DMDIntel Interprets LLM Predictions via Dynamic Mode Decomposition

DMDIntel decomposes LLM hidden states into prominent modes to perform input attribution for classification tasks. Experimental results show it outperforms PCA, Integrated Gradients, and SHAP in ranking token importance across multiple model families.