How to Fine-Tune an AI Model on New Tasks Without Losing Its Existing Abilities

Apple researchers have developed DynaMiCS, a method for fine-tuning AI models on new tasks without losing their core abilities, instruction-following or safety knowledge.

Why do existing approaches fall short?

Fine-tuning an LLM on several domains at once is routine work for any lab that ships AI products. The goal sounds simple: improve the model on target domains (medicine or law, for example) while preserving the so-called "constraint domains", core abilities that must not be degraded under any circumstances.

In practice, this balancing act is hard. Current data-mixing strategies offer two unsatisfying options. The first is fixed heuristics: predetermined data proportions per domain that stay constant throughout training and do not react to the model's current state. The second is adaptive rules: algorithms that adjust weights based on metric signals, but without explicit formal guarantees.

Neither approach can explicitly keep the constraint domains at a given level. As a result, teams either tune the ratios by hand or accept unpredictable losses on key safety and instruction benchmarks.

How DynaMiCS works

DynaMiCS reformulates fine-tuning as a constrained optimization problem. Instead of setting the trade-off between domains softly, through a combined loss or manual weights, the algorithm separates objectives from constraints explicitly.

At each training step, DynaMiCS works in three stages:

• It launches short "probing runs" on each domain, estimating the local dynamics of the metrics.
• It builds a slope matrix, an estimate of how changing the share of one domain's data affects the performance of all the others.
• It solves an optimization problem: find the data ratio that improves the target domains as much as possible while strictly satisfying the constraints on the constraint domains.

The key novelty lies in explicit mathematical constraints rather than soft regularization. The algorithm does not "try" to preserve safety or instruction-following; it guarantees this formally at every iteration. The probing runs are fast and do not require a full training cycle, so the method is practical without a significant increase in compute cost.

Where it applies

The problem DynaMiCS addresses arises every time a company adapts a base model to a specific product:

• Instruction-following is a core skill that must not degrade even under narrow specialization.
• General knowledge: an assistant must answer questions correctly outside its specialized domain.
• Safety: safety evaluation results must be preserved through any fine-tuning, especially in regulated industries.

For Apple, which is actively building language models into Apple Intelligence, the system AI layer of iPhone, iPad and Mac, such a tool has direct practical value. Models must specialize for the tasks of individual apps while staying general-purpose and passing internal safety checks. DynaMiCS offers a formal way to meet both requirements at once.

The approach is relevant beyond Apple: any company adapting models for enterprise tasks with strict safety and instruction requirements faces the same balancing problem.

What this means

DynaMiCS offers a mathematically rigorous approach to a problem most ML teams have handled by hand or by trial and error. The Apple ML Research publication shows that work on core fine-tuning infrastructure continues at large tech companies with closed models. If the method is reproducible and scales to real model sizes, it may find a place in the standard pipelines of companies working with multi-domain LLMs and strict requirements for preserving core abilities.

What is AI fine-tuning?

It is the process of training a ready-made language model on new data to specialize it for specific tasks and domains.

How do you avoid losing quality during fine-tuning?

You need to optimize the data mixture correctly, which preserves the model's core abilities. This is what the DynaMiCS method does.