🤖 machine learning

Resolving Interference (RI): Disentangling Models for Improved Model Merging

The paper proposes Resolving Interference (RI), a lightweight framework that uses unlabeled auxiliary data to disentangle expert models into functionally orthogonal spaces, thereby reducing cross-task interference and significantly improving the performance and generalization of model merging without requiring task-specific data.

Original authors: Pratik Ramesh, George Stoica, Arun Iyer, Leshem Choshen, Judy Hoffman

Published 2026-03-17

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Pratik Ramesh, George Stoica, Arun Iyer, Leshem Choshen, Judy Hoffman

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a team of world-class chefs. One is a master of Italian pasta, another is a sushi expert, and a third is a BBQ pitmaster. Each of them has their own unique set of tools, recipes, and muscle memory developed over years of practice.

Now, imagine you want to create a "Super Chef" who can cook all of these dishes perfectly at the same time. The obvious idea is to just take their aprons, mix their ingredients together, and hope for the best. This is what Model Merging is in the world of AI.

However, there's a problem. If you just mash their aprons together, the Italian chef's knife skills might get confused by the sushi chef's delicate handling of fish. The BBQ chef's heavy-handed seasoning might ruin the pasta. In AI terms, this is called Cross-Task Interference. The "Super Chef" ends up making mediocre pasta, bad sushi, and dry ribs because the skills are fighting each other.

The Problem: The "Clashing Aprons"

In the paper, the authors explain that when you combine AI models trained on different tasks (like recognizing cats vs. recognizing dogs), the internal "brain waves" of the models clash. The features learned for one task overwrite or distort the features needed for another.

Existing solutions try to fix this by:

Ignoring the data: Just averaging the numbers (like mixing the aprons blindly). This often fails because it doesn't understand how the skills interact.
Using the original data: Trying to retrain the mix using the specific recipes (e.g., thousands of cat photos). But often, you don't have access to those original photos (maybe they are private or too expensive to get).

The Solution: "Resolving Interference" (RI)

The authors propose a clever, lightweight fix called Resolving Interference (RI).

Think of it as a rehearsal before the Super Chef goes on stage.

Instead of trying to mix the chefs' skills directly, RI takes each chef individually and gives them a special training session using a random, unlabeled pile of pictures (like a bucket of random photos from the internet—cars, trees, clouds, etc.). This pile is called Auxiliary Data.

Here is the magic trick of the rehearsal:

The "Keep Your Skills" Rule: The Italian chef looks at a random photo of a car and must still be able to cook pasta if asked. They must keep their core identity.
The "Don't Get Confused" Rule: When that same Italian chef looks at the car photo, they must not accidentally start thinking about sushi or BBQ. They must learn to ignore the other chefs' domains.

By doing this, the Italian chef learns to be functionally orthogonal (a fancy way of saying "at a right angle") to the other chefs. Their skills become distinct and separate, like different lanes on a highway that don't cross over.

Once every chef has gone through this rehearsal, they are finally merged. Because they have learned to keep their skills distinct and not get confused by the others, the resulting "Super Chef" is much better at cooking everything.

Why is this a big deal?

No Secret Recipes Needed: You don't need the original, private data (the thousands of cat/dog photos) to do this. You just need any random pictures. This is huge for privacy and data-scarce situations.
It's Cheap and Fast: The "rehearsal" is very short. It only takes a few minutes on a standard computer chip, even for large models.
It Works Everywhere: The paper shows that this method improves the performance of almost every existing merging technique. It makes the "Super Chef" up to 3.8% better at their jobs and 2.3% better at handling completely new types of food (unseen domains) they've never seen before.
It's Robust: Even if you don't tune the settings perfectly, the method still works well. It's like a recipe that tastes good even if you don't measure the salt perfectly.

The Analogy in a Nutshell

The Old Way: Mixing three different languages into one sentence and hoping it makes sense. (Result: Gibberish).
The RI Way: Teaching each language speaker to speak their own language clearly, while ignoring the others, before they try to speak together. (Result: A harmonious conversation where everyone is understood).

In short, Resolving Interference is a simple, data-free "detox" for AI models that clears up the confusion between their different skills, allowing them to work together as a true team rather than a chaotic mess.

1. Problem Statement

Model Merging involves combining parameters from multiple independently trained "expert" models (each specialized in a specific task) into a single unified model. While this offers a cost-effective way to create multitask models, it often suffers from Cross-Task Interference.

The Core Issue: When models trained on distinct tasks are merged, their learned features can conflict. For example, parameters optimized for distinguishing dog breeds might overwrite or distort features necessary for cat breeds.
The Consequence: This interference causes the merged model's output representations to drift away from those of the constituent experts, leading to significant performance degradation on individual tasks.
Limitations of Existing Solutions:
- Gradient-free methods (e.g., weight averaging, TIES) operate without data, limiting their ability to understand how parameters interact with inputs.
- Gradient-based methods often require access to the original task-specific training data, which is frequently unavailable due to privacy constraints or data scarcity.
- Routing-based methods add capacity and inference costs.
The Gap: There is a need for a lightweight, gradient-based adaptation strategy that reduces interference without requiring task-specific data and without increasing inference cost.

