Research, Interpretability, Code
March 2026
Your LLM Understood the Code — Then Forgot the Answer
We use Patchscopes to dissect how Code LLMs process code layer-by-layer, revealing a counterintuitive phenomenon: models often compute the right...
Read More