Steering LLMs with Psychological Variables

Finding a psychologically meaningful direction inside a language model, pushing on it, and seeing what moves — a first test on coping potential.

1. What is steering?

At Heura Lab, we're exploring whether psychologically meaningful variables, not just demographics, not just topics, but the actual internal states that drive human behavior, can be identified and directly manipulated inside language models. If they can, it opens a path toward simulating populations that don't just sound like different people, but behave differently for the same underlying psychological reasons real people do.

Steering is the technique at the center of that bet. It's part of a broader effort called mechanistic interpretability, work that opens up deep neural networks to understand what's actually happening inside them, rather than treating the model as a black box.

What makes steering powerful is that it operates one level below the prompt. Prompting is indirect: you describe what you want in words and hope the model interprets it correctly. Steering skips that translation. You reach directly into the model's internal activations, the actual numbers it's computing with as it generates, and push them toward the concept you want. No retraining, no fine-tuning, just a precise nudge at the exact point where the model's internal state is formed.

At a given layer l and generation step t:

hₗ,ₜ is the model's activation at that point. v is a direction in activation space associated with some concept or property. α controls how strongly we push the model in that direction.

The rest of the model keeps computing as usual, just from this nudged starting point.

That's the mechanism. The interesting part is what direction is worth finding, and what it actually buys you once you have it.

2. Finding a direction: coping potential

We now have a mechanism for intervention. The next question is what to intervene on.

Appraisal theory gives us a place to start. It holds that emotions aren't triggered directly by events, they're produced by how a person evaluates the event. Two people can lose the same job and land in different emotional states, because one appraises it as “I have no options” and the other as “I will find a way forward.” Same event, different appraisal, different outcome.

One of these evaluations, coping potential, the perceived ability to influence or manage a situation, is a strong first target: it isolates cleanly. Hold the triggering event fixed, vary only the response, and any resulting shift in the model's behavior can be attributed to that one dimension.

Run both through the model, record the activation at a chosen layer, repeat across several dozen pairs, average each side, subtract:

This method, Contrastive Activation Addition, requires no fine-tuning and no learned probe, just the measured difference between two sets of internal activations.

We ran these experiments on Qwen2.5-7B-Instruct (28 transformer layers, 3,584-dimensional hidden states), on a single T4 GPU via Google Colab. The steering vector was extracted and injected at layer 14, the model's approximate midpoint, consistent with prior work showing that psychological and emotional content in transformer models tends to consolidate in the middle layers rather than at the very start or end of the network. We calibrated the intervention strength on a held-out set of 20 labeled scenarios, generating text at increasing α until we found a range that reliably shifted output toward the intended label without breaking coherence. That calibrated range is what powers the results below.

3. Results

Tested on a prompt the vector never saw, coffee spilled on a laptop before a presentation, the effect is exactly what the theory predicts:

One internal direction, extracted from thirty contrastive examples, moves the model from resignation to resourcefulness on a problem it's never encountered, without altering the facts of the situation. The laptop stays wet in both versions. What changes is only what the person does about it, which is precisely the lever coping potential is supposed to be.

This is the proof of concept: a psychologically grounded variable, extracted with a lightweight, inference-time method, causally shifts model behavior in the direction psychological theory predicts.

4. The open problem

The coffee-spill result held cleanly. Push the same calibrated vector across a wider range of situations, and it starts to reveal exactly where the current method's precision runs out, which is the frontier Heura is now working to close.

• Severity calibration: on an unrelated deadline threat, low coping didn't produce proportionate resignation, it escalated into unrelated catastrophe. The magnitude of the intervention isn't yet tied to the magnitude of the situation.
• Constraint preservation: on a genuinely unfixable setback, a broken leg before a marathon, high coping didn't produce a better response to the injury, it erased the injury. A precise variable should change behavior given a constraint, not delete the constraint.
• Signal interaction: on an emotionally loaded prompt, a friend's betrayal, high coping collided with the prompt's own emotional content rather than resolving cleanly, producing internal contradiction.
• Relevance detection: applied to a prompt with no obstacle at all, a promotion, the vector still imposed a coping narrative, with no mechanism yet to recognize when the dimension isn't the relevant one.

These failures suggest that coping potential may not be represented as a single linear direction. The extracted vector may entangle coping with sentiment, narrative resolution, agency, or other correlated features.

This defines the next research question: can we recover the causal structure underlying psychological variables inside the model, rather than merely directions correlated with them?

5. Why this matters

This small experiment, one dimension, one lightweight method, already shows something important: LLMs appear to contain internal structure consistent with appraisal dimensions from the data they're trained on, and that structure can be extracted and manipulated through steering. The results also give a clear picture of exactly where that extraction currently breaks down.

That combination is what makes this a foundation rather than a dead end. At Heura, the next step is to move from individual steering directions toward causal structure: identifying multiple psychological variables, separating them from correlated representations, intervening on them independently, and mapping how they interact to produce behavior. Coping potential is the first dimension tested. The broader goal is to recover enough of this internal psychological structure to build agents whose behavior changes for the same underlying reasons human behavior does.