Not All Thoughts Are Equal *

Developed a flip test to identify the critical reasoning steps in LLMs' chain-of-thoughts and studied the internal representation of these steps using residual stream analysis. Currently working on a mechanism to restore the correct reasoning path when critical steps are corrupted.