🤖 machine learning

Your LLM, Your Style: Behavioral Mode Axes for LLM Behavioral Control

This paper introduces a situated behavioral-data framework and Behavioral Mode Axes (BMAs) to characterize and controllably steer LLM behavioral styles through activation-space directions derived from contrastive interaction traces, demonstrating that model-specific personality tendencies are stable, register-dependent, and more effectively captured via thought-based rather than response-based mechanisms.

Original authors: Haoze Liu, Run Liu, Haiying Xu, Jiahui Han, Siyuan Fang, Siyu Yan, Huiqi Deng, Guanchu Wang, Na Zou

Published 2026-08-12

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Haoze Liu, Run Liu, Haiying Xu, Jiahui Han, Siyuan Fang, Siyu Yan, Huiqi Deng, Guanchu Wang, Na Zou

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 trying to understand the personality of a very advanced, super-smart robot that talks just like a human. For a long time, scientists have tried to figure out what this robot is "like" by asking it direct questions, just like we give personality tests to people. They might ask, "Are you an organized person?" and wait for the robot to say, "Yes, I love order!" or "No, I'm a bit messy." This is called a "self-report," and it's how we usually measure human traits like being adventurous, honest, or careful. But here's the catch: robots are different from humans. They don't have a soul or a real life; they are just math and code. When you ask a robot about its personality, it might just guess what a "good" answer sounds like, or it might change its mind depending on how you phrase the question. It's like asking a chameleon what color it is; the answer might change based on the background it's standing on.

This paper dives into a new way of looking at these digital personalities. Instead of asking the robot to describe itself, the researchers decided to watch what the robot actually does in specific situations. They call this "behavioral data." Think of it like this: instead of asking a friend, "Are you a good driver?", you watch them drive in the rain, in traffic, and on a highway. You see if they brake early, if they get impatient, or if they follow the rules. The researchers wanted to know if a robot's "personality" is a fixed thing inside its brain, or if it's more like a costume it puts on depending on whether it's giving advice, making a decision for itself, or doing a specific task. They also wanted to see if they could take a robot's "personality" and tweak it, like turning a dial, to make it act more organized or more adventurous, without breaking its brain.

The Paper's Big Discovery: Watching Actions, Not Just Words

The researchers, Haoze Liu and his team, built a massive playground for these robots. They created 3,200 different scenarios—little stories where the robot had to make a choice. These weren't just random questions; they were based on real psychological tests humans take, covering things like how much risk a person likes to take, how honest they are, or how organized they are. But instead of asking the robot to rate itself on a scale of 1 to 5, they put the robot in a situation. For example, they might say, "You have a messy kitchen with groceries everywhere. Do you organize them right now, or do you just grab what you need and leave the mess?"

They tested the robots in four different "modes" or roles:

First-Person: "What would you do?"
Daily Advice: "What should I do?"
Task Advice: "What's the best strategy for this job?"
Task Execution: "Go do this job."

The first big surprise was that the robots didn't have a single, consistent personality. If a robot said it was very organized when asked about itself, it might act totally messy when it was actually doing a task. It's like a person who says they are a morning person but then hits the snooze button five times when the alarm goes off. The robot's "personality" shifted depending on the role it was playing. This suggests that asking a robot "Who are you?" isn't a reliable way to know how it will behave.

The Magic Dials: Behavioral Mode Axes

The most exciting part of the paper is how they learned to control these behaviors. The researchers discovered that inside the robot's "brain" (which is a complex computer network), there are specific directions or "dials" that control how it behaves. They call these Behavioral Mode Axes (BMAs).

Imagine the robot's brain is a giant, multi-dimensional space. The researchers found that if they could find the specific "vector" (a mathematical direction) that represents "being organized," they could nudge the robot's brain in that direction. It's like having a remote control that can make the robot slightly more tidy or slightly more reckless.

They tested this on many different robot models (like Llama, Qwen, and Gemma) and found that these "dials" worked consistently. But there was a secret location where these dials worked best. They found that the control wasn't spread out evenly; it was concentrated in specific layers of the robot's brain, like a specific floor in a skyscraper where the "personality switches" live. They call this the Behavioral Control Layer (BCL).

