LLM Digest

Story

arxiv_cs_ai · Oct 30, 2026 · paper

Source brief

Dynamical Memory Transformer: Selective Recurrent Memories with Constant Cost Attention

arxiv.orgOct 30, 2026
original source linked

In brief

Transformers retain an explicit history of the sequence, allowing attention to directly retrieve past representations. However, the size of the KV-Cache and the cost of a new token increase with the context length. State Space Models compress the past into a fixed-size recurrent

Read the original at arxiv.org →Open in live feed Read that day’s brief

Dynamical Memory Transformer: Selective Recurrent Memories with Constant Cost Attention

Earlier in this thread 4 items

Fewer Heads, Fewer Memories: Pruning Redundant Attention Heads Suppresses Verbatim Memorisation

Smarter cost controls for serverless

CopyLoop: Fast Looped Transformer Decoding via Context Copying and Shallow Screening

Causal Attention on CPU: A Cache-Aware and Load-Balanced Scheduling Kernel for Prefill