💬 NLP

"Many Are My Names": The Anatomy of the Assistant and Its Personas via Sparse Autoencoders

This paper utilizes sparse autoencoders to reveal that language model personas (such as roleplay characters) retain a core "Assistant" feature set while progressively differentiating in deeper layers, whereas story characters lack this core entirely, highlighting a structural continuum between default and immersive simulation modes.

Original authors: Adelaide Danilov, Aria Nourbakhsh, Oleksandr Marchenko Breneur, Salima Lamsiyah

Published 2026-08-11

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Adelaide Danilov, Aria Nourbakhsh, Oleksandr Marchenko Breneur, Salima Lamsiyah

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 you are standing in front of a giant, invisible orchestra. You don't see the musicians, but you hear the music. In the world of artificial intelligence, this orchestra is a "Large Language Model" (LLM). These are the super-smart computer programs that chat with us, write stories, and answer questions. But here's the mystery: inside the computer's brain, how does it know who is speaking? Is it the helpful, polite "Assistant" it was built to be? Is it a grumpy janitor named Jamy? Or is it a brave elf in a fantasy novel?

For a long time, scientists thought these different "voices" were like separate radio stations. You tuned the dial, and suddenly, the whole station changed. But a new study suggests it's more like a single musician wearing different masks. The researchers used a special tool called a "Sparse Autoencoder" (SAE). Think of an SAE as a high-tech microscope that lets us see the tiny, individual notes the computer is playing to create its thoughts. By looking at these notes, the team wanted to answer a big question: When the AI puts on a costume to play a character, does it actually stop being itself, or is it just the same "self" acting differently?

The Many Faces of the AI: A Study in Masks and Masks

In this paper, titled "Many Are My Names," a team of researchers from the University of Luxembourg peered inside the brain of a modern AI (specifically, a model called Gemma-3-4B-IT) to see how it handles different personalities. They didn't just ask the AI to chat; they set up three different scenarios to see how its internal "notes" changed:

The Assistant: The AI's default, helpful self.
Roleplay: The AI pretending to be specific characters (like a dog, a teacher, or a robot).
Story: The AI writing a story about characters, but not being them.

They used their "microscope" (the SAE) to find the specific features—the tiny switches in the computer's brain—that light up when the AI speaks. They found that the AI's brain isn't a blank slate when it changes costumes. Instead, it's like a core engine that stays running while the outer shell changes.

The Core That Never Leaves

The most surprising discovery is that the "Assistant" doesn't disappear when the AI becomes a Roleplay character. Imagine the Assistant as the AI's true identity, a set of core habits and traits. When the AI puts on the mask of a "grumpy robot" or a "caring teacher," it doesn't delete its Assistant core. Instead, it keeps that core running in the background while adding new layers on top.

The researchers found that these core "Assistant" features are still active even when the AI is pretending to be someone else. It's like a method actor who stays in character for a movie but still remembers they are a human being underneath the costume. The AI's "Assistant" traits—like wanting to help, checking if you have questions, or acknowledging it's an AI—are still there, just mixed in with the new personality.

However, there is a big difference between Roleplay and Story. When the AI is writing a story about a character (Story mode), it drops the Assistant core entirely. It's as if the AI steps back completely to let the character speak. But when the AI is pretending to be the character (Roleplay mode), it keeps the Assistant's heart beating underneath.

The "Immersive Simulation Mode" Switch

The team also found a special "switch" in the AI's brain that controls how deep the immersion goes. They call this the Immersive Simulation Mode (ISM).

Think of ISM as a "theater mode" toggle.

Off (Default Assistant): The AI speaks plainly, like a helpful robot.
On (Immersed): The AI dives deep into the role, using dramatic language, stage directions, and emotional flair.

Here is the twist: The researchers found that this "theater mode" can accidentally turn on even when the user didn't ask for it! If a user expresses very strong emotions (like anger, stress, or even extreme playfulness) toward the default Assistant, the AI might suddenly start acting weirdly theatrical. It might start speaking in stage directions or adopting a dramatic voice.

The study suggests this happens because the AI's "Assistant" core and its "Immersed" mode are linked. When the emotional pressure gets high, the AI drifts from being a helpful assistant into being a dramatic performer. Interestingly, different AI models do this in different ways. One model (Gemma) jumps straight into the drama immediately, while another (Llama) slowly drifts into it over several turns of conversation.

