https://sparsenotes.com/posts/2026/11/papers/ Recent content in Weekly CS Paper Digest — 7 – 13 November 2026 on Sparse Notes https://sparsenotes.com/images/og-default.png https://sparsenotes.com/images/og-default.png Hugo -- gohugo.io Sat, 14 Nov 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/11/papers/access-path-selection-selinger-1979/ Sat, 14 Nov 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/11/papers/access-path-selection-selinger-1979/ Seminal paper of the week: the System R optimizer paper that made SQL practical. Pat Selinger and colleagues combined a cost formula, selectivity estimates from catalog statistics, and dynamic programming over join orders with 'interesting orders'. Nearly every relational optimizer since has used the same structure. https://sparsenotes.com/posts/2026/11/papers/raven-unified-file-management-lakehouse/ Sat, 14 Nov 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/11/papers/raven-unified-file-management-lakehouse/ University of Maryland and Dolby put a lakehouse catalog's transactional metadata (which files are visible, who references them) in Postgres and leave the bulky per-file statistics in Iceberg or Delta Lake. That one split gives cheaper small commits, multi-table snapshot-isolation transactions, and reachability-based garbage collection that won't delete files a clone or a running job still needs. https://sparsenotes.com/posts/2026/11/papers/unbundling-multimodal-vision-encoding/ Sat, 14 Nov 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/11/papers/unbundling-multimodal-vision-encoding/ Meta FAIR and NYU split the vision encoder into three separate design choices (a pretrained prior, dedicated vision parameters, and placement before fusion) and test each one with every model trained from scratch at matched compute. Only the dedicated parameters matter, and they work just as well as vision-only branches inside the first eight layers of the language model. https://sparsenotes.com/posts/2026/11/papers/eliminating-servers-state-machine-replication/ Sat, 07 Nov 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/11/papers/eliminating-servers-state-machine-replication/ UIUC and Wisconsin, with shared-log veteran Mahesh Balakrishnan, push state machine replication one step past disaggregation: the command log and a single materialized copy of state both live on blob storage, and clients cooperate to compact it. Their AirDB key-value store survives every server failure by having no servers, cold-starts 10–45× faster than server-based SMR, and costs about 22× less at low load. https://sparsenotes.com/posts/2026/11/papers/hinted-scaling-laws/ Sat, 07 Nov 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/11/papers/hinted-scaling-laws/ Stanford's fix for downstream scaling laws that break when small models score zero: show the model part of a reference solution, measure how accuracy rises with the hint, and fit one law over scale and hint level. Fit on models up to 10^20 FLOPs, it forecasts a 10^23 FLOP model's GSM8K score within 3 points, and it cuts median forecast error from 21 points to 4.3. https://sparsenotes.com/posts/2026/11/papers/backpropagation-rumelhart-hinton-williams-1986/ Sat, 07 Nov 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/11/papers/backpropagation-rumelhart-hinton-williams-1986/ Seminal paper of the week: a four-page Nature letter showed that gradient descent through hidden layers, computed with one forward and one backward pass, makes networks invent their own internal features. The 1986 paper didn't invent the chain rule or reverse-mode differentiation, but it showed what they could do and set off the field that became deep learning.