Standard inference optimizations underperform on Mixture-of-Experts (MoE) models. Evaluating three techniques across multiple MoE architectures, fused Triton kernels reach 5.6x–9.0x speedups in isolation yet only 0.999x end to end as kernel launch overhead dominates; INT4 quantization causes minimal quality loss despite swapping experts, suggesting experts are substitutable rather than specialized; and removing torch.compile graph breaks paradoxically slows execution. Together these results show that routing fidelity and output quality are separable optimization objectives in MoE inference.

RNA foundation models are increasingly used for structure prediction and design, but downstream benchmarks can blur genuine biological learning with task-specific memorization. REDIAL addresses this with a zero-shot, unsupervised diagnostic that extracts coevolutionary signals directly from RNA language model embeddings. By probing models layer by layer, it reveals what they have internalized about RNA structure, showing that current RNA LMs are often overparameterized for available sequence diversity while structure-guided pretraining improves learned base-pair coupling signals.