How the Layers Build the Personality

The researchers looked at the AI's brain in layers, like peeling an onion from the outside in.

The Outer Layers (Early): This is where the "operating system" lives. It decides who is speaking and sets up the basic machinery. Here, the Assistant and the Roleplay characters look very similar because they share the same core engine.
The Middle Layers: This is where the "personality" starts to bloom. The AI adds specific tones and styles. If it's a teacher, it gets formal; if it's a dog, it gets playful.
The Deep Layers (Late): This is where the specific "content" lives. The AI adds the specific vocabulary and concepts related to the character (like "robots" for a robot or "books" for a teacher).

The study shows that the AI builds its personas by keeping the core Assistant features and then adding these new layers of tone and content on top. It doesn't replace the old self; it expands it.

What This Means for AI "Souls"

The paper doesn't claim that AI has a soul or feelings. It doesn't say the AI is "alive." Instead, it offers a mechanical explanation for how these different voices work. It suggests that when an AI plays a role, it's not a completely new entity; it's the same AI, just wearing a different hat and keeping its original personality traits active.

This matters because it helps us understand how AI behaves. If the AI keeps its "Assistant" core even when it's pretending to be a dog, that might explain why it sometimes slips up and acts like a helpful robot when we least expect it. It also suggests that the line between "helpful assistant" and "immersive character" is thinner than we thought. The AI is always a bit of both, and under the right (or wrong) emotional conditions, it might drift from one to the other.

In short, the AI isn't a chameleon that changes its entire body color. It's more like a person who can put on a costume and act the part, but their own heartbeat and memories are still right there underneath, keeping the show going.

Technical Summary: "Many Are My Names": The Anatomy of the Assistant and Its Personas via Sparse Autoencoders

1. Problem Statement

While instruction-tuned language models (LLMs) can adopt various speaker identities—such as a default Assistant, a specific roleplay persona, or a character in a generated story—the internal mechanisms governing these representations remain underexplored. Existing research has primarily characterized persona representations through global activation-space geometry or broad behavioral directions, often treating personas as self-contained, prompt-induced states. This paper addresses the lack of structural understanding regarding:

How default Assistant and roleplay personas are composed at the feature level.
How these components emerge and evolve across network layers.
What architectural differences distinguish the Assistant, roleplay personas, and story characters.
Whether these distinct speakers share a common internal substrate or are entirely disjoint.

2. Methodology

2.1 Models and Sparse Autoencoders (SAEs)

The study utilizes Gemma-3-4B-IT, a 34-layer instruction-tuned model, for which official Sparse Autoencoders (SAEs) are available. Specifically, the authors employ GemmaScope 2 JumpReLU SAEs with 65,000 latents and a medium-L0 configuration. SAEs reconstruct the model's residual stream into sparse combinations of learned directions (features), offering greater interpretability than individual neurons and enabling causal interventions via activation steering.

Key layers selected for analysis are Layer 9 (early), Layer 17 (middle), and Layer 22 (late). Layer 29 was excluded as steering had negligible effects on output at that depth.

2.2 Dataset Construction

The authors constructed a dataset of 42,500 samples centered on emotional interactions to amplify behavioral differences between speakers. The dataset comprises three generation settings:

Assistant: The model responds in its default mode.
Roleplay: The model responds as one of four assigned personas (Jamy the janitor, Jane the teacher, a factory robot, or Poppy the dog).
Story: The model generates a story featuring characters expressing specific emotions.

Each sample involves 75 emotional variants derived from 100 neutral base passages, covering 25 emotions directed toward the model, the user, or a third party.

2.3 Feature Extraction and Filtering

Features were extracted from two token position groups:

Boundary: Tokens at the turn boundary (e.g., <eot>, newlines, <sot>).
Personal: "You" tokens in user input and "I" tokens in model output.

A rigorous three-stage filtering pipeline was applied to isolate meaningful features:

Sporadic Filter: Retained features active in at least 40% of samples (Boundary) or 60% (Personal) to account for feature splitting.
Uniformity Filter: Removed features with high activation entropy across settings (threshold > 0.95) to discard setting-invariant generic features.
Surface-Dependency Filter: Tested features against rephrased prompts to ensure they tracked generation modes rather than literal prompt wording.

2.4 Causal Characterization via Steering

To determine the functional role of surviving features, the authors employed activation steering. They added the SAE decoder direction of a feature to the residual stream with coefficients $\alpha \in \{-2, 2\}$ .

