Task Vectors

A task vector is formally defined as the difference between the weights of a model fine-tuned on a target task and the weights of the original base model: θ_task = θ_fine-tuned - θ_base. This additive parameterization isolates the learned change, representing it as a directional vector in the high-dimensional parameter space. The vector's magnitude indicates the strength of the update, while its direction encodes the specific functional shift required for the task.

The linear structure of task vectors allows for their combination through simple arithmetic operations, a core capability known as task arithmetic. This enables powerful model editing and multi-task adaptation without retraining.

• Addition: Combining task vectors (e.g., θ_sentiment + θ_formality) can create a model proficient in a composite task (e.g., formal sentiment analysis).
• Negation & Interpolation: Applying θ_base - θ_toxic can reduce unwanted behavior, while interpolating between vectors (e.g., 0.7*θ_taskA + 0.3*θ_taskB) blends model behaviors.
• This treats model adaptation as an algebraic operation in weight space.

Task vectors are inherently parameter-efficient. They are derived from delta tuning methods like LoRA or (IA)^3, where only a small subset of parameters (the delta) is updated. The resulting vector is therefore compact, often representing updates to <1% of the total model parameters. This makes storage, sharing, and combining multiple task vectors for a single base model highly efficient, avoiding the cost of maintaining multiple full-model copies.

Task vectors facilitate a modular adaptation paradigm. Each vector acts as a self-contained, plug-and-play module representing a discrete skill or capability. A single frozen base model can be dynamically adapted by applying different vectors at inference time, akin to selecting software modules. This architecture is central to systems like AdapterSoup, where an ensemble of adapter-based task vectors is averaged for robust multi-task performance.

By isolating the change from a base model, task vectors offer a window into what the model learned for a specific task. Analyzing vector magnitudes across layers can reveal which parts of the network were most critical for adaptation. This supports model editing—making precise, targeted adjustments to model behavior (e.g., correcting factual errors, removing bias) by surgically adding or subtracting carefully constructed vectors, providing more control than black-box fine-tuning.

The standard workflow involves:

1. Fine-tuning a base model on a target task using a delta tuning method.
2. Extracting the task vector by subtracting the base weights from the tuned weights.
3. Applying the vector to the base model for inference: θ_applied = θ_base + θ_task.

Key challenges include:

• Catastrophic forgetting: Applying a new vector can interfere with previously learned tasks.
• Interference: Vectors for conflicting tasks may not combine linearly.
• Scale sensitivity: The effectiveness of arithmetic operations can depend on vector normalization.

The overarching paradigm where a model is adapted by learning a small, task-specific parameter change (a delta), applied to a subset of the pre-trained model's frozen weights. Task vectors are a direct product of this paradigm.

• Core Principle: Learn ΔW, not W.
• Efficiency: Updates a tiny fraction of total parameters (often <1%).
• Foundation: Enables techniques like LoRA, (IA)^3, and the creation of task vectors.

A technique for model editing and multi-task adaptation that uses task vectors as its fundamental operands. It treats adaptation as a linear algebraic operation.

• Operations: Task vectors are added, subtracted, or interpolated.
• Example: Vector(Sentiment) + Vector(Sarcasm) ≈ Vector(Irony).
• Application: Enables zero-shot task composition and behavior steering without additional training.

The mathematical framework that defines an adapted model's weights as the sum of the original frozen weights and a learned delta matrix. This is the formal representation of a task vector.

• Formula: W_adapted = W_base + ΔW_task.
• Key Insight: The task vector ΔW_task encapsulates the directional change needed for adaptation.
• Implementation: Used explicitly in LoRA (ΔW = BA) and implicitly in other delta methods.

A class of techniques for making precise, localized updates to a neural network's knowledge or behavior post-training. Task vectors provide a mechanism for coarse-grained, layer-wise editing.

• Scope: Ranges from editing single factual associations to full task adaptation.
• Task Vectors vs. Local Editing: Task vectors are global edits applied across many layers, unlike methods that target specific neurons.
• Goal: Achieve desired behavioral change with minimal side effects on unrelated capabilities.

A PEFT approach that extends a base model with small, self-contained, and composable neural modules (like adapters) tuned for specific tasks. Task vectors can be seen as a weight-space instantiation of a module.

• Composability: Independent modules can be combined or switched.
• Relation to Task Vectors: A trained adapter's weights, when subtracted from the base model's weights, can form a task vector.
• Systems: Enables AdapterFusion and AdapterSoup for multi-task learning and inference.

A secondary neural network that generates the weights or parameters for a primary model. In PEFT, a hypernetwork can be trained to output task-specific delta weights or adapter parameters conditioned on a task descriptor.

• Dynamic Generation: Produces adaptation parameters on-the-fly without storing them.
• Intrinsic Task Vectors: The hypernetwork's conditioning input (e.g., a task embedding) acts as a compressed, intrinsic task vector.
• Efficiency: Shares hypernetwork parameters across many tasks, further reducing storage.