💬 NLP

Internal narratives parameterise affective states

Through two large-scale studies utilizing large-language-model representations, this research demonstrates that the structure and dynamics of internal narratives not only predict depression severity but also causally influence affective states, suggesting that affect functions as a computational state that constrains narrative construction and integrates context.

Original authors: Jakub Onysk, Quentin J. M. Huys

Published 2026-05-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jakub Onysk, Quentin J. M. Huys

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

The Big Idea: Your Feelings Have a "Shape"

Imagine your mood isn't just a vague feeling like "sad" or "happy." Instead, think of it as a hidden blueprint or a master recipe that dictates how you think, what you remember, and how you describe your day.

The authors of this paper propose that this "mood blueprint" acts like a filter. When you are in a low mood, that filter makes it hard to see positive things and easy to see negative things. They wanted to see if they could measure this invisible blueprint by looking at the words people write.

To do this, they used Large Language Models (LLMs)—the same kind of AI that powers chatbots. They treated the AI not just as a chatbot, but as a "digital microscope" that can look inside the structure of human language to find the hidden patterns of depression.

The Experiment: Two Studies

The researchers ran two big studies with over 1,200 people to test their theory.

Study 1: The "Translation" Test

The Setup:
Imagine you have a standard medical checklist for depression (like the PHQ-9), where you check boxes like "I feel tired" or "I feel hopeless."

The Twist: Instead of just checking boxes, the participants were asked to write free-form stories about how they felt regarding those specific topics.
The AI's Job: The researchers fed these written stories into an AI. They asked the AI: "Based on this story, if this person had to fill out the standard checklist, what would they check?"

The Findings:

The AI was surprisingly accurate. When the AI read a person's story about feeling "hopeless," it could predict their score on the standard checklist with high accuracy.
The "Geometry" of Depression: This is the most important part. The researchers found that depression isn't just a list of separate symptoms. It's a connected web. For example, if someone writes about feeling tired, the AI predicted they would also write about feeling bad about themselves.
The Metaphor: Think of depression symptoms like a mobile hanging from the ceiling. If you push one piece (tiredness), the whole structure sways in a specific, predictable way. The AI didn't just guess the answer; it understood the shape of the mobile. If the AI got the shape wrong, it couldn't predict the answers correctly. This proved that the AI was actually capturing the underlying "structure" of the person's mood, not just memorizing words.

Study 2: The "Mood Weather" Test

The Setup:
If your mood is a "blueprint," does it change when the environment changes? The researchers wanted to test this.

They asked participants to listen to "diary entries" read aloud. Some entries were very sad (like someone describing a terrible day), and some were very happy (like someone describing a great day).
The participants had to "act" these out by typing their own diary entries in the same style.
Then, they had to re-evaluate their own feelings.

The Findings:

The Mood Shift: When people listened to the sad stories and wrote in that style, their own self-reported mood and memory became more negative. When they listened to happy stories, they became more positive.
The AI as a Tracker: The researchers used the AI to measure the "severity" of the stories the participants wrote during the experiment.
The Result: The AI detected the change in the participants' internal "blueprint" immediately. The more deeply a participant got into the role (the more they matched the style of the sad or happy story), the more their own mood shifted.
The Metaphor: Think of the mood blueprint like a river current. If you throw a leaf (a person) into a fast, dark current (sad stories), the leaf gets swept along that path. The AI could measure exactly how fast the leaf was moving and in which direction, proving that the "current" of the story changed the "flow" of the person's mind.

What Does This Mean? (According to the Paper)

The paper argues that our emotions work like a computational state.

Constraint: Your mood limits what you think about. If you are sad, your brain narrows its focus to sad things, just like a spotlight that only shines on dark corners.
Sensitivity: Your mood is sensitive to new information. If you hear a sad story, your "spotlight" shifts to match that story.

The researchers found that AI models are surprisingly good at seeing this spotlight. Because AI models are trained to predict the next word in a sentence based on context, they naturally learn to mimic how human moods constrain our thoughts.

Summary

The Problem: We don't fully understand how our feelings shape our thoughts and words.
The Solution: Use AI to read our free-form writing and map out the hidden "shape" of our depression.
The Discovery: The AI successfully predicted standard depression scores from free writing and showed that depression has a specific, consistent structure (like a mobile).
The Dynamic: When we change the stories we tell ourselves (or listen to), our internal "mood blueprint" shifts, and the AI can measure that shift in real-time.

The paper concludes that by treating emotions as "computational states" that constrain language, we can use AI to quantify and understand the mechanics of human feelings, specifically in the context of depression.

Technical Summary: Internal narratives parameterise affective states

Problem Statement

The mapping between internal narratives (verbalized thoughts and feelings) and underlying affective states, particularly in the context of depression, remains poorly understood. While psychological assessments rely heavily on verbal descriptions to gauge mental health, the structural and dynamic relationship between these narratives and the latent geometry of affective disorders is not well quantified. Previous work has established that affective states bias attention, memory, and language use, and that moods can be viewed as computational states that constrain behavior. However, a framework that treats affective states as computational entities capable of being parameterized by the structure of internal narratives is lacking. This paper addresses the question: Can Large Language Models (LLMs), which function by maximizing context-consistent token prediction, serve as a computational proxy to quantify and predict the latent structure and dynamics of human affective states?

Methodology

The authors conducted two large-scale studies (total $N=1257$ ) utilizing open-ended verbalizations of depressive symptoms and large-language-model representations to test three hypotheses.