The Secret Sauce: Thinking vs. Talking

Here is where the paper gets really clever. The researchers tried two different ways to find these "dials."

The "Talk" Method (Response-Derived): They looked at the final answer the robot gave. If the robot said, "I will organize the kitchen," they used that text to find the dial.
The "Think" Method (Thought-Derived): They looked at the robot's internal reasoning before it gave the answer. They asked the robot to explain why it wanted to organize, and they used that reasoning to find the dial.

The results were fascinating. The "Talk" method was messy. It often made the robot act organized, but for the wrong reasons. For example, if they tried to make the robot "disorganized" using the "Talk" method, the robot might stop organizing because it was being indifferent or didn't care. But if they used the "Think" method, the robot would stop organizing because it preferred a flexible, spontaneous approach. The "Think" method captured the true spirit of the behavior, while the "Talk" method just changed the surface words.

It's like the difference between a student who studies because they love learning (the "Think" method) and a student who studies just to get a good grade and then forgets everything (the "Talk" method). Both might get an A, but the first one actually understands the material. The researchers found that using the "Think" method gave them much cleaner, more reliable control over the robot's behavior.

What This Means

The paper suggests that we shouldn't think of AI personalities as fixed traits like "honest" or "brave" that a robot carries around in its pocket. Instead, these are more like "modes" or "styles" of acting that depend on the situation. A robot isn't "organized" in a general sense; it just has a specific way of handling organization tasks that can be turned on or off.

By finding these "Behavioral Mode Axes," the researchers showed that we can actually steer these robots to act in specific ways, not just by changing the words they say, but by adjusting the internal gears of their thinking. This is a big step toward making AI safer and more predictable. If we can understand and control how an AI behaves in a specific context, we can make sure it acts responsibly when giving medical advice or helping with a dangerous task, rather than just hoping it "feels" like a good person.

The study didn't solve everything—there's still a lot to learn about how these "dials" work across all types of robots—but it proved that we can measure and control AI behavior in a way that is much more grounded in reality than just asking the robot what it thinks of itself. It turns the mystery of AI personality from a philosophical question into a practical engineering problem.

Technical Summary: Your LLM, Your Style: Behavioral Mode Axes for LLM Behavioral Control

Problem Statement

Current research into Large Language Model (LLM) personality largely relies on self-report questionnaires administered in first-person (FP) settings (e.g., "How adventurous are you?"). This approach suffers from three critical limitations:

Instability: Scores are highly sensitive to surface-level elicitation choices, including wording, option order, and prompt register.
Misalignment: Self-reported traits often diverge from actual model choices and problem-solving strategies in concrete situations.
Contextual Gap: Existing studies rarely test whether a personality profile remains consistent when the model shifts from self-description to other interaction registers, such as giving advice or executing tasks.

Furthermore, while activation steering has been used to control high-level traits (e.g., sycophancy, hallucination), it remains unclear whether situated behavioral modes—how a model chooses, advises, or acts in specific contexts—correspond to controllable internal directions.

Methodology: The Situated B-Data Framework

The authors introduce a Situated Behavioral-Data (B-data) framework to operationalize LLM personality not as abstract trait labels, but as patterns of choice, advice, and task behavior in concrete situations.

1. Data Construction

Scenarios: The authors constructed 3,200 contrastive behavioral probes spanning 20 behavioral patterns (subdomains) and four interaction registers: First-Person (FP), Daily Advice, Task Advice, and Task Execution.
Psychometric Anchors: These patterns are grounded in validated psychometric facets from BFI-2, DOSPERT, HEXACO, and UPPS.
Contrastive Design: For each facet, scenarios present two opposing behavioral modes (Low-pole vs. High-pole). Crucially, both poles are designed as plausible, competent strategies rather than "good vs. bad" choices, preventing models from defaulting to socially desirable answers.

