💻 computer science

Friends and Grandmothers in Silico: Localizing Entity Cells in Language Models

This paper identifies and validates sparse, causally actionable MLP neurons in early layers of language models that selectively encode specific entities, demonstrating that activating these localized cells can reliably retrieve entity-specific factual knowledge and support robust canonicalization across various linguistic forms.

Original authors: Itay Yona, Dan Barzilay, Michael Karasik, Mor Geva

Published 2026-04-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Itay Yona, Dan Barzilay, Michael Karasik, Mor Geva

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 massive library where a librarian (the AI) knows the answer to almost every question you ask. But here's the mystery: Where exactly does the librarian keep the answer? Is it written on a specific page, stored in a specific drawer, or is it a vague feeling that builds up as the librarian reads your question?

This paper, titled "Friends and Grandmothers in Silico," goes inside the brain of an AI to find the specific "switch" that turns on knowledge about a specific person or thing.

Here is the breakdown using simple analogies:

1. The "Grandmother Cell" Idea

In neuroscience, there's a famous (and slightly controversial) idea called the "Grandmother Cell." The theory suggests that somewhere in your brain, there is a single neuron that fires only when you see your grandmother. It doesn't fire for your grandfather, your aunt, or a picture of a cat. It's just for Grandma.

The researchers asked: Do AI models have "Grandmother Cells" for facts?

Is there a specific tiny switch inside the AI that lights up only when the AI thinks about "Barack Obama"?
Or is the concept of "Obama" spread out across thousands of neurons like a fog?

2. The Detective Work: Finding the Switch

The researchers acted like detectives. They took an AI model (specifically one called Qwen2.5) and asked it thousands of questions about 200 different famous people and places (like "Who is the spouse of X?" or "Where is X located?").

They watched the AI's internal "brain waves" (activations) to see which tiny switches (neurons) fired up every single time the AI thought about a specific entity, regardless of how the question was phrased.

The Discovery:
They found that for many popular entities, there is a specific "switch."

Location: These switches are mostly found in the early layers of the AI's brain (the beginning of the processing chain). It's like the AI recognizes "Who is this?" almost immediately, before it starts doing the heavy lifting of answering the question.
Stability: The same switch fired for "Barack Obama," "Obama," "Brock Obma" (a typo), and even when the name was written in different languages. This suggests the switch isn't just recognizing the letters of the name, but the identity of the person.

3. The "Amnesia" Test (Turning the Switch Off)

To prove these switches were actually important, the researchers tried to break them. They used a technique called "negative ablation," which is like flipping a switch to the "off" position or even reversing its polarity.

The Result: When they turned off the "Obama switch," the AI suddenly forgot everything about Obama. It couldn't name his wife or his presidency.
The Control: However, the AI could still talk perfectly fine about other people (like Donald Trump). It didn't go crazy; it just developed a specific case of amnesia for that one person.

This proved that the switch wasn't just a bystander; it was the key to accessing that specific knowledge.

4. The "Magic Wand" Test (Turning the Switch On)

Next, they tried the opposite. They took a question where the AI didn't know the answer (because they replaced the name with a placeholder like "X"). Then, they manually forced the "Obama switch" to turn on.

The Result: Suddenly, the AI remembered! Even though the name was missing, the AI started spitting out facts about Obama.
The Takeaway: You don't need to reprogram the whole AI. Just flipping one single switch was often enough to make the AI recall the correct facts.

5. Why This Matters (The "Canonical" Identity)

The most fascinating part is that these switches work even if you misspell the name or use an acronym.

If you type "FBI," the AI uses the same switch as if you typed "Federal Bureau of Investigation."
If you type "Paris" in French, Chinese, or Hebrew, it uses the same switch.

This suggests the AI isn't just memorizing strings of letters. It has built a canonical identity card for each entity. The "Grandmother Cell" is the ID card reader. Once the ID is scanned, the rest of the AI's brain knows exactly who to talk about.

The Catch

This isn't true for every entity in the world.

