Could Writable Ngram Embedding Tables Enable Inference-Time Learning? A Reddit Proposal

RapidRaid · reddit · 2026-09-12

A Reddit poster proposes adding a writable "live" table on top of pretrained read-only ngram embeddings (as seen in models like Qwen and DeepSeek): new knowledge writes new embeddings, frequent entries persist, unused ones decay — mimicking memory formation and forgetting during inference. Open discussion, no implementation yet.

Original post →

More from Research

Research channel →