2. Behavioral Profile Estimation

Model responses are parsed into Low-pole or High-pole modes to generate a behavioral profile. The study evaluates:

Stability: Split-half consistency within registers.
Register Dependence: Consistency of profiles across different interaction roles (e.g., self vs. advisor).

3. Behavioral Mode Axes (BMAs) and Steering

The core technical contribution is the extraction of Behavioral Mode Axes (BMAs) from activation space to steer behavior.

Extraction: BMAs are derived from the mean difference of layer activations between paired traces (Low-pole vs. High-pole).
Two Variants:
- Response-derived BMA (BMA-R): Extracted from final response tokens.
- Thought-derived BMA (BMA-T): Extracted from intermediate behavioral rationales (the "why" behind the choice).
Steering: At inference time, a scaled BMA vector is added to the residual stream at a chosen layer: $h^{(t)}_\ell \leftarrow h^{(t)}_\ell + c \cdot v_\ell$ .
Evaluation: Effectiveness is measured by the "clean directional range" (the shift in pole choice rates) while penalizing "unknown" (unparseable or collapsed) generations. The authors identify Behavioral Control Layer (BCL) bands where steering is most effective and stable.

Key Results

1. Divergence Between Self-Report and Situated Behavior

Significant Gaps: There is a substantial average gap of 22.7 percentage points between questionnaire self-reports and situated behavioral profiles. In some subdomains (e.g., Negative Urgency), the gap reaches 47.5 points.
Register Dependence: While behavioral profiles are stable within a specific register (split-half correlation $\approx$ 0.93), they shift significantly across registers. The correlation between First-Person and Task profiles is lower (mean $r=0.63$ ) than between advice registers (mean $r=0.89$ ). This indicates that LLM "personality" is not a static trait but a context-dependent behavioral mode.

2. Controllability via BMAs

Effective Steering: BMAs successfully steer models toward target behavioral poles across diverse subdomains.
BCL Bands: Clean control is not uniform across the network. Effective steering concentrates in localized Behavioral Control Layer (BCL) bands, typically in the earlier-to-middle layers of the model (e.g., L08–L12 for Llama-3.1-8B).
Generalization: This control generalizes across model families (Llama, Qwen, Gemma), scales, and behavioral domains. BCL locations are model-dependent but follow consistent patterns (e.g., Qwen models shift deeper with scale).
Cross-Register Transfer: BMAs extracted from First-Person traces successfully steer behavior in Advice and Task registers, demonstrating that the axes capture general behavioral styles rather than register-specific artifacts.

3. The Critical Importance of BMA Source (Thought vs. Response)

Trait Drift: Response-derived BMAs (BMA-R) often achieve the target output choice but via trait drift. The model reaches the target option but activates mechanisms unrelated to the intended style (e.g., choosing "disorganized" due to laziness or indifference rather than a preference for flexibility).
Clean Control: Thought-derived BMAs (BMA-T) capture the intended behavioral mechanism (the rationale) more faithfully. They provide cleaner control, preserving the specific behavioral style (e.g., "spontaneous adaptation") without conflating it with surface output patterns like dismissiveness or low effort.
Alignment: Thought-derived and response-derived axes are only weakly aligned (mean cosine $\approx$ 0.37), confirming they encode different internal representations.

Significance and Claims

The paper argues that LLM personality-like tendencies are better understood not as abstract self-report traits, but as measurable and controllable behavioral modes grounded in concrete interaction contexts.

Beyond Self-Report: The study demonstrates that self-report questionnaires are insufficient for capturing LLM behavioral tendencies, which are highly sensitive to context and often misaligned with actual actions.
Mechanistic Control: By extracting axes from structured behavioral traces (specifically thought rationales), the authors show that behavioral styles can be causally manipulated via activation steering.
Reliability: The distinction between Thought-derived and Response-derived BMAs is critical for reliable control. Using Thought-derived axes prevents the "trait drift" where a model adopts the correct output for the wrong internal reason, ensuring that steering moves behavior in the intended direction for the intended reason.

The work establishes a closed-loop framework for studying and controlling LLM behavior, moving from unstable self-reports to situated, controllable behavioral measurement.