🧬 biology

An evaluative axis organises mental representations of prospective experiences and predicts choice

By analyzing similarity judgments across diverse scenarios, this study reveals that mental representations of prospective experiences are organized along a multidimensional space dominated by a unidimensional evaluative axis (valence) that emerges independently of preference tasks yet strongly predicts subsequent choices.

Original authors: Gabriel Brandolini, Axel Cleeremans

Published 2026-08-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Gabriel Brandolini, Axel Cleeremans

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Every day, our minds run a quiet simulation of what might happen next. We weigh the quiet comfort of a book against the noise of a party, or the absurdity of fighting a giant duck against the chaos of a hundred tiny horses. These choices feel effortless, yet they pose a deep puzzle for scientists: how can the brain compare two things that seem to have nothing in common? To make a decision, the mind must first translate these wildly different experiences into a shared language, a way of placing them side by side so they can be measured against one another. While we know the brain can map the similarities between physical objects or sounds, it remains unclear how it organizes the vast, abstract landscape of future possibilities. Does the mind rely on a single, universal scale of "good" and "bad" to make these comparisons, or does it use a complex web of unrelated features?

A team of researchers at the Université libre de Bruxelles set out to map this invisible landscape of prospective experiences. They gathered a diverse collection of two hundred scenarios, ranging from the mundane, like eating pancakes, to the deeply unpleasant, such as having lemon juice poured on a wound, and the fantastical, like discovering a mythical sea monster. To understand how people mentally organize these options without being told to judge them as good or bad, the researchers asked 150 participants to perform a simple task: they were shown three scenarios at a time and asked to pick the one that felt most different from the other two. This "odd-one-out" game was played 30,000 times, creating a massive dataset of how people perceive the relationships between these imagined futures.

Using a computer model to analyze these thousands of judgments, the researchers reconstructed the mental map the participants were using. The model identified seven distinct dimensions that described how the scenarios related to one another. Some dimensions captured specific themes, such as physical discomfort, social status, or the presence of superhuman abilities. However, when the researchers looked at the overall structure of this map, a striking pattern emerged. The most powerful line running through the entire space was not about how realistic a scenario was, or how disgusting it might be, but rather how positive or negative it felt. This primary axis, which the researchers call an evaluative axis, stretched from experiences that felt like success and wealth to those that felt like suffering and misery.

What makes this finding so significant is that the participants never explicitly rated the scenarios as good or bad. They were only asked to judge similarity. Yet, the very structure of their mental representations was dominated by this feeling of value. The researchers then tested whether this hidden map could predict what people would actually choose. In a second phase, 100 participants were asked to choose which of two scenarios they would prefer to experience. The results were clear: the position of a scenario along that single evaluative axis was the strongest predictor of choice. In fact, once the researchers accounted for this feeling of "goodness" or "badness," no other feature of the mental map added any extra power to predict the decision. The other dimensions, while real and distinct, did not drive the choice on their own.

The study suggests that when we imagine the future, our minds do not just store neutral details. Instead, the representation of a possible experience is already imbued with an emotional value, a built-in sense of how it would feel to us. This evaluative information appears to be intrinsic to the way we simulate the world, emerging even when we are not actively trying to make a decision. It acts as a common currency, a single scale that allows us to compare a book, a party, or a mythical beast, and ultimately guides us toward the path we prefer. The research indicates that while our mental world is rich and multidimensional, the thread that ties it all together and drives our choices is the simple, profound question of whether an experience feels good or bad.

Technical Summary: An Evaluative Axis Organises Mental Representations of Prospective Experiences and Predicts Choice

Problem Statement
People frequently make choices between prospective experiences that appear radically dissimilar (e.g., attending a party versus staying home, or fighting a horse-sized duck versus duck-sized horses). While the ability to compare such incommensurate options is fundamental to decision-making, the mental architecture enabling this comparison remains unclear. Specifically, it is unknown how the mind represents hypothetical scenarios such that they can be compared, and whether these representations are organized by a shared evaluative metric (a "common currency") or by disparate, non-evaluative features. Previous research has mapped representational spaces for objects, words, and emotions, but the dimensions organizing the space of prospective experiences (simulations of future scenarios) are largely uncharacterized. A central theoretical question is whether valence (positive/negative affect) is an intrinsic property of these mental representations or merely a construct generated only at the moment of decision.

Methodology
The study employed a multi-phase approach combining representational geometry, computational modeling, and neuroeconomic-style choice prediction.

