🤖 AI

Emergent Analogical Reasoning in Transformers

This paper formalizes analogical reasoning in Transformers through the lens of category theory functors, demonstrating via synthetic tasks and mechanistic analysis that it emerges from the geometric alignment of relational structures and functor application, with findings that hold true for pretrained large language models.

Original authors: Gouki Minegishi, Jingyuan Feng, Hiroki Furuta, Takeshi Kojima, Yusuke Iwasawa, Yutaka Matsuo

Published 2026-05-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Gouki Minegishi, Jingyuan Feng, Hiroki Furuta, Takeshi Kojima, Yusuke Iwasawa, Yutaka Matsuo

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: How AI Learns to "Get the Joke"

Imagine you are teaching a robot to understand the world. You show it a picture of a Sun and a Planet orbiting it. Then, you show it a picture of a Proton and an Electron orbiting it.

A human immediately sees the connection: "Oh! The Sun is like the Proton, and the Planet is like the Electron." We didn't learn this because the Sun looks like a Proton (they are very different). We learned it because they play the same role in their respective systems. This is called analogy.

This paper asks: How do AI models (Transformers) learn to make these "leaps" of logic? Do they just memorize facts, or do they actually understand the underlying structure?

The Experiment: A Synthetic Playground

To find out, the researchers built a "toy world" for the AI.

The Setup: They created two separate groups of characters (let's call them Group A and Group B).
The Rules: In Group A, every character has specific relationships with others (e.g., "Alice is the boss of Bob"). In Group B, the characters are different (e.g., "X is the boss of Y"), but the pattern of relationships is identical to Group A.
The Test: The AI is trained on Group A. Then, it is asked: "If Alice is the boss of Bob, and X is the boss of Y, who is the boss of Z?" The AI has to figure out that "X" corresponds to "Alice" and "Y" corresponds to "Bob" just by looking at the structure of the relationships, not the names.

What They Discovered: The Three-Stage Growth Spurt

The researchers watched the AI learn over time and found it happens in three distinct stages, like a child growing up:

Memorization (The Rote Learner): First, the AI just memorizes the specific facts it sees. If it sees "Alice bosses Bob," it remembers that exact pair.
Composition (The Puzzle Solver): Next, it learns to chain facts together. If "Alice bosses Bob" and "Bob bosses Carol," it can figure out "Alice bosses Carol."
Analogy (The Insightful Thinker): Finally, and most importantly, it learns to map the structure from Group A to Group B. It realizes, "I don't need to know who X is; I just need to know X plays the same role as Alice."

Key Finding: This final stage of "insight" is very fragile. It doesn't happen automatically just by making the AI bigger. It requires the right amount of data, the right training speed, and specific settings. If the training is too messy or the data is too sparse, the AI never makes the leap.

The Secret Sauce: How the AI Actually Does It

The researchers didn't just watch the AI solve the problem; they looked inside its "brain" (its internal math) to see how it solved it. They found two magical steps happening inside the model:

1. The "Geometric Dance" (Structural Alignment)

Imagine the AI has a giant 3D map where every character is a dot.

Before learning: The dots for Group A and Group B are scattered randomly. They don't look related.
After learning: The AI rearranges the dots. Suddenly, the dot for "Alice" and the dot for "X" move closer together in 3D space. The dot for "Bob" and "Y" move closer too.
The Metaphor: It's like two separate dance floors. At first, the dancers are moving randomly. Then, the music changes, and the dancers from both floors start mirroring each other's moves perfectly. The AI has aligned the geometry of the two groups.

2. The "Magic Vector" (The Functor)

Once the dots are aligned, the AI uses a simple math trick to solve the puzzle.

It treats the relationship between the two groups as a vector (a direction and distance).
It essentially does this math: Target Character = Source Character + Magic Direction.
The Metaphor: Imagine you have a map of London. You want to know where Paris is relative to London. You don't need to memorize every street in Paris. You just need a "translation arrow" that says "Move 200 miles East." The AI finds this "translation arrow" (which they call a functor) and adds it to the source character to find the answer.

Does This Happen in Real AI?

The researchers tested this on massive, pre-trained AI models (like Gemma and Llama) that were never trained on their specific toy game.

The Result: Yes! Even though these big models weren't explicitly taught the toy game, when you ask them to solve an analogy, they go through the same process.
The Layer-by-Layer Journey: In the toy model, this happened over time (as training steps increased). In the big models, it happens over depth (as the data moves through the layers of the network). The early layers are messy, but by the time the data reaches the final layers, the "geometric dance" has aligned, and the "magic arrow" is applied to give the correct answer.

Summary

This paper shows that analogical reasoning isn't just a vague "feeling" AI has. It is a concrete, mechanical process where the AI:

Aligns the shapes of two different worlds in its internal map.
Finds a simple "translation arrow" to move between them.
Uses that arrow to jump from one world to another.

This proves that modern AI can do more than just memorize; it can learn to see the hidden structures that connect different ideas, much like a human does.

