A few-MB overlay can inject new facts into conditional-memory LLMs like DeepSeek
teortaxesTex · x · 2026-10-09
teortaxesTex notes that models with conditional memory (n-gram lookup) architectures, such as DeepSeek's, can have their n-gram blocks edited via a cheaply derived overlay of just a few MBs to introduce new declarative facts. The method is primitive—limited to [subject, relation, object] triples—but enables per-user continual learning.
Related event: EngramEdit Enables Decoupled Fact Editing in LLMs via Conditional Memory(3 posts)→
More from Research
- InfiLoop residuals let a 7M looped model improve past 20,000 test-time steps, 97.9% on Sudoku — Pengxiang Li · 2026-10-09
- TokenRouter Serving System Boosts Token-Level LLM Routing Throughput up to 64x — nics-efc · 2026-10-09
- DreamTrue robot world model cuts interaction defect rate from 48.12% to 6.25%, tops AgiBot challenge — Junyan Li · 2026-10-09
- SpatialOPSD self-distills coding-agent traces so MLLMs reason spatially without tools — ATH-MaaS · 2026-10-09
- NTU's Trace2Env Builds Agent Environment Simulations from Historical Traces — nanyang-technological-university-singapore · 2026-10-09
- RISEBench++: 65 reasoning-based visual editing tasks; best model GPT-Image-2.5 hits only 56.6% — VisionXLab · 2026-10-09