AppleAnalysisAI ModelsNovember 11, 2026

Apple researchers propose weighted state souping for long-context retrieval

Read original source →machinelearning.apple.com

Apple's method learns query-dependent weights to recombine cached SSM state encodings, improving single- and multi-hop document QA, needle-in-a-haystack retrieval, variable tracking and tool use. A sparse variant keeps encoding cost near-constant as context length grows.

1 source

More stories today

Open the live feed