Study 1: Structural Parameterization

Participants: $N=704$ fluent English speakers with varying degrees of depression severity.
Procedure: Participants answered eight open-ended questions derived from the PHQ-9 (Patient Health Questionnaire) regarding specific depressive symptoms (e.g., mood, energy, sleep). They subsequently completed standard multiple-choice questionnaires: PHQ-9, GAD-7 (anxiety), and SDS (Self-rating Depression Scale).
LLM Analysis:
- Prediction: The authors constructed prompts containing a participant's open-ended response and asked various LLMs (Gemma2 2B/9B, Llama-3.1/3.2, Mistral-7B) to predict the corresponding multiple-choice label (A-D) via logit sampling.
- Latent Structure: To test if LLMs capture the covariance of symptoms (the latent geometry), the authors compared the pairwise item correlations of the ground-truth questionnaires against the correlations predicted by the LLMs based on open-ended inputs.
- Sparse Auto-Encoder (sSAE): A supervised sparse auto-encoder was trained on the LLM's hidden states (averaged across the open-ended responses) to predict z-scored PHQ-9 scores. This allowed for the extraction of a "depression symptom subspace."
- Perturbation: The authors perturbed the latent sSAE representations in specific directions (increasing or decreasing symptom severity) and observed the resulting shifts in the LLM's sampled multiple-choice responses to test sensitivity.

Study 2: Temporal Dynamics and Contextual Sensitivity

Participants: $N=553$ participants.
Procedure: Participants underwent a mood induction protocol using emotionally evocative "diary entries" (representing high or low mood extremes).
- Task: Participants listened to and recreated these diary entries, then wrote new continuations. A subset (autobiographical) was asked to recall similar life events, while the other (non-autobiographical) acted out a character.
- Measures: Baseline and follow-up assessments included momentary mood (Visual Analogue Scale), emotional word recall (sentiment analysis), and open-ended re-evaluations of mood/energy.
Analysis: The authors used the sSAE model trained in Study 1 to quantify the severity of the "Feeling down, depressed, or hopeless" symptom in the participants' continuation diaries. They analyzed how changes in these internal narratives correlated with changes in self-reported affective measures.

Key Results

1. Prediction of Self-Reported Scores

The LLMs successfully predicted standard multiple-choice depression scores based solely on open-ended verbalizations.

Correlation: The best-performing model (Gemma2 9B) showed moderate-to-high correlations ( $r = 0.56 - 0.78$ ) between predicted and ground-truth scores across all eight symptom items. This performance is comparable to the test-retest reliability of the PHQ-9 itself.
Bias: A systematic bias was observed where the model overestimated severity for "no symptom" responses and underestimated it for high-severity responses, likely due to the lack of explicit negation in open-ended text.

2. Recovery of Latent Symptom Structure

The study demonstrated that LLM representations capture the underlying covariance structure of depression.

Correlation Matrices: The pairwise item correlations predicted by the LLMs from open-ended inputs closely matched the ground-truth correlation matrices of the PHQ-9, GAD-7, and SDS.
Structural Similarity: Metrics including Singular Value Decomposition (SVD) bases and projection measures confirmed that the LLM-derived structures aligned with the ground-truth latent geometry. Poorer-performing models failed to capture this structure, suggesting that capturing the covariance pattern is essential for construct validity.

3. Latent Perturbation and Sensitivity

The sSAE latent space proved sensitive to specific perturbations.

Steering: By perturbing the hidden state representations in the direction of specific symptoms (e.g., increasing "hopelessness"), the authors successfully steered the LLM's sampled responses toward higher or lower severity scores.
Effect Size: Significant Cohen's $d$ effect sizes were observed for most items, confirming that the latent space contains a meaningful, manipulable representation of symptom severity.

4. Dynamics and Contextual Sensitivity

Study 2 confirmed that affective states are sensitive to contextual changes and that these changes are reflected in internal narratives.

Induction Success: Mood induction successfully altered momentary mood, word recall sentiment, and self-reported symptom severity.
Narrative Tracking: The severity of the continuation diaries, as quantified by the sSAE, predicted the magnitude of affective change.
Adherence: The degree to which participants adhered to the emotional tone of the intervention (similarity of their text to the induced state) predicted the extent of mood change.
Baseline Prediction: The baseline severity of a participant's internal narrative predicted the magnitude of their subsequent affective change, supporting the hypothesis that affective states constrain the space of possible thoughts.

Significance and Claims

The paper claims to provide a computational framework for understanding human affective states by drawing parallels between the function of affect and the mechanics of transformer-based LLMs.

Affective States as Computational Constraints: The authors argue that affective states function to maximize temporal and contextual consistency, constraining the space of possible language responses. This aligns with how LLMs maximize next-token prediction given context.
Quantification of Latent Geometry: The study demonstrates that internal narratives, when quantified via LLM representations, capture the latent geometry of depression. This suggests that the "structure" of a disorder is not just a statistical artifact of questionnaires but is embedded in the semantic relationships of how individuals verbalize their experiences.
Therapeutic Implications: The ability to quantify affective states and their dynamics offers a potential tool for:
- Assessment: Gauging depression severity from any free-text source.
- Therapy Monitoring: Tracking the process and outcomes of verbally rich psychological therapies.
- Precision Psychiatry: Identifying individual components of psychological change to design targeted interventions.
Meta-Learning Parallel: The paper posits that LLMs, trained on vast human text, have meta-learned the input-output dependencies of human affective states (e.g., low self-worth preceding hopelessness). This emergent parallel allows LLMs to act as a probe for human meta-reasoning and the "intractable problem" of deciding what to think about.

The authors remain modest, noting limitations such as the specific focus on depression (PHQ-9), the potential variability in open-ended response quality, and the need for validation across diverse populations and other mental health syndromes. They do not claim to replace clinical diagnosis but rather to offer a scalable, computational method for parameterizing and understanding the dynamics of affective states.