Engram: RL Teaches LLMs to Use Notes, Cheap to Train and Robust Across Millions of Token Edits
scott_linderman · x · 2026-11-05
Research company Engram published a new blog on combining RL with text-based memory: RL lets models learn how to use their notes, with training surprisingly cheap compared to building the notes, and remaining robust even through millions of tokens of edits. A memorable analogy: the LLM equivalent of rereading discrete math notes until they finally make sense.
Related event: Engram: $373 RL run teaches models to use their notes(7 posts)→
More from Research
- LongTails: 25+ new optimization results with Lean 4-verified proofs, improving known complexity bounds — IgorCarron · 2026-11-05
- The Economist's Babbage podcast asks how science should be reimagined for the AI age — james_y_zou · 2026-11-05
- NYU and Meta FAIR introduce ID-WM, a hierarchical world model beating baselines by 38.5% — ylecun · 2026-11-05
- Guidelight ships Capability Testing standard v1.0 with sandbagging-detection practice — dfrsrchtwts · 2026-11-05
- mjorbit: open-source space robotics simulator runs thousands of parallel worlds on one GPU — ChongZzZhang · 2026-11-05
- CMU paper unifies pretraining, SFT and RL into one likelihood-based algorithm — burny_tech · 2026-11-05