🤖 machine learning

On the Power of Foundation Models

This contribution uses category theory to show that while prompt-based learning is strictly limited to representable tasks, a sufficiently powerful foundation model with fine-tuning can theoretically solve any downstream task within the category defined by its pretraining task and generalize to unseen objects through structural mapping.

Original authors: Yang Yuan

Published 2026-04-27

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Yang Yuan

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 Question: Can a Super-Brain Do Everything?

Imagine a student who has read every book in the universe, has infinite time to learn, and never makes a single mistake on practice tests. The authors pose a simple question: If this student is perfect on their practice tests, can they automatically solve every new problem you present to them?

In the world of AI, this "student" is a Foundation Model (like those behind ChatGPT or image generators). The "practice tests" are called Pretext Tasks (e.g., predicting the next word in a sentence or guessing how an image was rotated).

The paper argues that existing theories (regarding the amount of data or the size of the computer) cannot answer this. Instead, the authors use a branch of mathematics called Category Theory (think of it as the "grammar of structures") to determine what these models can and cannot do.

They discovered two main rules for how these models learn new things:

Rule #1: The "Prompt" Limit (The Library Card Analogy)

The Scenario: You want the model to perform a new task (e.g., identifying cats in photos), but you do not want to retrain it. You simply give it a specific instruction or a "prompt" (like a library card that says: "Find cats").

The Insight: The model can only solve the new task if that task is already "hidden" within the structure of its original training.

The Analogy: Imagine the model as a Library.

The Pretext Task (the training) is the way the library organized its books. If the library organized books only by "color," the books are sorted into Red, Blue, and Green.
Prompt Tuning is like asking the librarian: "Can you find me a book about History?"
The Result: If the library was organized only by color, the librarian cannot find a book about history, even if they have the best memory in the world. The category "History" simply does not exist in the library's structure.
The Paper's Proof: The authors prove that if the training task (like guessing image rotations) only teaches the model about "rotating," the model cannot be made to perform complex tasks like "segmentation" (cutting out) objects, because the concept of "cutting out" was not built into the library's structure.

Conclusion: If you only teach the model to play chess, you cannot simply make it play soccer via a "prompt." The new task must be representable through the old structure.

Rule #2: The Power of "Fine-Tuning" (The Master Key Analogy)

The Scenario: You are willing to give the model a little new training (fine-tuning) with a small dataset for the new task.

The Insight: This is much more powerful. If the model learned the "structure" of the world well enough during its original training, it can learn any new task, provided you give it enough resources (data and computing power) to connect the dots.

The Analogy: Imagine the model as a Master Key filed to fit the locks of a specific building (the Pretext Task).

Fine-Tuning is like taking this master key and filing it slightly so it fits a new door in a different building.
The Result: As long as the master key captured the essence of the first building's locks, you can reshape it to open almost any door in the world.
The Paper's Proof: The authors show that if the model is "ideal" (having perfectly learned the training structure), it contains all the necessary information to solve any downstream task. The training data for the new task merely serves as a guide to show the model how to reshape this information.

Conclusion: Fine-tuning is not limited by the "representability" of the prompt. If the foundation is strong, the model can be adapted to almost anything.

Rule #3: The "Magic Bridge" (The Translator Analogy)

The Scenario: What happens when we combine different types of models, such as one that understands text and one that understands images (Multimodal Learning)?

The Insight: The paper presents a "Generalization Theorem." It states that if you build a perfect bridge (a mathematical mapping) between two different worlds (e.g., text and images), the structure of one world is preserved in the other.

The Analogy: Imagine a Dictionary that perfectly translates "Text" into "Images."

In the text world, you have the concept of an "Avocado Chair" (a chair shaped like an avocado). This object might not exist in the real world, and no photo of it exists in your training data.
However, since the text world has a logical structure where "Avocado" and "Chair" can be combined, and your dictionary (the model) perfectly preserves this structure when translating into the image world...
The Result: The model can "hallucinate" or generate a perfect image of an Avocado Chair, even though it has never seen one. It did not memorize the image; it understood the relationship between the words and applied that structure to the image world.

The Paper's Proof: This explains why models like DALL-E 2 or CLIP can create images of things that do not exist in their training sets. They do not just copy; they use the structural logic of language to build new images.

Summary of the Paper's Claims

Prompting is limited: You can only prompt a model to do things that are already "built-in" to the way it was trained. If the training did not teach it the structure of the new task, a simple prompt will not work.
Fine-tuning is powerful: If you are willing to do a little extra training, a model that has learned a good structure can be adapted to solve almost any task.
Structure creates the new: By perfectly mapping the structure of one domain (like text) onto another (like images), models can generate things that have never existed before (like an Avocado Chair) because they follow the logical rules of the structure rather than just memorizing images.

The authors emphasize that they are not talking about specific network sizes or data amounts, but rather the logical structure of the tasks themselves. They use mathematics to prove that the "shape" of the training task determines the "shape" of what the model can ultimately do.

1. Problem Statement

The paper addresses a fundamental question in the era of Foundation Models: Given infinitely high-quality data, infinite compute, and an infinitely large model with a perfect training algorithm (guaranteeing zero generalization error on the pretraining task), can such a model solve any downstream task?

