ICML Paper Reveals LLM Memory Bottleneck: 95% Facts Learned, 30% Unretrievable

新智元 · wechat · 2026-08-14

Google's ICML 2026 paper Empty Shelves or Lost Keys? reveals that frontier LLMs like GPT-5 and Gemini 3 store 95%–98% of facts in their parameters, yet fail to recall 26%–34% when directly queried.

The study introduces a 'Knowledge Profile' framework categorizing fact states into five types. Errors often stem not from a lack of learning, but from retrieval failures, which are particularly prevalent with obscure facts and reverse queries.

Notably, enabling 'thinking' (chain-of-thought) recovers 40%–65% of these inaccessible facts, mirroring human 'spreading activation.' The research concludes that simply scaling parameters primarily improves storage, and future gains must come from better post-training and inference-time retrieval strategies.

Original post →

More from Models

Models channel →