The Emergent Symbolic Structure of Artificial Neural Networks
Paper · Text & code
Shows network vector representations can be closely approximated by symbolic structures, replaced by a closed-form equation with behavior largely preserved
StatusResearch
Evidence
- Sep 2Targeted interventions on LLM internal representations, guided by the symbolic approximation, change behavior in predictable ways
- Sep 2Commenter says section 3.5 shows limited generalizability and robustness so far; method is early compared with mechanistic interpretability