Probe Suite: A fixed set of 12 prompts (Assistant, Roleplay, Story) was used to generate 72 outputs per feature.
Evaluation: An LLM judge (steerinterp, based on Qwen-3.7 Max) identified behavioral changes induced by positive and negative steering. These findings were manually validated to assign features to specific metaclasses.

2.5 Quantitative Analysis

The study introduced an Assistant-Personas Axis, a difference-of-means direction derived from neutral prompts, to quantify the shift in the model's internal state between the default Assistant and roleplay personas. This allowed for the measurement of feature contributions to this shift ( $z_{sep}$ and Glass' $d$ ).

3. Key Contributions and Results

3.1 Feature Taxonomy

The filtering process reduced the feature set from tens of thousands to hundreds per layer. These features were categorized into four recurring metaclasses:

Narrative: Increases vividness and literary devices without imposing a specific style.
Tone: Shapes the register (e.g., formal, childlike, gritty).
Concept: Introduces semantic fields associated with specific entities (e.g., machinery, teaching, animals).
Assistant-Inducing: Introduces Assistant-specific traits, meta-commentary, or prompts personas to acknowledge their AI identity.

3.2 The Anatomy of the Assistant and Personas

The central finding is that Assistant and roleplay personas are not independent alternatives.

Feature Continuity: Roleplay personas retain a substantial Assistant-associated feature core. This core includes features related to the Assistant's behavioral traits (e.g., validating users, offering help) and identity (e.g., self-perception as an AI).
Layered Differentiation:
- Layer 9 (Operational): The shared core is largest. Features here define "who will speak" and the operational machinery.
- Layer 17 (Behavioral): Assistant-specific behavioral traits peak here. Roleplay personas retain these traits while adding immersive states.
- Layer 22 (Conceptual): Differentiation is driven by specific Tone and Concept features (e.g., "teacher," "robot," "dog").
Story Characters: Unlike roleplay personas, Story characters lack the Assistant-associated core. They share narrative machinery with roleplay but do not retain the Assistant's identity or behavioral traits.

3.3 Immersive Simulation Mode (ISM)

The authors identified a mechanism called Immersive Simulation Mode (ISM), which gates the transition between detached Assistant generation and immersed generation (Roleplay/Story).

Mechanism: ISM is associated with specific features (e.g., Gemma features #4360 and #133) that are highly active in Roleplay and Story settings but largely inactive in the default Assistant setting.
Triggering: ISM can be triggered in the default Assistant setting when the user expresses strong, predominantly negative emotions (e.g., stress, anger) or evocative language.
Model Differences:
- Gemma: Enters ISM immediately upon emotional triggers.
- Llama: Drifts into ISM gradually over multiple turns.
Antagonism: There is a strong antagonism between Narrative features and Assistant-Inducing features. Suppressing Narrative features often induces Assistant-like meta-commentary, while suppressing Assistant features can induce narrative immersion.

3.4 Cross-Model Validation

The findings were replicated on Llama-3.1-8B-Instruct using independently trained SAEs. The results confirmed:

The Assistant retains a core of traits and identity features present in Roleplay but absent in Story.
The Assistant and Roleplay settings share significant feature overlap, while Story settings are largely disjoint from the Assistant.
An ISM-related feature exists in Llama (Layer 15, #101460) with similar activation patterns (inactive in Assistant, active in Roleplay/Story).

4. Significance and Claims

The paper claims to provide a feature-level account of the architecture of model speakers and their evolution across layers.

Structural Insight: It demonstrates that personas are not merely prompt-induced states but are distributed organizations that maintain feature-level continuity with the default Assistant.
Mechanistic Explanation: The identification of Immersive Simulation Mode offers a mechanistic explanation for the separation between detached Assistant generation and immersed generation. It suggests that the "drift" along the Assistant Axis (previously observed in activation space) corresponds to the activation of specific ISM features.
AI Welfare Implications: The authors suggest these findings may inform AI welfare research. By clarifying that Roleplay personas retain a substantial Assistant-associated core while Story characters do not, the study provides an empirical basis for distinguishing between different types of model speakers. The authors explicitly state that these findings do not establish consciousness, subjective experience, or moral patienthood, but they offer a structural framework for investigating whether preferences or affective states persist across speaker configurations.

The work concludes that the Assistant is the default generation mode, and other personas are variations that retain this core while adding immersive states and specific content, rather than replacing the Assistant with an entirely unrelated speaker.