LLM Compression with Jointly Optimizing Architectural and Quantization choices

In three linesDifferentiable NAS framework for LLM compression jointly optimizing architecture and mixed-precision quantization of linear layers. Results: 1.4x faster inference than sequential NAS-then-quantization baseline, or 6% higher average accuracy across seven reasoning tasks at equivalent latency.

Summary generated by Claude — human-verified