+\epsilon),$

where $C$ bounds the norm of $\tau_i$ (Zeng et al., 3 Feb 2025).

These properties underlie the theoretical reliability of task arithmetic in merging neural optima.

3. Practical Workflows: Implementation Paradigms and Extensions

3.1 Model Specialization and Merging

Fine-tuning is performed independently for each task to produce $\theta_{t_i}$ . Task vectors $\tau_{t_i}$ are computed and stored, enabling post hoc merging with arbitrary coefficients without joint retraining (Ilharco et al., 2022, Tao et al., 2024).

In multimodal or continual learning settings employing adapter strategies such as LoRA, low-rank updates $\Delta W_t = B_tA_t$ for each task are merged entrywise: $W^{\text{merge}} = W^{\text{PT}} + \sum_{i=1}^N \lambda_i \Delta W_i$ (Chitale et al., 2023, Cheng et al., 2024).

3.2 Augmented and Layer-aware Task Arithmetic

Empirical studies reveal that naive merging ("vanilla" task arithmetic) can cause interference (e.g., language confusion in multilingual ST, overload in LLM instruction-following). Solutions include:

Augmented Task Arithmetic with Language Control Vector:

Incorporates an auxiliary language control (LC) task, producing additional vector $\tau^{LC}$ and regularization weight $\lambda_{LC}$ to nudge outputs toward correct language tokens:

$\theta^{\text{merge}} = \theta^{\text{PT}} + \sum_{i} \lambda_i \tau^{ST}_{L_s \to L_{t_i}} + \lambda_{LC} \tau^{LC}$

Optimizing $\lambda_{LC}$ minimizes the language-confusion rate (Cheng et al., 2024).

Layer-Aware Task Arithmetic (LATA):

Assigns layer-wise weights to task vectors based on their cosine alignment with an instruction-following vector, amplifying task-specific and suppressing generic updates (Chen et al., 27 Feb 2025).

3.3 Task Vector Compression

For scalable deployment, collections of task vectors are approximated by linear combinations of a compact learned basis $\{b_m\}_{m=1}^M$ : $\tau_j \approx \sum_{m=1}^M W_e[j,m]\;b_m,$ where $W_e$ is learned via an autoencoder. Arithmetic transfers directly to this subspace, reducing storage and computation while retaining addition and unlearning guarantees up to the reconstruction error (Zeng et al., 3 Feb 2025).

3.4 Per-Parameter and Trust-Region Approaches

Per-parameter Task Arithmetic (PerTA): Elementwise weighting of subtracted task vectors (for unlearning) using gradient or Fisher curvature estimates:

$\theta_{\text{unlearned}} = \theta_{\text{original}} - W \odot \mathrm{TV},$

where $W$ is defined by e.g., absolute gradient norms or diagonal Fisher information from "forget" and "retain" datasets, balancing retention and removal (Cai et al., 29 Jan 2026).

Trust Region Merging (TATR): Projects each task vector orthogonally to the gradient subspaces of all other tasks before merging, mitigating cross-task knowledge conflicts:

$\theta_{\text{merge}} = \theta_{\text{pre}} + \sum_j P^{\perp}_{G_{-j}}\Delta_j,$

where $P^{\perp}_{G_{-j}}$ projects onto the orthogonal complement of other task gradients (Sun et al., 25 Jan 2025).

4. Application Domains

Task arithmetic provides a modular toolkit for various advanced tasks:

Domain	Core Mechanism	Reference
Multilingual speech translation	Merge one-to-one systems, add language-control, analogy-based zero-shot expansion	(Cheng et al., 2024)
Continual learning	LoRA adapter superposition, few-shot buffer	(Chitale et al., 2023)
Visual explainability transfer	Transfer explainability vector from source to target (zero-shot)	(Yoshikawa et al., 6 Jul 2025)
Zero-shot IR/domain adaptation	Domain/task injection via parameter shifts	(Braga et al., 1 May 2025)
Unlearning	(Per-)Parameter subtraction with per-coordinate weighting	(Cai et al., 29 Jan 2026)
Model merging for LLMs	Scaling/combining vectors with closed-form, model-exclusive weights	(Zhou et al., 2024)
Molecular property manipulation	Negate property-direction (task vector) from negative samples	(Özçelik et al., 23 Jul 2025)
Federated learning	Theoretically identical to one-shot FedAvg	(Tao et al., 2024)

In each scenario, task vectors provide a direct means to encode, combine, or remove model capabilities robustly, often with guarantees about interference minimization or preservation of utility.

