This paper identifies a three-stage emergent symbolic architecture within large language models (LLMs) that underpins their abstract reasoning capabilities. The research investigates the internal mechanisms supporting abstract reasoning in LLMs to address ongoing debates regarding the robustness of these capabilities and their reliance on structured reasoning.

The core methodology of this study involves the mechanistic interpretation of LLMs to uncover how they implement abstract reasoning. The identified emergent architecture comprises a sequential series of three distinct computational stages, each performed by specialized "heads" within different layers of the neural network:

1. Symbol Abstraction Heads: These mechanisms are primarily located in the early layers of the LLM. Their function is to transform input tokens into abstract variables. This abstraction is achieved by analyzing and encoding the relational properties between the input tokens, effectively mapping concrete linguistic elements to more generalized, abstract representations or symbols.
2. Symbolic Induction Heads: Positioned in the intermediate layers of the LLM, these heads operate on the abstract variables generated by the symbol abstraction heads. Their role is to perform sequence induction over these abstract variables. This implies recognizing patterns, inferring rules, or predicting continuations within the sequence of abstract symbols, essentially carrying out symbolic manipulation and reasoning tasks.
3. Retrieval Heads: Found in the later layers of the LLM, the final stage involves these heads. After abstract reasoning has been performed by the symbolic induction heads, the retrieval heads are responsible for predicting the next token. This prediction is made by retrieving the concrete value or token that corresponds to the abstract variable predicted by the preceding stages, effectively translating the abstract reasoning outcome back into a tangible output token.

This emergent symbolic architecture suggests a resolution to the longstanding debate between symbolic and neural network approaches to artificial intelligence. The findings indicate that sophisticated reasoning in neural networks, despite their generic architectural origins, can arise through the spontaneous emergence of symbol-processing mechanisms during training. This work contributes to the field of mechanistic interpretability by illuminating how LLMs develop complex, specialized circuits for abstract reasoning through learning.