It works best for popular things (like famous presidents or big cities).
It works best in certain AI models (like the Qwen family). Other models have these switches too, but they are messier and harder to find.
It's not a perfect 100% rule; sometimes the "fog" (distributed knowledge) is still involved.

The Big Picture Analogy

Imagine the AI's brain is a giant, dark warehouse.

Old Theory: To find a box labeled "Obama," you have to search the whole warehouse, gathering clues from every corner until you find the box.
This Paper's Finding: There is actually a light switch on the wall near the entrance. If you flip that specific switch, the lights turn on, and the "Obama" box instantly becomes visible and accessible. If you break that switch, the box disappears into the dark, even though the warehouse is still full of other boxes.

In short: The researchers found that for many facts, AI models rely on sparse, specific "keys" located early in their processing. These keys act like a master ID card, allowing the AI to instantly retrieve a consistent identity for a person or place, regardless of how you ask about them.

1. Problem Statement

The paper addresses a fundamental question in mechanistic interpretability: How do Large Language Models (LLMs) store and retrieve factual knowledge about specific entities (e.g., "Barack Obama" or "Paris")?

The Core Question: Does the model construct entity meaning gradually across many layers, or does it retrieve a compact, stable representation via specific, localized internal mechanisms?
The Hypothesis: Drawing an analogy to the "grandmother cell" hypothesis in neuroscience (the idea that single neurons can represent complex concepts), the authors propose the existence of "entity cells": sparse, MLP (Multi-Layer Perceptron) neurons that act as stable access points for specific entities.
The Gap: While prior work has localized fact-specific neurons (e.g., "Paris is the capital of France"), it remains unclear if there are reusable, entity-centric neurons that serve as anchors for retrieving any attribute about a subject, regardless of the specific relation queried.

2. Methodology

The authors developed a localization-and-intervention pipeline to identify and validate these entity cells across multiple models.

A. Localization (Finding the Candidate)

Data: They used a curated subset of PopQA-200 (200 popular entities including people, locations, and organizations).
Prompting: For each entity, they generated templated prompts varying the attribute (e.g., "The origin of [Entity]," "The role of [Entity]").
Activation Extraction: They extracted the down-projection activations of MLP neurons at the final token of the entity span.
Stability Scoring: They defined a stability score ( $S_{\ell j}$ $S_{ℓ j}$ ) to rank neurons. The score favors neurons that:
- Activate strongly (high mean activation).
- Activate consistently across different prompts for the same entity (low coefficient of variation).
- Formula: $S_{\ell j} = \frac{(E[z])^2}{\text{Std}[z] + \epsilon}$ , where $z$ is the standardized activation.
Selection: The top-ranked neuron for each entity is designated as the candidate entity cell.

B. Causal Validation (Testing the Mechanism)

To prove these neurons are causally responsible for factual recall (rather than just correlational), they performed two interventions:

Negative Ablation (Necessity): They scaled the activation of the candidate neuron by a negative factor ( $\alpha < 0$ $α < 0$ ).
- Goal: To induce "entity-specific amnesia." If the neuron is critical, the model should fail to recall facts about that specific entity while maintaining fluency and knowledge of other entities.
Controlled Injection (Sufficiency): They replaced the entity mention with a placeholder token (e.g., "X") and injected the candidate neuron's activation vector at that position.
- Goal: To see if activating a single neuron is sufficient to steer the model toward the correct entity-consistent answer, even without the entity name in the prompt.
- Controls: Compared against "mean-entity" initialization (average vector) and "wrong-cell" injection (neuron from a different entity).

C. Robustness Checks

They tested if the localized cells remained stable across surface-form variations:

Aliases/Spelling: (e.g., "Barack Obama" vs. "Obaama").
Acronyms: (e.g., "FBI" vs. "Federal Bureau of Investigation").
Multilingual: (e.g., "Paris" in English, Hebrew, Chinese, Arabic).

3. Key Results

The study was conducted on 7 models from 5 families (Qwen, OLMo, Llama, Mistral, OpenLLaMA), with the strongest results observed in Qwen2.5-7B base.

