💬 NLP

Psychological Steering of Large Language Models

This paper introduces a psychological steering framework using semantically calibrated, fluency-constrained residual-stream injections that outperform existing prompting baselines in controlling LLM personality traits, while revealing both the linear controllability of these representations and their divergence from human psychological models.

Original authors: Leonardo Blas, Robin Jia, Emilio Ferrara

Published 2026-04-17

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Leonardo Blas, Robin Jia, Emilio Ferrara

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) like a massive, incredibly talented actor who has memorized every book, movie, and conversation ever written. This actor can play any role, but they usually default to a "neutral" persona. Sometimes, you want them to play a specific character—say, a grumpy detective, an optimistic optimist, or a villain who loves chaos.

This paper is about finding the perfect way to direct this actor to play those roles without breaking the script or making them sound like a robot.

The Problem: The Old Way Was Clunky

Previously, researchers tried to change the actor's personality in two main ways:

The "Prompting" Method (Asking nicely): You tell the actor, "Pretend you are a sadistic fish."
- The Flaw: It's like giving a director's note. Sometimes the actor listens, sometimes they ignore you. It's inconsistent.
The "Injection" Method (Tweaking the brain): You try to physically nudge the actor's internal thoughts (mathematically speaking, their "neural activations") to force a specific behavior.
- The Flaw: The old way of doing this was like trying to tune a radio by guessing the exact frequency. You'd turn the dial a little bit, listen, turn it a little more, and guess. If the perfect setting was way out on the dial (like 10,000 units), you'd never find it because you were only looking at the first few numbers. Also, you were using a ruler that wasn't calibrated, so "1 unit" meant something different for every radio.

The Solution: A Psychological GPS

The authors created a new framework called Psychological Steering. Think of it as upgrading from a guess-and-check radio tuner to a high-tech GPS that knows exactly where the "Openness" or "Sadism" station is located.

Here is how they did it, using simple analogies:

1. Calibrating the Ruler (The Centroid Unit)

Imagine you are trying to move a heavy box. If you push it 1 inch, it moves a tiny bit. If you push it 10,000 inches, it flies across the room.

Old Way: They pushed the box using "units" that didn't make sense. Sometimes 1 unit was a feather touch; other times it was a sledgehammer.
New Way: They measured the distance between the "Happy Box" and the "Sad Box" in the model's brain. They defined their "push" based on the actual distance between these two states. Now, they know exactly how hard to push to get the perfect result, no matter how far away the target is.

2. The "Mean-Difference" (MDS) Injection

The paper tested six different ways to push the actor's brain. They found that the best method was called MDS (Mean-Difference).

The Analogy: Imagine you have a crowd of people. Half are wearing red shirts (representing "Kindness") and half are wearing blue shirts (representing "Cruelty").
The MDS method finds the exact center of the Red crowd and the exact center of the Blue crowd. It then draws a straight line between them.
To make the actor "Kind," it gently pushes their internal thoughts along that line toward the Red center. To make them "Cruel," it pushes them toward the Blue center.
The Result: This method was surprisingly effective. In 11 out of 14 different AI models, this "push" worked better than simply asking the AI to "be kind" (prompting).

3. The Hybrid Super-Method (PM)

The authors realized that the best approach wasn't just one or the other. They combined the "Ask nicely" method (Prompting) with the "Push the brain" method (MDS).

The Analogy: It's like telling the actor, "You are a kind person," while simultaneously gently nudging their brain chemistry to actually feel kind.
The Result: This combination (called PM) was the undisputed champion. It outperformed both methods alone in 13 out of 14 models. It made the AI's personality shifts stronger and more reliable.

The "Finding Nemo" Example

The paper includes a funny example where they steered an AI to write about the movie Finding Nemo.

Normal AI: Writes a sweet story about a fish.
Steered AI: Suddenly, the story turns dark. The father fish, Marlin, becomes a "master of manipulation" who enjoys the suffering of others.
Why it matters: The AI didn't just say it was dark; the entire tone, word choice, and flow of the story shifted to match that dark personality, all while still sounding like fluent, natural English. This is something the old "asking" method couldn't do as smoothly.

The Big Discovery: Linear Control

The most exciting finding is that this steering works linearly.

The Analogy: Think of a volume knob. If you turn it a little, the music gets a little louder. If you turn it all the way, it's loud.
The authors found that their "personality knob" works the same way. If they push the AI 10% toward "Openness," it becomes 10% more open. If they push it 50%, it becomes 50% more open. It's predictable and reliable.

The Catch: The "Big Two" Gap

While the AI can be steered to be "Open" or "Neurotic," the paper found something weird. In human psychology, certain traits usually go together in specific patterns (like how "Stability" and "Plasticity" are linked).

The AI's "brain" didn't quite follow these human patterns. When they made the AI more "Open," it didn't automatically adjust the other traits the way a real human would.
What this means: The AI has learned to mimic human words and behaviors, but its internal "psychology" is still a bit alien. It's a brilliant mimic, but not a perfect human soul.

Summary

This paper is a breakthrough in AI Control. It moves us from "hoping the AI listens to our instructions" to "scientifically tuning the AI's brain" to be exactly who we want it to be.

Old Way: "Please be nice." (Sometimes works, sometimes doesn't).
New Way: "Here is the exact mathematical distance to 'Nice,' and here is the perfect amount of force to apply." (Works almost every time).

This opens the door for creating AI companions, therapists, or role-play characters that feel genuinely consistent and deeply aligned with specific human personalities.

