Spectral Explainer

Spectral Explainer (SPEX) is a framework for discovering influential interactions in large ML models, especially LLMs. SPEX treats interaction discovery as a sparse‑recovery problem by exploiting sparsity (few true interactions) and low‑degeneracy (interactions involve few features). It uses strategically selected ablations of input features, training data, or model components, then decodes combined signals with efficient algorithms to isolate specific interactions. ProxySPEX extends SPEX by leveraging hierarchy, reducing required ablations by ~10×. Experiments on feature, data, and component attribution show SPEX matches the faithfulness of existing interaction methods while scaling to thousands of features. NVIDIA collaborates with AI2F in France AI2F in France collaborated with NVIDIA Inception program for four months Doyub Kim used GPT-6 Astra Alice&Bob integrated NVIDIA NVQLink platform for superconducting cat qubits Quandela is using NVIDIA NVQLink to integrate DGX Spark system with photonic qubits Grayscale report reports Bitcoin’s one‑year realized volatility at 43% TikTok blocks Musk documentary Microsoft released MAI-Base-2, a pre‑trained 3T‑parameter base model OpenAI has Ramamohan Paturi as a shepherd for the computer science breakthrough OpenAI has Ramamohan Paturi as a shepherd for the computer science breakthrough· Entity summary generated by Graphiti