Finding 1: Early-Layer Concentration

Observation: Localized entity cells are not uniformly distributed. 99% of top-ranked cells for the 200 entities are concentrated in the early layers (0–5).
Implication: This supports the theory that entity identity is canonicalized early in the forward pass, before deep reasoning layers.

Finding 2: Causal Necessity (Ablation)

Result: In Qwen2.5-7B, 131 out of 200 localized cells were "trustworthy."
Effect: Negatively ablating these cells caused a drastic drop in recall for the target entity (e.g., dropping from 1.0 to 0.123 relative score) while leaving control entities (e.g., "Trump" when ablating "Obama") near baseline.
Conclusion: These neurons are functionally necessary for entity-specific factual retrieval.

Finding 3: Causal Sufficiency (Injection)

Result: On the subset of "trustworthy" entities where the model already knew the answer, injecting the single localized neuron at a placeholder token restored the correct answer with 63.3% Pass@5 accuracy.
Comparison: This was significantly higher than mean-entity (1.8%) or wrong-cell (1.8%) controls.
Sufficiency: A single neuron was often sufficient (41/79 entities passed with top-1 injection), suggesting compact retrieval rather than distributed gradual enrichment.

Finding 4: Surface-Form Robustness

Result: The same top-ranked neuron was recovered for spelling variants, acronyms, and multilingual forms (e.g., the same cell for "Barack Obama," "Brock Obma," and "FBI").
Implication: These cells provide access to a canonical identity representation rather than a specific token string.

Cross-Model Generalization

Qwen Family: The phenomenon is highly robust in Qwen2.5-7B and Qwen2.5-7B-Instruct (post-training preserved the map). Qwen3-8B showed similar localization patterns but weaker causal effects.
Other Models: In Llama, Mistral, and OLMo, early-layer candidates could be found, but the causal validation was weaker and less consistent. This suggests entity cells are a reproducible but model-dependent phenomenon.

4. Key Contributions

Methodological Framework: Introduced a stability-based method to localize entity-sensitive neurons using templated prompts, moving beyond relation-specific fact localization.
Empirical Evidence for "Entity Cells": Provided causal evidence that sparse, early-layer MLP neurons act as stable anchors for factual retrieval, supporting a "grandmother cell" hypothesis for entities in LLMs.
Causal Manipulation: Demonstrated that these neurons are causally actionable: suppressing them causes amnesia, and injecting them restores knowledge, often with a single neuron.
Canonicalization Insight: Showed that these cells abstract entity identity across surface forms (typos, languages, acronyms), suggesting they represent a semantic "vocabulary" of entities.

5. Significance and Limitations

Significance:

Interpretability: Offers a concrete mechanism for how LLMs access factual knowledge, shifting the view from "distributed gradual enrichment" to "sparse, localized retrieval."
Control & Editing: Identifies specific "handles" (neurons) that can be manipulated to edit or suppress specific entity knowledge without retraining the model, potentially aiding in model editing and safety (e.g., forgetting sensitive entities).
Model Architecture: Highlights that different model families (even within the same architecture type) may implement factual retrieval differently, with Qwen showing a particularly clear "entity cell" architecture.

Limitations:

Not Universal: Not every entity admits a reliable single-neuron handle; coverage is higher for popular entities.
Model Dependency: The strong causal effects were most prominent in Qwen2.5-7B; other models showed weaker signals, suggesting the mechanism is not yet a universal law of all LLMs.
Metric Constraints: Reliance on first-token probabilities may understate multi-token factual competence.
Data Scope: Focused on PopQA; generalization to other knowledge domains requires further study.

Conclusion

The paper successfully identifies sparse, causally actionable "entity cells" in LLMs, primarily concentrated in early layers. These neurons function as compact access points for entity identity, robust to surface variations. While the phenomenon is strongest in the Qwen family, the findings provide a new lens for understanding, analyzing, and potentially controlling how language models retrieve factual knowledge.