Fetching from the wire…
Research2026-09-26 · source-backed
arXiv 2609.28870 replayed production agentic traces from two companies against 14 algorithms in both HBM-constrained and large-pool settings. The sophisticated policies give little over LRU despite a large gap to Belady, and the cause is structural: active sessions resend growing context at a regular pace, making recency unusually predictive. Their recommendations are keeping LRU as the base and adding quick demotion for one-hit prefixes, compute-aware partial eviction for expensive misses, and capacity-dependent granularity. Traces and simulator are promised.
Each link below shares sources, entities, or timing with this story.
arXiv 2607.24667 recasts eviction as estimation of a hidden reuse signal along a commit-lag axis, with StreamingLLM/H2O/SnapKV at lag 0 and Belady's optimum at full future knowledge, then fills the middle: wait a bounded number of steps, observe what a correct near-future pred...
agentic-kv-cache simulates cross-request prefix caching against real traces, not synthetic ones: 68,266 requests across 393 Claude Code sessions at 64-token blocks, plus Mooncake traces at 512-token blocks. It models prefix-contiguous hits, radix eviction constraints and pinne...
Reports surfacing August 7 say Samsung, SK Hynix and Micron have sold their entire 2027 allocation for both DRAM and HBM. Consumer impact is already visible: 32GB DDR5 kits are well over $400 versus roughly $100 in September 2025, a 4x move after DRAM contract prices jumped 50...
The approach recasts expert-activation prediction as sequence-to-sequence modeling for multi-step multi-layer forecasts, then treats prefetching as job sequencing with deadlines and adds probabilistic Belady eviction. At 45% residency it averages a 96.97% hit rate. The offload...
An agent proposes changes to a training pipeline, runs it, and keeps edits improving a verifiable in-loop metric. Looks like reliable progress. The authors name algorithmic mode collapse: surface edit diversity stays stable while semantic and mechanism-level diversity collapse...
This one landed sideways on a belief I have been operating on for months. MemTrapBench (arXiv 2608.20202, submitted August 20, from a Zhejiang-affiliated team led by Mengru Wang and Ningyu Zhang) tests something the memory-layer boom has mostly assumed away: whether *correct*...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.