2. Methodology: Resolving Interference (RI)

The authors propose Resolving Interference (RI), a lightweight adaptation framework that disentangles expert models before they are merged.

A. Formal Definition of Interference

The paper formally defines Cross-Task Interference ( $\xi$ ) as the deviation between the output representations of the merged model and the original expert models when evaluated on their respective tasks:
$\xi = \sum_{i} \mathbb{E}_x [\text{dist}(h_i(f(x|\theta_i)), h_i(f(x|\theta_m)))]$
Where $\theta_i$ is the expert, $\theta_m$ is the merged model, $h_i$ is the task head, and $\text{dist}$ is a distance metric (e.g., KL-Divergence). Minimizing $\xi$ ensures the merged model retains the capabilities of its constituents.

B. The RI Framework

Instead of merging first and then fixing, RI adapts each expert model individually to be functionally orthogonal to other tasks before merging.

Task-Preservation Objective: The adapted task vector $\tau^*_i$ must preserve the output of the original expert $i$ on its own task.
Interference-Reduction Objective: The adapted task vector $\tau^*_i$ must have zero influence on the outputs of all other task heads ( $j \neq i$ ).
Data Constraint: Crucially, this adaptation is performed using unlabeled auxiliary data ( $D_{aux}$ ) rather than task-specific data. The assumption is that enforcing orthogonality on general data transfers to the specific task distributions.

C. Optimization Loss

The method optimizes a composite loss function ( $L_{RI}$ ) for each expert $i$ :
$L_{RI} = L_1 + \frac{\alpha}{N-1} L_2$

$L_1$ (Task-Preservation): Minimizes the distance between the original expert's output and the adapted expert's output on $D_{aux}$ using the specific task head $h_i$ .
$L_2$ (Interference-Reduction): Minimizes the distance between the adapted expert's output and the pretrained backbone's output on $D_{aux}$ when passed through other task heads $h_j$ .
$\alpha$ : A hyperparameter balancing the two objectives (default set to 1).

D. Workflow (Algorithm 1)

Initialize adapted vectors $\tau^*_i$ from original task vectors $\tau_i$ .
Sample unlabeled data from $D_{aux}$ .
Compute $L_1$ and $L_2$ and update $\tau^*_i$ via gradient descent.
Repeat until convergence (typically 2,500 steps).
Merge the adapted vectors $\{\tau^*_1, ..., \tau^*_N\}$ using standard merging techniques (e.g., Task Arithmetic, TIES).

3. Key Contributions

Formalization of Interference: Introduced a quantitative metric ( $\xi$ ) to measure cross-task interference, providing a principled diagnostic for merge quality.
RI Framework: Proposed a novel, lightweight adaptation strategy that disentangles expert models into disjoint functional subspaces using only unlabeled auxiliary data.
Data-Efficiency: Demonstrated that high-quality merging is possible without access to proprietary or scarce task-specific training data.
Robustness: Showed that RI-merged models are significantly less sensitive to the tuning of merging hyperparameters compared to standard methods.

4. Experimental Results

The authors evaluated RI on CLIP vision encoders (ViT-B/32, ViT-B/16, ViT-L/14) across 8, 14, and 20 vision classification tasks, as well as out-of-distribution (OOD) benchmarks.

In-Domain Performance:
- RI consistently improved state-of-the-art (SOTA) merging methods.
- Gains: Up to 3.8% improvement in average accuracy over baseline merging methods (e.g., TSV-M, Iso-CTS).
- Specific improvements were notable for Task Arithmetic (up to 9.6% gain) and TIES.
Out-of-Distribution (OOD) Generalization:
- Tested on the DomainNet benchmark (6 domains).
- RI improved generalization to unseen domains by up to 2.3% compared to standard merging.
- This suggests that reducing interference facilitates better positive transfer of shared representations.
Ablation Studies:
- Distance Metric: KL-Divergence was found to be the most effective metric for the loss function.
- Auxiliary Data: RI works with diverse data sources (synthetic noise, ImageNet, MSCOCO). While task-specific data yields the best results (Oracle), general visual diversity in auxiliary data is sufficient for significant gains.
- Hyperparameter Sensitivity: Models adapted with RI showed a 0.4% performance variance when hyperparameters were tuned vs. default, compared to 1.8% for non-adapted models. This indicates RI creates a flatter, more stable optimization landscape.
- Scalability: The method is computationally efficient (sub-linear time growth with task count) and requires constant memory usage (~4.8 GB VRAM for ViT-B/32).

5. Significance

This paper addresses a critical bottleneck in the democratization of multitask AI: the inability to merge models without access to their original training data.

Practical Impact: RI enables the creation of robust, high-performing multitask models in data-scarce or privacy-constrained environments (e.g., medical imaging, proprietary enterprise data).
Theoretical Insight: It establishes that "disentangling" experts into orthogonal functional subspaces is a viable strategy to mitigate parameter conflicts, offering a new perspective beyond simple weight averaging or pruning.
Efficiency: By being lightweight and requiring no additional inference capacity, RI is a practical drop-in enhancement for existing model merging pipelines.