https://sparsenotes.com/series/weekly-papers-2026-06-06/ Recent content in weekly-papers-2026-06-06 on Sparse Notes https://sparsenotes.com/images/og-default.png https://sparsenotes.com/images/og-default.png Hugo -- gohugo.io Sat, 06 Jun 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/06/papers/2026-06-06-dynamo-amazon-key-value-store/ Sat, 06 Jun 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/06/papers/2026-06-06-dynamo-amazon-key-value-store/ DeCandia et al.'s SOSP 2007 paper introduced a system that traded strong consistency for availability under partition — and codified the design vocabulary (consistent hashing, vector clocks, sloppy quorums, hinted handoff, read repair) that almost every modern NoSQL store still uses. https://sparsenotes.com/posts/2026/06/papers/2026-06-06-smt-rnn-without-recurrence/ Sat, 06 Jun 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/06/papers/2026-06-06-smt-rnn-without-recurrence/ Supervised Memory Training (SMT) sidesteps BPTT by reducing RNN training to one-step memory-transition labels supplied by a Transformer-based predictive-state encoder — recovering O(1) gradient paths and time-parallel training. https://sparsenotes.com/posts/2026/06/papers/ Sat, 06 Jun 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/06/papers/ Three blog-length entries from this week's CS paper digest: the Dynamo paper as Seminal pick, plus top picks on training nonlinear RNNs without BPTT and on cross-layer sparse attention for long-context LLMs. https://sparsenotes.com/posts/2026/06/papers/2026-06-06-clsa-cross-layer-sparse-attention/ Sat, 06 Jun 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/06/papers/2026-06-06-clsa-cross-layer-sparse-attention/ CLSA shares the top-k routing index across cross-decoder layers in a YOCO-style KV-sharing architecture, amortizing token-sparse routing cost to deliver 7.6× decoding and 17.1× throughput at 128K context.