💻 computer science

From Memorization to Parameter Interference: How Overtraining Experts Harms Model Merging

This paper challenges the assumption that optimizing expert models for individual performance benefits downstream model merging, demonstrating instead that overtraining leads to memorization and negative parameter interference that degrade merged performance, a problem effectively mitigated by task-dependent early stopping.

Original authors: Stefan Horoi, Guy Wolf, Eugene Belilovsky, Gintare Karolina Dziugaite

Published 2026-06-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Stefan Horoi, Guy Wolf, Eugene Belilovsky, Gintare Karolina Dziugaite

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

The Big Idea: "Less is More" (Even for AI)

Imagine you have a team of expert chefs. Each chef has been trained on a specific recipe: one makes the perfect pizza, another the perfect sushi, and another the perfect cake.

In the world of AI, we often take a "base" model (a general knowledge chef) and fine-tune it to become an expert in one specific task. Then, we want to combine all these different expert chefs into one "Super Chef" who can do everything at once. This process is called Model Merging.

The common belief was: "The better the individual chef gets at their specific recipe, the better the Super Chef will be."

This paper proves that belief is wrong. In fact, if you train your individual experts for too long, the resulting Super Chef actually becomes worse.

The Problem: The "Overtrained" Expert

The authors discovered a phenomenon they call "Overtraining."

Think of it like a student cramming for a test.

Early in training: The student learns the general rules of math. They understand the concepts.
Late in training: The student stops learning new concepts and starts memorizing the specific, weird, difficult questions from the practice test. They memorize the exact wording of the hardest problems, even the ones with typos or confusing phrasing.

In AI terms, when an expert model is trained for too many steps, it stops learning general patterns and starts memorizing a small set of very difficult, weird examples from its training data.

The Collision: Why Merging Fails

Now, imagine you try to combine these "memorizing" chefs into a Super Chef.

Chef A memorized that "Pizza" always has a specific, weird stain on the crust because of one bad photo in their training data.
Chef B memorized that "Sushi" always has a specific, weird texture because of one noisy image.

When you try to merge them, their brains (the computer parameters) start fighting. Chef A says, "The crust must have a stain!" Chef B says, "No, the texture must be weird!"

Because they memorized specific, idiosyncratic details rather than general rules, their instructions cancel each other out. This is called Negative Parameter Interference.

The result? The Super Chef forgets the hard stuff entirely. It can still make simple pizza and simple sushi (the easy examples), but it completely fails at the difficult tasks because the conflicting "memorized" instructions destroyed the knowledge.

The Evidence: The "Hard Example" Trap

The researchers used a tool to measure how "hard" a training example was. They found:

Early in training: The model learns the easy examples quickly.
Late in training: The model spends all its time obsessing over the hardest 10% of examples.
The Merging Disaster: When they merged models that had been trained for a long time, the "hard examples" were the first to be forgotten. The model lost the ability to handle difficult cases because the conflicting memorized data wiped them out.

Interestingly, they found that you actually need some memorization to get good results. If you remove the hard examples entirely, the model fails. But if you let the model memorize them too much (by training too long), it breaks the merging process.

The Solution: Stop Early!

The paper suggests a simple fix: Aggressive Early Stopping.

Instead of training your expert models until they are perfect at their specific task, you should stop training them much earlier.

The Old Way: Train until the expert gets 90% accuracy. (Result: Bad Super Chef).
The New Way: Train until the expert gets 85% accuracy and then stop. (Result: Great Super Chef).

By stopping early, the experts haven't had time to memorize those weird, difficult examples. They retain the general rules that work well for everyone. When you merge them, they agree on the general rules, and the Super Chef becomes much more capable.

Key Takeaways

Better Experts $\neq$ Better Merges: Just because an individual AI model is better at its specific job doesn't mean it will help build a better combined model. Sometimes, "good enough" is actually better for the team.
Memorization is the Enemy of Merging: The more a model memorizes specific, difficult data points, the harder it is to combine with other models.
Stop Training Sooner: Whether you are using full model training or efficient "LoRA" adapters (a lightweight way to train), stopping the training process early yields better results when merging.
It Works Everywhere: This happens with both image models (like recognizing cars or cats) and language models (like answering questions), and it happens regardless of the size of the model.

In short: Don't let your AI experts become too obsessed with the hardest details. Keep them focused on the big picture, and when you combine them, they will work together much better.

Technical Summary: From Memorization to Parameter Interference

Problem Statement

Modern deep learning relies heavily on a pipeline where open-weight foundation models are pre-trained on general data and subsequently fine-tuned on specialized tasks to create "expert" models. These experts are often merged to combine capabilities into a single, more capable model. A prevailing assumption in this workflow is that improvements at the fine-tuning stage (i.e., achieving higher accuracy on the expert's specific task) propagate downstream to improve the performance of the merged model.

This paper challenges that assumption. The authors investigate the relationship between the duration of expert fine-tuning and the efficacy of model merging. They posit that "overtraining" experts—optimizing them to convergence or beyond on their individual tasks—may actually degrade the performance of the resulting merged model, despite the experts performing better on their native tasks.

