Tag
LaMR introduces a structured pruning framework for coding agents that decomposes code relevance into semantic evidence and dependency support dimensions, using dedicated CRFs and a mixture-of-experts gate to reduce token usage by up to 31% while maintaining or improving task performance.