Testing Stimulus Equivalence in Transformer-Based Agents

This study investigates the ability of transformer-based models (TBMs) to form stimulus equivalence (SE) classes. We employ BERT and GPT as TBM agents in SE tasks, evaluating their performance across training structures (linear series, one-to-many and many-to-one) and relation types (select–reject, select-only). Our findings demonstrate that both models performed above mastery criterion in the baseline phase across all simulations ( n = 12). However, they exhibit limited success in reflexivity, transitivity, and symmetry tests. Notably, both models achieved success only in the linear series structure with select–reject relations, failing in one-to-many and many-to-one structures, and all select-only conditions. These results suggest that TBM may be forming decision rules based on learned discriminations and reject relations, rather than responding according to equivalence class formation. The absence of reject relations appears to influence their responses and the occurrence of hallucinations. This research highlights the potential of SE simulations for: (a) comparative analysis of learning mechanisms, (b) explainability techniques for TBM decision-making, and (c) TBM bench-marking independent of pre-training or fine-tuning. Future investigations can explore upscaling simulations and utilize SE tasks within a reinforcement learning framework.