Methodology

The study employs a systematic empirical analysis across vision and language modalities, utilizing various model families (ViT-B-32, ViT-L-14, T5-Base/Large/3B, BERT-Base) and adaptation regimes (Full Fine-Tuning and Low-Rank Adaptation/LoRA).

Experimental Setup:

Datasets: 8 image classification tasks (e.g., Cars, MNIST, SVHN) and 7 NLP tasks (e.g., QASC, Winogrande).
Training Duration: Experts were fine-tuned for varying numbers of steps ( $s \in \{2, \dots, 2048\}$ ).
Merging Methods: The authors evaluated multiple merging techniques, including simple Parameter Averaging, Task Arithmetic (TA), TIES, DARE, and SVD-based methods (Iso-C, Iso-CTS, TSV-M).
Evaluation: The primary metric is the raw, un-normalized average test accuracy across all tasks of the merged model. The authors explicitly avoid normalized accuracy metrics to prevent inconsistencies arising from varying expert performance across experiments.

Analytical Tools:
To explain the observed phenomena, the authors utilize data difficulty scores (specifically the EL2N score) to categorize training examples into bins of difficulty (easiest to hardest). They analyze:

Memorization: Using the Feldman memorization framework to measure how much a model's confidence on a specific example depends on having seen that example during training.
Parameter Interference: Quantifying the conflict between task vectors using metrics such as sign conflict percentage, parameter overlap, magnitude ratio, and per-parameter variance.
Causal Isolation: Conducting experiments where specific subsets of data (easy, medium, hard) are held out to measure their individual impact on model behavior and merging outcomes.

Key Findings

1. Overtraining Harms Merging

Contrary to the assumption that better experts yield better merges, the authors find that longer fine-tuning leads to degraded merging performance.

Vision & Language: In both domains, merging accuracy peaks at intermediate training steps (e.g., 256 steps for T5, 128 for ViT-LoRA) and declines significantly as training continues to 2048 steps.
Magnitude of Impact: The degradation is severe. For LoRA-adapted models, merging accuracy can drop by over 16% (e.g., TIES method) when moving from optimal early stopping to overtrained checkpoints.
Expert vs. Merged Performance: While expert accuracy on their specific tasks continues to improve with longer training, the merged model's performance deteriorates. For instance, in NLP, expert accuracy improved by 3.7% from 256 to 2048 steps, yet merging accuracy dropped by at least 2.4% for all methods.

2. The Mechanism: Memorization and Interference

The paper traces this degradation to the memorization of difficult examples during late-stage training.

Memorization of Hard Examples: Early training is dominated by learning common features (easy examples). Later training stages are driven by the memorization of a small fraction of "hard" examples (outliers or noisy labels).
Idiosyncratic Updates: The memorization of these hard examples leads to idiosyncratic parameter updates that do not generalize across tasks.
Negative Parameter Interference: When merging, these task-specific, memorized updates cause negative interference. The authors show that hard examples are disproportionately forgotten during the merging process (over 50% of forgotten data points belong to the hardest 30% of examples), while easy examples remain correctly classified.
Quantitative Evidence: Parameter interference metrics (sign conflict, overlap, variance) increase systematically with training duration, confirming that longer training creates more conflicting parameter updates.

3. The Necessity of Some Memorization

The authors demonstrate that while excessive memorization is harmful, some memorization is necessary for optimal merging. Removing the hardest examples from training consistently hurts merging performance, suggesting that a balance is required where generalizable features are learned without overfitting to idiosyncratic noise.

4. Early Stopping as a Solution

The paper proposes aggressive early stopping as a general principle to mitigate overtraining.

Task-Dependent Strategies: By stopping training when validation accuracy plateaus (rather than training for a fixed number of steps), the authors achieve merging performance that matches or exceeds the best results obtained by fixed-step training, even though the individual experts are less accurate on their specific tasks.
LoRA Specifics: Early stopping is particularly effective for LoRA adapters, recovering optimal merging accuracy that is otherwise lost with standard long-duration training.

Significance and Claims

The paper claims to introduce a critical new dimension to the field of model merging: the duration of expert training.

Paradigm Shift: It challenges the standard practice of training experts to convergence before merging. The authors argue that the "best" expert for merging is often an "undertrained" expert that has not yet memorized difficult, idiosyncratic examples.
Practical Guidance: The work provides actionable guidance for practitioners:
1. Publish Intermediate Checkpoints: Since the optimal merging point often precedes convergence, releasing intermediate checkpoints is crucial.
2. Prioritize Early Stopping: When training experts for merging, aggressive early stopping can yield better downstream results than full convergence.
Efficiency: By reducing the training duration required for optimal merging, these findings can significantly reduce the computational and environmental costs associated with training and merging AI models.
Generalizability: The phenomenon is shown to be robust across model scales (from 220M to 3B parameters), architectures (ViT, T5, BERT), and adaptation methods (FFT, LoRA), as well as alternative upcycling strategies like Model MoErging.

The authors conclude that careful management of expert training dynamics, specifically avoiding the late-stage memorization of hard examples, is essential for unlocking the full potential of model merging pipelines.