Technical Summary: Emergent Analogical Reasoning in Transformers

Problem Statement

While large language models (LLMs) have demonstrated remarkable progress in compositional reasoning—where complex conclusions are derived by chaining simpler, local steps—the mechanisms underlying analogical reasoning remain poorly understood. Analogical reasoning differs qualitatively from composition; it involves identifying shared relational structures across distinct domains to enable "leaps" in inference (e.g., mapping the solar system to atomic structure) rather than chaining local relations. Although analogy is considered a central faculty of human intelligence and a source of creativity, it is unclear when and how Transformer architectures acquire this capability, or whether it relies on the same mechanisms as compositional reasoning.

Methodology

The authors propose a controlled framework to study the emergence of analogical reasoning, inspired by the concept of functors in category theory.

1. Formalization and Synthetic Task

The authors formalize analogy as the inference of correspondences between entities across categories. They design a synthetic task involving:

Entities and Relations: Two disjoint sets of entities ( $E_1, E_2$ ) representing two categories. A directed complete graph of relations is constructed on $E_1$ .
Functor Mapping: A bijection $F: E_1 \to E_2$ defines a one-to-one correspondence, transferring the relational structure from $E_1$ to $E_2$ such that $r(F(e_s), F(e_t)) = r(e_s, e_t)$ .
Fact Types:
- Atomic Facts: Single relations $(e_s, r, e_t)$ .
- Compositional Facts: Two-hop relations derived from atomic facts.
- Analogical Facts: Cross-category mappings $(e_s, f, F(e_s))$ , where $f$ is a special functor token.
Evaluation: The model is trained on in-distribution (ID) facts and evaluated on out-of-distribution (OOD) compositional and analogical facts. Success requires generalizing beyond memorization to capture the underlying relational structure.

2. Experimental Setup

Models: Lightweight causal Transformers (1 layer, 1 head) for mechanistic analysis, and pretrained LLMs (Gemma-2-2B/9B, LLaMA) for external validity.
Training: Trained with cross-entropy loss on the final token. Hyperparameters (weight decay, batch size, learning rate) and data characteristics (number of entities, relations, OOD ratios) are systematically varied.
Mechanistic Analysis: The authors employ:
- Dirichlet Energy: To quantify the geometric alignment of entity embeddings in the representation space.
- Logit Lens & Linear Probes: To track the evolution of predictions and representational structure across Transformer layers.
- Attention Analysis: To measure information retrieval from source entities to the functor token.

Key Contributions and Results

1. Training Dynamics and Sensitivity

The study reveals a distinct three-stage learning dynamic for analogical reasoning:

Memorization: The model fits in-distribution facts.
Compositional Reasoning: The model learns to chain relations.
Analogical Reasoning: The model finally acquires the ability to map across categories.

Crucially, unlike compositional reasoning, the emergence of analogical reasoning is highly sensitive to:

Data Characteristics: It requires sufficient relational diversity (number of relations) and fails if the OOD ratio is too high. It emerges later as the number of entities increases.
Optimization: Moderate weight decay accelerates emergence, but excessive weight decay prevents it. Larger batch sizes generally help.
Model Scaling: Performance does not improve monotonically with model size. Very wide models (e.g., $d_{model}=512$ ) or deeper models often fail to acquire analogical reasoning under default settings, suggesting it relies on specific structural alignments rather than just capacity.

2. Mechanistic Decomposition

The authors decompose the implementation of analogy in Transformers into two geometric components:

Structural Alignment: Before analogy emerges, the embedding space undergoes a reorganization where entities from different categories become geometrically aligned based on their relational roles. This is quantified by a significant decrease in Dirichlet Energy.
Functor Application: Once aligned, the Transformer applies the functor via a vector arithmetic operation. The mechanism involves:
- The functor token $f$ attending to the source entity $e_s$ .
- Residual connections integrating this information such that the target entity representation approximates $e_t \approx e_s + f$ .
- This is evidenced by increasing attention scores from $f$ to $e_s$ and high parallelism between the vector difference $(e_t - e_s)$ and the functor embedding $f$ .

3. Presence in Pretrained LLMs

The study validates these findings in pretrained LLMs (Gemma-2, LLaMA) using in-context learning:

Layer-wise Emergence: Instead of occurring over training steps, the geometric alignment (decreasing Dirichlet Energy) and the subsequent rise in target probability occur progressively across the depth of the network layers.
Robustness: These signatures persist across different model families and varying numbers of entities, indicating that the geometric mechanism for analogy is a fundamental property of Transformer inference.

Significance and Claims

The paper claims to move the study of analogy from an abstract cognitive notion to a concrete, mechanistically grounded phenomenon in neural networks.

Distinct Mechanism: The work demonstrates that analogical reasoning is qualitatively distinct from compositional reasoning, relying on specific geometric alignments in the embedding space rather than just increased model capacity or simple weight-norm regularization.
Interpretability: By linking the abstract concept of a "functor" to concrete vector arithmetic ( $e_t \approx e_s + f$ ) and attention mechanisms, the paper provides a mechanistic explanation for how Transformers perform cross-domain inference.
Foundation for Future Work: The authors position this work as an initial step toward understanding how models might achieve sample-efficient learning and creativity through analogy. They acknowledge limitations, such as the use of explicit functor tokens and perfect structural isomorphism in synthetic tasks, noting that real-world analogies are often partial and lack explicit cues. However, they argue that the identified geometric signatures offer a basis for investigating more naturalistic analogical reasoning settings.