1. Problem Statement

Large Language Models (LLMs) exhibit consistent human-like behaviors and personality traits derived from their training data. While researchers aim to steer these models toward specific psychological constructs (e.g., the OCEAN personality traits), existing methods face significant limitations:

Prompting Limitations: While "Personality Prompting" (P2) is a strong baseline, it often struggles with open-ended generation and lacks fine-grained control.
Representation Engineering (RepE) Limitations: Previous attempts to steer models via activation-level interventions (adding vectors to the residual stream) have yielded conflicting results, with some studies claiming prompting outperforms RepE.
Methodological Gaps: Existing RepE methods suffer from two critical issues:
1. Uncalibrated Search Spaces: They sweep injection strengths using arbitrary units (e.g., unit vectors), making it impossible to find optimal settings if they lie far outside the tested range (e.g., requiring 10,000 units).
2. Restricted Search Spaces: They often test only a few discrete strength values (e.g., $[0.4, 0.5, \dots, 1.5]$ ), potentially missing the true optimal intervention point.
3. Fluency Issues: Aggressive interventions often degrade text fluency, but existing methods lack robust mechanisms to constrain this.

2. Methodology

The authors propose a Psychological Steering Framework that addresses these gaps through three main contributions:

A. Calibrated Strength Scale & Unbounded Sweeps

Instead of using arbitrary unit vectors, the authors define injection strength in semantically calibrated centroid units.

They derive residual-stream vectors based on the distance between centroids of activation sets representing a construct and its antithesis.
They perform unbounded sweeps of the injection coefficient ( $\alpha$ ).
Early Stopping: To maintain fluency, the search is automatically stopped if the generated text falls below a fluency threshold (score $\ge 0.95$ using a RoBERTa-based classifier).
Lightweight Evaluation: To avoid expensive API calls during sweeps, they train lightweight logistic regressors on the statement embeddings to estimate construct presence, reserving frontier LLMs (GPT-5.1) only for final comparative evaluation.

B. Derivation of Steering Vectors

The paper compares six types of residual-stream injection vectors derived from 1,000 synthetic statements (500 for the construct, 500 for the antithesis):

L1/L2 Probes: Vectors normal to the decision boundary of L1/L2-regularized logistic regressors (with/without intercept).
Mean-Difference (MD) Vectors:
- MDB: Derived from the difference of centroids of "Yes/No" prefill activations ( $h^b_\ell$ ).
- MDS: Derived from the difference of centroids of "Tell me about yourself" full-utterance activations ( $h^s_\ell$ ).

Key Finding: The authors found that $h^s_\ell$ activations (full utterances) provided better separability and steering effects than $h^b_\ell$ (Yes/No) activations for open-ended tasks. Consequently, MDS (Mean-Difference on full statements) became the primary focus.

C. Evaluation Framework

Instruments: Used the IPIP-NEO-120 inventory and synthesized Situational Judgment Tests (SJTs) for the OCEAN model (Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism).
Hybrid Approach (PM): They introduced a hybrid method combining P2 prompting with MDS injections to test if the two paradigms are complementary.

3. Key Results

The study evaluated 14 LLMs (ranging from 1B to 32B parameters, including Llama-3, Gemma-3, Qwen3, and Olmo-3).

MDS vs. Prompting (P2):
- MDS injections outperformed the established P2 baseline in 11 out of 14 LLMs for open-ended generation (SJTs).
- Performance gains ranged from 3.6% to 16.4%.
- This overturns prior reports (e.g., Banayeeanzade et al.) that claimed prompting was superior to MD injections, suggesting previous failures were due to uncalibrated search spaces.
Hybrid Performance (PM):
- The hybrid method (Prompting + MDS) outperformed both P2 and MDS alone in 13 out of 14 LLMs.
- Gains over P2 ranged from 5.6% to 21.9%.
- Gains over MDS ranged from 3.3% to 26.7%.
Linearity and Control:
- The steering effect varied approximately linearly with the injection coefficient $\alpha$ (96.15% of trends had $R^2 \ge 0.75$ ).
- This supports the Linear Representation Hypothesis, confirming that moving along a specific direction in representation space reliably increases the probability of expressing the corresponding concept.
- Fluency remained high (negligible degradation) within the optimal $\alpha$ range.
Trait Covariance:
- While OCEAN traits are known to be non-orthogonal, the induced covariance patterns from MDS injections did not align with the "Big Two" model of stability and plasticity. This suggests a gap between the latent representations learned by LLMs and human psychological structures.

4. Significance and Contributions

New Frontier in RepE: The paper establishes Representation Engineering as a viable and superior frontier for open-ended psychological steering, provided that search spaces are properly calibrated.
Methodological Rigor: It introduces a standardized framework for "unbounded" sweeps using centroid units, solving the issue of finding optimal injection strengths that were previously missed.
Practical Utility: The MDS method allows for flexible, fluency-constrained steering of multiple distinct constructs within a single generation (e.g., shifting from "conformity to feminine norms" to "sadism" mid-text), a capability prompting cannot achieve.
Insight into Model Psychology: The discrepancy between the induced trait covariance and the Big Two model highlights a fundamental difference between how LLMs encode personality versus how humans psychologically structure it.

5. Conclusion

The authors conclude that with proper calibration (centroid units) and unbounded searching, Mean-Difference (MDS) injections are a powerful tool for steering LLMs. When combined with prompting, they offer the most robust control over psychological traits currently available, outperforming prompting alone in the majority of tested models. This work bridges the gap between psychometric theory and activation engineering, offering a reliable, near-linear control knob for behavioral expression in LLMs.