Constructive Circuit Amplification: Improving Math Reasoning in LLMs via Targeted Sub-Network Updates

Researchers propose Constructive Circuit Amplification, a new method for improving LLM mathematical reasoning by directly targeting and strengthening specific neural network subnetworks (circuits) responsible for particular tasks. This approach builds on findings that model improvements through fine-tuning often result from amplifying existing circuits rather than creating new capabilities.

• →LLMs contain sparse subnetworks called circuits that are responsible for specific tasks like mathematical reasoning.
• →Fine-tuning improvements often come from strengthening existing circuits rather than creating new ones.
• →Constructive Circuit Amplification allows direct intervention on circuits for precise, task-targeted model updates.
• →The method identifies pivotal tokens to improve mathematical reasoning capabilities in language models.
• →This approach could lead to more efficient and targeted AI model improvements without full retraining.