A Mechanistic Interpretation Of Syllogistic Reasoning In Auto-regressive Language Models

Kim Geonhee, Valentino Marco, Freitas André. Arxiv 2024

[Paper]    
Attention Mechanism Ethics And Bias Model Architecture Reinforcement Learning Training Techniques

Recent studies on logical reasoning in auto-regressive Language Models (LMs) have sparked a debate on whether such models can learn systematic reasoning principles during pre-training or merely exploit superficial patterns in the training data. This paper presents a mechanistic interpretation of syllogistic reasoning in LMs to further enhance our understanding of internal dynamics. Specifically, we present a methodology for circuit discovery aimed at disentangling content-independent reasoning mechanisms from world knowledge acquired during pre-training. Through two distinct intervention methods, we uncover a sufficient and necessary circuit involving middle-term suppression that elucidates how LMs transfer information to derive valid conclusions from premises. Furthermore, we investigate how belief biases manifest in syllogistic reasoning, finding evidence of partial contamination from additional attention heads responsible for encoding commonsense and contextualized knowledge. Finally, we explore the generalization of the discovered mechanisms across various syllogistic schemes and model sizes, finding that the identified circuit is sufficient and necessary for all the schemes on which the model achieves high downstream accuracy (\(\geq\) 60%). Overall, our findings suggest that LMs indeed learn transferable content-independent reasoning mechanisms, but that, at the same time, such mechanisms do not involve generalisable and abstract logical primitives, being susceptible to contamination by the same world knowledge acquired during pre-training.

Similar Work