Researchers let a 1B model read in chunks and extend its context with memory

alex_verem · x · 2026-07-25

• Researchers from MBZUAI, Princeton, and the Weizmann Institute redesigned how a 1B model reads long text.
• Instead of attending to everything at once, it processes 1,024-token chunks and carries forward important information through associative memory.
• The setup is explicitly inspired by human memory: chunk-level full attention acts like short-term memory, while the cross-chunk module acts like long-term memory.
• The model is trained progressively, starting with short/easy text and gradually increasing difficulty.
• The approach reportedly cuts the GPU-memory burden on long documents and avoids the sharp collapse seen in the baseline model beyond its limit.

Original post →

More from Research

• AI slop is already clogging PR review and weakening the credit system behind science — rbhar90 · 2026-07-27
• ICML 2026 oral paper replication scores stay middling after a stricter re-scoring — profjamesevans · 2026-07-27
• Long-running agents will need immutable event logs, this thread argues — sebpaquet · 2026-07-27
• Seed IQ navigates Doom II, prompting questions about benchmarks beyond ARC-AGI — Fit_Transition8824 · 2026-07-27
• Agentic Data Science in Practice: Agents Write Code but Answer Wrong Questions — hugobowne · 2026-07-27
• A concise canon of foundational papers in ML, systems, NLP, speech, and audio — deliprao · 2026-07-27

Research channel →