Stimulus Curation: The authors curated a semantically diverse set of 200 prospective scenarios (gerund-form descriptions) drawn from a larger dataset of "Would You Rather" questions. The set spanned mundane, aversive, and fantastical domains, selected via farthest-point sampling in a 384-dimensional semantic space (using all-MiniLM-L6-v2) to ensure maximal diversity.
Phase 1: Mapping the Space (Odd-One-Out Task):
- Data Collection: 150 participants performed 30,000 triplet odd-one-out similarity judgments. In each trial, three scenarios were presented, and participants selected the one most dissimilar to the other two.
- Computational Modeling: A Sparse Positive Similarity Embedding (SPoSE) model was fitted to the judgments. This model learns a set of latent, non-negative dimensions where each scenario is represented as a vector of weights. The model was trained to minimize cross-entropy loss between predicted and observed similarity judgments.
- Dimensionality Selection: Through reproducibility analysis (20 random initializations) and dimension elimination, the model retained seven reproducible dimensions from an initial 10. The model achieved 47.2% accuracy on held-out trials (chance = 33.3%, noise ceiling = 72%).
Phase 2: Characterizing the Space:
- Labeling: 24 participants freely labeled the seven dimensions and two principal components (PCs) derived from the embedding.
- Rating Task: 95 participants rated the 200 scenarios on five candidate measures: Valence, Prestige, Disgust, Restriction, and Unreality.
- Representational Similarity Analysis (RSA): The authors constructed Representational Dissimilarity Matrices (RDMs) for the embedding dimensions, PCs, rating measures, and semantic embeddings. They correlated these RDMs to determine structural correspondences.
Phase 3: Predicting Preference:
- Choice Task: 100 participants made 10,000 two-alternative forced-choice (2AFC) decisions, selecting which of two scenarios they would prefer to experience.
- Prediction Models: Mixed-effects logistic regression and leave-one-participant-out (LOPO) cross-validation were used to predict choices based on the embedding dimensions, PCs, and rating measures.
- Decoder Alignment: Linear decoders were trained to map the embedding to valence ratings and to choice outcomes. The alignment between the weight vectors of these decoders was analyzed.

Key Results

Recoverable Structure: The SPoSE model successfully recovered a low-dimensional structure (7 dimensions) that predicted similarity judgments above chance and generated a similarity matrix correlated with empirical data ( $r = .41$ ).
Emergence of an Evaluative Axis: Despite the task being non-evaluative (similarity judgment), the first principal component (PC1) of the representational space was strongly aligned with valence ( $r = .73$ ) and prestige ( $r = .81$ ). PC1 accounted for the largest source of variance and was interpreted by participants as a broad evaluative axis (ranging from success/wealth to suffering/misery). The second component (PC2) corresponded to "unreality" or the spectrum from ordinary to fantastical.
Valence as the Dominant Predictor of Choice:
- In predicting 2AFC choices, valence was the strongest single predictor ( $\beta = 2.01$ ), outperforming all other rating measures, embedding dimensions, and principal components.
- When all rating measures were included in a single model, valence remained the strongest predictor, suggesting other measures (prestige, disgust, restriction) largely share variance with valence.
- Information Exhaustion: Adding the seven embedding dimensions to a valence-only model did not improve the prediction of held-out choices ( $\Delta \text{log loss} = 0.0011, p = .134$ ). Conversely, adding valence to a model containing only the seven dimensions significantly improved prediction ( $\Delta \text{log loss} = 0.0615, p < .001$ ). This indicates that the choice-relevant information in the entire representational space is exhausted by valence.
Subspace Identity: The weight vector of the valence decoder and the weight vector of the choice decoder were nearly identical (cosine similarity = .991; angle = $7.82^\circ$ ). The subspace capturing valence is effectively the same subspace driving choice.
Double Dissociation: Removing the valence direction from the embedding barely impaired the model's ability to predict similarity judgments (accuracy dropped from .459 to .450), yet it exhausted the model's ability to predict choices. This suggests valence is encoded in the representation but is not the primary driver of similarity judgments; however, it is the sole driver of preference.

Significance and Claims
The paper claims that the mental representation of prospective experiences is organized by a unidimensional evaluative ordering that emerges even in the absence of explicit preference-based tasks.

Intrinsic Evaluation: The findings suggest that evaluative information (specifically valence) is intrinsic to the representation of prospective experiences, rather than being computed only at the moment of decision. This supports the hypothesis that the mind automatically imbues simulations of future states with affective consequences.
Common Currency: The results align with the "common currency" hypothesis of value-based decision-making. While the representational space is multidimensional (allowing for complex comparisons), it contains a dominant evaluative axis that allows for the ordering of otherwise incommensurate experiences.
Comparison to Object Representations: Unlike the representational space of physical objects (which is organized by categorical dimensions like "animal-relatedness" or "food-relatedness"), the space of prospective experiences is dominated by a global evaluative dimension. The authors attribute this to the nature of prospective simulation, which inherently involves the subject's relation to the experience ("how it would feel to me").
Modesty: The authors note that their dimensions were "fuzzier" and less discrete than those found for objects, likely due to the semantic diversity of the stimuli and the smaller dataset size. They do not claim to have proven that all phenomenal experience is valenced, but rather that prospective experiences are organized by value in a way that predicts choice.

In summary, the study demonstrates that while the mind represents prospective experiences along multiple dimensions, a single evaluative axis (valence) organizes this space and serves as the primary mechanism for comparing and choosing between disparate future possibilities.

More like this