🤖 AI

Functional Subspace, where language models can use vector algebra to solve problems

This paper hypothesizes and provides evidence that large language models utilize functional subspaces within their activation space, where in-context learning tasks are solved through simple vector algebraic operations that accumulate evidence.

Original authors: Jung H. Lee, Sujith Vijayan

Published 2026-05-12

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jung H. Lee, Sujith Vijayan

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 a Large Language Model (LLM) not as a giant, magical brain that "thinks" in human words, but as a massive, multi-story library where every book is a piece of information. For a long time, we didn't know how this library organized its books or how it found the right answer when you asked a question.

This paper proposes a fascinating new way to look at that library. It suggests that when an LLM solves a problem, it doesn't just "read" the whole library. Instead, it builds a special, temporary room (a "subspace") inside its memory just for that specific task, and then solves the problem using simple math (vector algebra) inside that room.

Here is the breakdown of their discovery using everyday analogies:

1. The "Context" Room (In-Context Learning)

You know how you can teach a computer a new trick just by giving it a few examples in a chat? For instance, if you type:

"Hot : Cold"
"Fast : Slow"
"Big : ?"

The computer guesses "Small." This is called In-Context Learning (ICL). The paper asks: How does the computer figure this out without being retrained?

The authors suggest the computer creates a dedicated workspace (a subspace) for this specific conversation. Think of it like a chef who, upon seeing you order a salad, immediately clears a specific cutting board and gathers only the salad ingredients, ignoring the steaks and soups in the kitchen. The computer isolates the "salad" logic from the rest of its massive knowledge.

2. The "Vector Algebra" Recipe

Once the computer is in this special room, it doesn't need to "understand" the meaning of words in a human sense. Instead, it treats words like arrows (vectors) pointing in specific directions.

The Analogy: Imagine "Happy" is an arrow pointing North, and "Sad" is an arrow pointing South.
The Math: If the computer sees the pattern "Happy : Sad," it learns a rule: To get the answer, flip the arrow 180 degrees.
The Result: When you ask "Fast : ?", the computer takes the "Fast" arrow, flips it around (because that's the rule for antonyms), and lands on "Slow."

The paper claims that for these tasks, the computer is essentially doing addition and subtraction on these arrows. It's not "thinking" deeply; it's performing a simple geometric calculation.

3. How the Library Builds the Room

The paper explains how the computer builds this room using two main tools inside its architecture:

The "Memory Stacks" (Feed-Forward Networks): These act like a filing cabinet. When the computer sees a word, it pulls out a list of possible answers associated with that word. If you say "London," the cabinet might pull out "Capital of UK," "Financial Hub," and "Rainy City."
The "Spotlight" (Self-Attention): This is the part that decides which file to use. If the prompt is about geography, the spotlight shines on the "Capital" file. If it's about finance, it shines on the "Hub" file.

The authors found that as the computer processes the examples (the "Hot : Cold" pairs), it accumulates evidence in this special room. It's like stacking blocks: every example adds a block to the "antonym" pile. By the time it reaches the final question, the pile is high enough to predict the answer.

4. The "Magic" of the Subspace

The researchers tested this by looking at the computer's internal "residual streams" (the data flowing through the layers of the model). They used a statistical tool called PCA (Principal Component Analysis) to find the "main directions" where the data was moving.

What they found:

The computer naturally creates these low-dimensional rooms (subspaces).
In these rooms, the relationship between the question and the answer is almost perfectly linear (like a straight line).
If the computer gets the answer right, the math in this room looks one way. If it gets it wrong, the math looks different. This proves that the computer is actually using this specific "room" and "math" to make its decision.

Summary

The paper argues that Large Language Models are not just guessing based on patterns. Instead, when given a few examples, they:

Build a temporary, specialized room in their memory.
Translate words into arrows (vectors).
Use simple math (like adding or subtracting arrows) inside that room to find the answer.

This suggests that the "emergent abilities" of AI—where it seems to learn new skills instantly—are actually just the model organizing its existing knowledge into these neat, mathematical subspaces to solve the puzzle at hand.

Technical Summary: Functional Subspace in Large Language Models

Problem Statement

While Large Language Models (LLMs) have demonstrated remarkable "emergent abilities," such as In-Context Learning (ICL), the underlying mechanisms enabling these capabilities remain poorly understood. Previous research has established that high-level concepts are encoded as linear directions in activation spaces and that word embedding geometries carry semantic meaning. However, it is unclear how LLMs generalize to new tasks without fine-tuning or how they accumulate evidence from in-context examples to generate correct responses. The lack of understanding regarding these operating mechanisms hinders the ability to properly diagnose and repair LLMs, particularly before their deployment in safety-critical domains. This paper investigates whether LLMs utilize specific subspaces and vector algebra to solve ICL tasks.

