mello-no-weight-modification
IN premise — summaries/2026/08/24/zhong-2023-mquake-s1-introduction.md
Created 2026-08-25T02:59:05+00:00
MeLLo keeps the LLM frozen and stores edited facts in an external memory, using iterative sub-question decomposition and consistency checking at inference time without any gradient updates
Summary
MeLLo corrects what a language model "knows" without ever retraining or tweaking the model itself; it simply writes the updated fact into a separate memory bank and, at question time, breaks the query into smaller pieces and cross-checks them for consistency. The practical upshot is that fact edits are cheap, reversible, and don't risk degrading the model's general abilities, but the tradeoff is a heavier, multi-step reasoning process every single time you ask it a question.