AI BriefingApple Research Proposes Weighted State Souping for Efficient Long-Context Retrieval
·2026.11.10 09:00
Key point
The method learns query-dependent weights to compose cached state encodings, improving performance on document QA, needle-in-a-haystack retrieval, and tool use.
Details
- Weighted state souping is a method that learns query-dependent weights for composing cached state encodings in state space models (SSMs).
- The approach substantially improves long-context performance across single- and multi-hop document question answering, needle-in-a-haystack retrieval, variable tracking, and tool use.
- An efficient sparse variant of the method yields near-constant encoding cost with respect to context length while preserving or improving performance.
- The paper includes theoretical arguments explaining why souping works and when it may fail, and was accepted at the Workshop on Personalized, Aligned, Long-term Memory for AI Systems (PALM) at NeurIPS 2026.
- Authors include William L. Tong, Aryo Lotfi, Emmanuel Abbe, Josh Susskind, Cengiz Pehlevan, and Eran Malach, with work done while at Apple and Harvard University.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Apple Research Proposes Two-Stage Distillation to Convert Transformers to MambaMemorilla Compresses LLM Memory into 16 Tokens for Better Long-Context Performance