A Unified Approach to Interpreting Knowledge Distillation for Large Language Models via Interactions

Collection of 10 new ML/LLM research papers (arXiv) covering: a unified interpretation of knowledge distillation via interaction sparsification and a proposed Complex Interaction Penalty (CIP); iLENS, an interpretable LLM-guided mixture-of-experts framework for Alzheimer’s survival analysis; signed symmetric quantization for few-bit integer weights that preserves symmetric runtime while reducing quantization error; StickyMoE, a routing-consistency loss to reduce expert switching in MoE models; Reward Transport, an optimal-transport-based control knob for molecular generation; Director, an on‑d