TALES (Tracking And Leveraging Engaging Story-elements) aims to uncover the computational bases of narrative pleasure — a key form of intrinsic reward underpinning motivation and mental health. Combining experimental psychology, behavioral economics, and artificial intelligence, the project will quantify how suspense, surprise, and emotional valence shape the enjoyment of stories. Three work packages structure the thesis: (1) experimental measurement of the links between these features and pleasure; (2) use of large language models to predict story success from temporal profiles of suspense, surprise, and valence; and (3) investigation of how sensitivity to these signals varies across individuals and relates to depression and addiction. TALES fits within the DIM C-BRAINS axis on cognition, behavior, and computational modeling.

What makes stories — in books, films, or games — so compelling that people devote hours to them, often at the expense of other activities? Stories are intrinsically rewarding: they provide pleasure for their own sake, not because of external outcomes. Yet, despite centuries of research in narratology and psychology, the quantitative features that make stories rewarding remain unknown. The TALES project aims to identify and model the computational features that drive narrative pleasure, focusing on three key dimensions: emotional valence, suspense, and surprise.

*Scientific objectives

TALES will combine approaches from experimental psychology, behavioral economics, and artificial intelligence to investigate how these dimensions interact to shape intrinsic pleasure. The central hypothesis is that suspense and surprise interact with affective valence to determine enjoyment: highly suspenseful narratives resolving positively, as well as (positive) surprising outcomes should be the most pleasurable on average. The project has three main objectives:

1)Quantify the contribution of suspense, surprise, and valence to subjective pleasure in controlled experimental tasks and their textual equivalents.

2) Test the ecological validity of these features by ausing large language models to published stories, assessing how suspense, surprise, and valence predict real-world success (downloads, ratings).

3) Investigate interindividual variability in sensitivity to these intrinsic rewards and its relation to mental health dimensions, particularly depression and addiction.

*Background and positioning

While narratology and psychology have proposed qualitative notions such as emotional valence, suspense, and surprise, these concepts lack formal definitions that allow quantitative testing. Recent developments in computational modeling provide a way forward. In economics, suspense can be defined as the expected change in belief — the difference between what we currently believe and what we expect to believe in the near future. Surprise corresponds to the actual change in belief once new information is revealed. These definitions can be formalized mathematically and applied to sequences of events, such as story plots.

Importantly, positive surprises (prediction errors associated with good outcomes) are known to be intrinsically rewarding, while suspense can be conceptualized as the anticipation of such belief changes. TALES will integrate these notions into a unified quantitative model of narrative pleasure. Beyond its contribution to fundamental knowledge, the project seeks to bridge cognitive modeling, natural language processing, and computational psychiatry, in line with the interdisciplinary vision of the DIM C-BRAINS initiative.

*Methodology and work plan

**Work Package 1 – Experimental quantification of intrinsic reward features

In WP1, participants will complete behavioral tasks designed to generate well-controlled sequences of events that vary in suspense, surprise, and valence. Using a card-based paradigm inspired by the Black-Jack game, the task will allow independent manipulation of these dimensions. Participants will rate their feelings of suspense (before outcomes), surprise (after outcomes), and pleasure (liking). The same paradigms will be adapted into narrative (text-based) versions to test generalization to story contexts. The main prediction is that objective suspense and positive surprise (as defined by computational models) will best predict subjective enjoyment.

A second experiment will assess wanting rather than liking: participants will freely choose between lotteries differing in expected value, variance, and ambiguity. Preference for suspenseful or surprising options will reveal the motivational component of intrinsic reward.

**Work Package 2 – Computational modeling of story pleasure using neural networks

WP2 will extend these principles to real-world stories. Using pre-trained Bayesian recurrent neural networks, the project will compute paragraph-level time series of suspense, surprise, and valence in large text corpora (e.g., Project Gutenberg, IMDb subtitles). The model will estimate suspense as the expected divergence between predicted and future sentence distributions, and surprise as the realized deviation. Valence will be fine-tuned using affective annotation datasets.

These features will be used to predict the success of published books and movies (e.g., downloads, ratings). The hypothesis is that stories with higher proportions of suspense and positively valenced surprises will achieve greater popularity, validating the intrinsic reward framework on ecological data.

**Work Package 3 – Individual differences and links to psychopathology

In WP3, a large-scale behavioral study (N > 200) will examine how individual sensitivity to suspense and surprise relates to mental health. Participants will perform the tasks from WP1, alongside standardized questionnaires assessing depression, apathy, anhedonia, and addictive behaviors. Dimensionality reduction (e.g., PCA) will be applied to identify latent components. It is hypothesized that hyposensitivity to suspense and positive surprise will correlate with depressive symptoms, while hypersensitivity will correlate with addictive tendencies. This approach aligns with the Research Domain Criteria (RDoC) framework, emphasizing computational mechanisms underlying transdiagnostic dimensions of psychopathology.

*Expected outcomes and significance

TALES will provide the first quantitative framework linking narrative structure, intrinsic reward, and mental health. The project will bring:

**A computational model of narrative pleasure, combining psychological theory with mathematical definitions of suspense and surprise.

**Validated neural network tools for estimating narrative features from text, applicable to cultural analytics and affective computing.

**New insights into mental health, by identifying how sensitivity to intrinsic rewards varies across individuals and contributes to depression or addiction.

Beyond academic impact, TALES addresses societal questions about how media engage human cognition and emotion. Understanding the mechanisms that make stories pleasurable could inform education and digital media design.

*Integration within DIM C-BRAINS

This PhD project fits squarely within the DIM C-BRAINS themes on cognition, behavior, and computational modeling. It bridges psychology, computational psychiatry, and AI to explore how cognitive and affective mechanisms drive engagement and well-being. Hosted at the Centre d’Économie de la Sorbonne (Paris 1) and involving collaborations with experts in cognitive psychology, decision theory, and language modeling, TALES embodies the interdisciplinary spirit of the DIM network.

By quantifying what makes stories rewarding and how this relates to mental health, TALES will contribute to understanding the neural and computational bases of intrinsic motivation—a central challenge for cognitive science and psychiatry.