5. Limitations, Regularization, and Empirical Phenomena

Local Linearity and Orthogonality Assumptions:

Task arithmetic relies on the local linearity of the parameter manifold and the near-orthogonality between different task vectors—supported empirically for related tasks and wide LLMs, but known to break down for highly heterogeneous or out-of-distribution tasks (Zhou et al., 2024, Zhou et al., 22 Aug 2025).

Knowledge Conflicts:

Direct addition of task vectors can cause interference if vectors are not sufficiently independent; trust-region and layer-aware algorithms partially alleviate this (Sun et al., 25 Jan 2025, Chen et al., 27 Feb 2025).

Regularization Strategies:

Grid search over vector scaling coefficients, pruning small-magnitude elements (parameter dropout), and task-vector compression are employed to improve robustness and utility preservation (Cheng et al., 2024, Zeng et al., 3 Feb 2025, Chen et al., 27 Feb 2025).

Data Efficiency and Privacy:

Merges require only fine-tuned parameters, not data, and are compatible with privacy/constrained federated settings (Tao et al., 2024, Zhou et al., 2024).

6. Empirical Outcomes and Evaluation Metrics

Translation Benchmarks:

On MuST-C and CoVoST-2 for speech translation, augmented task arithmetic yields BLEU improvements up to 4.92 and COMET gains up to 11.83 versus baselines (Cheng et al., 2024).

Continual Learning:

Minimal buffer fine-tuning after LoRA task arithmetic matches offline accuracy within 1–3% using only 2% parameter updates per task (Chitale et al., 2023).

Zero-shot IR:

Task arithmetic merging yields 18% relative gains in NDCG@10 for domain adaptation without additional fine-tuning (Braga et al., 1 May 2025).

Model Editing/Explainability:

Explainability or safety vectors constructed via arithmetic transfer successfully, with qualitative and quantitative alignment to target domains (Yoshikawa et al., 6 Jul 2025, Chen et al., 27 Feb 2025).

Unlearning:

Per-parameter task arithmetic approaches (PerTA) achieve strong trade-offs between forgetting effectiveness and retention of general utility, surpassing conventional retraining protocols (Cai et al., 29 Jan 2026).

Generalization:

Basis compression frameworks maintain or improve downstream performance with fewer stored vectors, meeting derived theoretical error bounds (Zeng et al., 3 Feb 2025).

Evaluation metrics are domain specific: BLEU, COMET, NDCG@10, forget quality/log-KS, average loss difference, and ablation studies on knowledge retention and removal are all standard.

7. Interpretive Summary and Future Directions

Task arithmetic represents a unifying geometric formalism for post hoc model editing, composition, and knowledge transfer in deep learning. Its efficacy derives from the locally linear structure of weight space near pre-trained optima, the approximately orthogonal specialization directions induced by task-specific fine-tuning, and the compatibility of linear updates with multitask optimization and distributed training paradigms. Under mild regularity, task arithmetic yields practical, theoretically-justified solutions to model merging, zero-shot adaptation, safety editing, explainability transfer, and privacy-aware unlearning, with continuing extensions in basis compression and trust-region regularization. Key open questions pertain to its limits for highly nonlinear or adversarial tasks, extension to more complex parameter manifolds (e.g., in the presence of large architectural shifts), and automated strategies for optimal task mixture weighting and selection.

References:

(Ilharco et al., 2022, Cheng et al., 2024, Chitale et al., 2023, Zhou et al., 22 Aug 2025, Cai et al., 29 Jan 2026, Zeng et al., 3 Feb 2025, Chen et al., 27 Feb 2025, Tao et al., 2024, Braga et al., 1 May 2025, Sun et al., 25 Jan 2025, Yoshikawa et al., 6 Jul 2025, Zhou et al., 2024, Özçelik et al., 23 Jul 2025).