Examining Reject Relations in Stimulus Equivalence Simulations

Alexis Carrillo, Asieh Abolpour Mofrad, Anis Yazidi, Moisés BetancortarXiv 2025 · DOI · arXiv

Introduction

In the field of behavioral psychology, Stimulus Equivalence (SE) is considered a foundational element of language acquisition and complex symbolic reasoning. Human learners possess a unique ability: if taught that "A equals B" and "B equals C," they can spontaneously derive that "A equals C" (transitivity), "C equals A" (equivalence), and "B equals A" (symmetry) without explicit instruction. While researchers have long used Artificial Neural Networks (ANNs) to simulate these cognitive processes, a significant debate remains regarding whether these models are truly forming relational "equivalence classes" or are simply utilizing clever associative shortcuts.

This paper investigates the role of "reject relations"—the process of learning what a stimulus is not—to determine if modern AI architectures, including Transformers like BERT and GPT, are capable of genuine relational reasoning or if they rely on simple exclusion strategies.

Experimental Methodology

The researchers conducted 72 simulations across 18 distinct experimental conditions. They compared three primary computational agents: standard Feedforward Neural Networks (FFNs), state-of-the-art Transformers (BERT and GPT), and a baseline Probabilistic Agent designed to use purely associative memorization.

The experimental design manipulated three key factors:

  1. Training Structure: The arrangement of taught relations, specifically Linear Series (A→B→C), One-To-Many (A→B, A→C), and Many-To-One (B→A, C→A).
  2. Relation Type: Whether the model was trained to identify the correct match ("Select-only"), identify incorrect matches ("Reject-only"), or both ("Select-Reject").
  3. Negative Comparison Selection: A "Standard" condition with few wrong choices versus a "Biased" condition that exposed the model to a high density of incorrect associations.
The three training structures compared in the study: Linear Series, One-To-Many, and Many-To-One.

The Role of Reject Relations in Matching-to-Sample Tasks

A core component of the study is the Matching-to-Sample (MTS) task. In this setup, a model is presented with a sample stimulus and must choose the corresponding stimulus from a set of comparisons. The researchers hypothesized that models might pass these tests not by understanding the relationship between the sample and the correct comparison, but by learning to reject the "distractor" stimuli.

An illustration of the Matching-to-Sample (MTS) task, highlighting the distinction between selecting a correct stimulus and rejecting incorrect ones.

When "reject" information is abundant, models can use an "exclusion" shortcut. For example, if a model knows that "A1" is definitely not "B3" or "B4," it will select "B1" by default. This strategy mimics successful relational reasoning without requiring the model to understand the underlying symbolic link.

Key Insights and Results

The simulations revealed several critical findings that challenge the perceived symbolic capabilities of modern AI:

  • Mastery is Rare: Out of 72 simulations, only 6 reached the 90% mastery criterion on equivalence tests. This suggests that emergent relational properties are not a default feature of these architectures.
  • The Exclusion Shortcut: Performance improved significantly in the "Biased" condition, where models were given more information about what stimuli were not related. This confirms that reject relations are a major confounding variable; when models were denied this negative information (the "Select-only" condition), accuracy often plummeted below 40%.
  • Transformers vs. Associative Agents: Despite their complexity, BERT and GPT frequently produced response patterns nearly identical to the simple Probabilistic Agent. This implies that even advanced Large Language Models (LLMs) may rely on basic associative strategies for symbolic tasks rather than human-like logic.
  • One-To-Many (OTM) Failure: While OTM training is highly effective for humans, the AI models struggled most with this structure. This divergence highlights a fundamental difference in how biological and artificial networks process hierarchical information.

Conclusion and Future Directions

The study demonstrates that current ANN architectures often rely on exclusion and associative mirroring rather than genuine stimulus equivalence. The "peak" performance observed in BERT (reaching 95-99% accuracy) occurred only under highly specific, biased conditions that favored exclusion strategies.

To bridge this gap, the authors suggest moving toward "online" learning models that learn trial-by-trial and exploring Reinforcement Learning (RL) architectures. Future research should also look beyond output behavior to analyze the internal hidden layers of these models, seeking to understand how—or if—they represent symbolic relationships internally.

Summary updated

BibTeX
@misc{carrillo2025examining,
  author = {Alexis Carrillo and Asieh Abolpour Mofrad and Anis Yazidi and Moisés Betancort},
  title = {{Examining Reject Relations in Stimulus Equivalence Simulations}},
  year = {2025},
  eprint = {2507.00265},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.00265},
}
Figure