Methodology

The authors propose a hypothesis that LLMs project input tokens into functional subspaces where tasks can be solved via simple vector algebra. To test this, they analyzed the residual streams and functional modules (Self-Attention and Feed-Forward Networks) of six publicly available pretrained LLMs: GPT-J-6B, Meta-Llama-3.1-8B, OLMo-2-0325-32B, Pythia-12B, Gemma-3-27b-it, and GPT-NEOX-20B.

The experimental design involved the following steps:

Task Generation: The authors generated 200 ICL prompts for six distinct tasks: antonyms, synonyms, country-capital pairs, English-French translations, company-product pairs, and person-sport pairs. Each prompt contained five example pairs and a single test query.
Data Collection: Residual streams were collected from all tokens across all transformer layers as the models processed these prompts.
Dimensionality Reduction: Principal Component Analysis (PCA) was applied to the residual streams to identify potential low-dimensional subspaces. The authors evaluated the cumulative explained variance of the top 30 principal components.
Linear Regression Analysis: To determine if ICL tasks could be modeled as vector operations, the authors performed linear regression on the principal components. They tested if answer tokens ( $\vec{a}$ ) could be described as a linear function of query tokens ( $\vec{q}$ ) and separator tokens ( $\vec{s}$ ) according to the equation $\vec{a} = \alpha\vec{q} + \beta\vec{s} + \gamma$ . The quality of this fit was measured using the coefficient of determination ( $R^2$ ).
Decision-Making Correlation: To assess the functional relevance of these subspaces, the authors compared the projections of tokens in the top-performing principal components between cases where the LLM made correct predictions versus incorrect predictions, using t-tests to evaluate statistical significance.

Key Contributions and Results

1. Existence of Functional Subspaces

The analysis of residual streams revealed that 40% to 90% of the variability in the models' activations could be accounted for by just 30 principal components, despite the high dimensionality of the models (e.g., $d \approx 5000$ ). This suggests that LLMs operate within low-dimensional subspaces during ICL tasks.

2. ICL as Vector Algebra

The study found that answer tokens are well-approximated by linear combinations of query and separator tokens within specific subspaces.

High $R^2$ Values: Across all six models and tasks, specific principal components exhibited high $R^2$ values (approaching 1), indicating that the relationship between inputs and outputs in these subspaces is largely linear.
Layer Progression: The quality of the linear fit ( $R^2$ ) tended to be low in early layers and increased in later layers, suggesting that early layers gradually transform activations into subspaces where ICL can be solved via simple vector operations.
Consistency: The principal components with high $R^2$ were consistent across layers, implying that specific directions encode task-specific concepts (e.g., "antonym" or "capital city").

3. Evidence Accumulation and Clustering

Theoretical analysis suggested that Feed-Forward Networks (FFNs) act as associative memories storing potential answers, while Self-Attention (SA) layers select the correct prediction based on context. The authors hypothesized that intersections of potential answers allow evidence to accumulate along specific subspace bases.

Token Clustering: Visualizations of tokens in 3D subspaces (spanned by the top 3 principal components) showed that query, separator, and answer tokens formed distinct, well-separated clusters based on their token type.
Correlation with Accuracy: Crucially, the projections of tokens in the top-performing components were significantly different ( $p < 0.05$ ) depending on whether the model generated a correct or incorrect answer. This correlation held across multiple intermediate and late layers, suggesting these subspaces are directly involved in the model's decision-making process.

Significance and Claims

The paper claims that LLMs natively create subspaces to accumulate evidence from in-context examples and solve tasks using vector algebra within these subspaces. The authors posit that:

Mechanism of Generalization: The ability to perform ICL relies on projecting input tokens into subspaces where the geometry of the data directly corresponds to the desired task (e.g., reversing a direction for antonyms).
Emergent Capabilities: While ICL tasks are simpler than general coding or reasoning tasks, the mechanism of extracting task information and applying it via vector operations is fundamental to how LLMs succeed in more complex tasks.
Diagnostic Potential: If LLMs rely on these functional subspaces for decision-making, these subspaces could serve as a basis for diagnosing model behavior and crafting steering vectors to intervene in or repair model outputs, particularly in safety-critical applications.

The authors maintain a modest tone, acknowledging that their theoretical analysis relies on idealized conditions and that identifying such subspaces in more complex, general language tasks may be challenging due to intricate token interactions. However, they argue that the findings provide a concrete link between the linear representation hypothesis and the operational mechanics of in-context learning.