AIBerkeley AI Research4 months ago
Identifying Interactions at Scale for LLMs
8 min read
SPEX and ProxySPEX are algorithms that identify critical interactions in large language models at scale. They use ablation to measure influence by observing changes when a component is removed. The goal is to isolate drivers of decisions by systematically perturbing the system.
Level
Hype check
Tap to vote and see what everyone thinks.