Existing theories of deep learning (representation, optimization, and generalization) are insufficient to answer this, as they typically assume specific data distributions, network architectures, or finite resources. The authors argue that the limitation lies not in the model's learning capacity, but rather in the structural representation of the tasks themselves. The core problem is to determine which downstream tasks can be solved by a Foundation Model solely based on the structure induced by its pretraining task (pretext task).

2. Methodology: A Categorical Framework

The authors employ Category Theory as the primary mathematical tool for analyzing Foundation Models. This approach enables a "model-independent" perspective that focuses on the structural relationships between data points, rather than on specific neural network architectures or loss functions.

Defined Key Concepts:

Category Induced by the Foundation Model ( $C_f$ ): A Foundation Model $f: D \to H$ $f : D \to H$ (mapping data to representations) induces a category, where:
- Objects: Data points modulo representation equivalence (i.e., $X \sim X'$ if $f(X) = f(X')$ ).
- Morphisms: Defined by a "data-independent readout" $k_f$ and composition maps derived from the operations of the pretext task (e.g., translation in feature space for contrastive learning or token appending for language modeling).
Yoneda Embedding ( $h_C$ ): The paper utilizes the Yoneda Lemma to interpret the model's representation $f(X)$ as a representable functor $h_C(X) = \text{Hom}_C(X, \cdot)$ . This maps the internal feature space to a category of functors ( $C^\vee$ ), effectively encoding the structural relationships of the data.
Task Representation: Downstream tasks are defined as functors $T \in C^\vee$ . Solving a task means the model can produce an output isomorphic to $T(X)$ .

3. Main Contributions and Results

The paper establishes three main theorems regarding the capabilities of Foundation Models under two adaptation strategies: Prompt Tuning and Fine Tuning.

A. The Limits of Prompt Tuning (Theorem 3)

Result: A Foundation Model can solve a downstream task $T$ via Prompt Tuning if and only if the task $T$ is representable within the category induced by the pretext task.
Implication: Prompt Tuning is strictly limited by the structural richness of the pretext task. If the pretext task (e.g., rotation prediction) induces only a simple group structure, the model cannot solve complex downstream tasks (e.g., semantic segmentation) via prompting alone, regardless of model size.
Corollary 1: Specifically, pretext tasks for rotation prediction do not guarantee the ability to solve content-related tasks like classification or segmentation via prompting, as the induced category lacks the necessary structural complexity.

B. The Power of Fine Tuning (Theorem 5)

Result: In contrast to Prompt Tuning, Fine Tuning has no such structural constraints. If a Foundation Model possesses the minimal required performance to solve the pretext task (up to symmetry) and is provided with sufficient resources (data and compute) for the downstream task, it can theoretically solve any downstream task defined within the category $C^\vee$ .
Mechanism: Fine Tuning enables the model to learn an extension functor (via the Yoneda Extension Theorem) that maps the pretrained representation space to the specific target task space.
Significance: This suggests that the "bottleneck" of Foundation Models is not the Fine Tuning capability itself, but the quality of the pretext task. A rich pretext task generates a rich category $C$ , which in turn expands the space of solvable downstream tasks.

C. Generalization and Multimodal Learning (Theorems 6 & 7)

Result: The paper presents a Generalization Theorem for Structural Learning. If a "feature-preserving functor" exists between two categories (e.g., text and image), the structural information of the source category is preserved in the feature space of the target category.
Application to "Unseen" Objects: This explains how models like CLIP or DALL-E 2 can generate or recognize objects not present in the training data (e.g., an "avocado chair"). The model learns the structural relationships (morphisms) between text and image categories. If the text "avocado chair" constitutes a valid composition in the text category, the feature-preserving functor maps this structure to a corresponding (potentially unseen) object in the image category.
Composition Theorem (Theorem 7): This framework extends to chains of categories and explains how multimodal models can connect disparate domains (e.g., Language A $\to$ Language B $\to$ Image) by composing functors.

4. Significance and Impact

Shift from Data/Architecture to Structure: The paper shifts the theoretical focus away from data volume or network depth toward the algebraic structure of tasks. It postulates that the "power" of a Foundation Model is determined by the category it induces, not merely by its parameter count.
Theoretical Justification for Pretext Task Design: It provides a rigorous mathematical framework for Self-Supervised Learning (SSL). It explains why certain pretext tasks work better than others: they induce categories with richer morphisms, enabling a broader range of downstream tasks.
Explanation of "Zero-Shot" Capabilities: The Yoneda-based framework offers a formal explanation for how Foundation Models generalize to unseen concepts (compositional generalization) by treating them as structural compositions within a category, rather than mere statistical correlations.
Distinction between Prompting and Fine-Tuning: It clarifies the theoretical boundary between Prompt Engineering (limited by representability) and Fine-Tuning (limited only by resources and the richness of the pretrained category).

Conclusion

The paper argues that with infinite resources, a Foundation Model's ability to solve tasks is fundamentally constrained by the categorical structure learned during pretraining. While Prompt Tuning is restricted to tasks that are structurally representable within this category, Fine Tuning can unlock any task within this structural space. Furthermore, the framework demonstrates that multimodal learning is essentially the alignment of functors between different categories, enabling the generation and recognition of new, unseen objects through structural composition.