Scale joint W/A screening to certifiable large-model deployments

Evaluate the population certification and computational cost of joint weight/activation screening, particularly for large language models, beyond the specified extension of the forward-only screening procedure.

Background

The paper introduces behavioral fiber mass sensitivity (BFMS) as a forward-only method for screening layer-precision choices and reports results on small controlled decoders. It also specifies a joint W/A extension, but the reported experiments do not establish whether that extension can provide population-level certification or remain computationally practical at large model scales.

Resolving this problem would determine whether the screening workflow can be used for joint weight and activation precision allocation in substantially larger LLMs, where exhaustive certification and search may be considerably more expensive.