Reproduce “A Human-Inspired Reading Agent with Gist Memory of Very Long Contexts” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kuang-Huei Lee, Xinyun Chen, Hiroki Furuta, John Canny, Ian Fischer
Google DeepMind's ReadAgent introduces an LLM agent that leverages human-inspired "gist memory" and interactive lookup mechanisms to efficiently process and reason over extremely long documents. It extended effective context length by up to 20x on NarrativeQA and consistently outperformed RAG and full-text baselines on QuALITY, NarrativeQA, and QMSum.
Key insights
- Modeling human cognitive processes, specifically gist memory and interactive lookup, can significantly extend the effective context length and improve LLM performance on long documents.
- LLM-driven "intelligent" pagination and context-aware retrieval (deciding what to look up) are more effective than uniform chunking or external retrieval mechanisms.
- Operating primarily on compressed gist memories and retrieving raw details only when necessary can reduce overall token consumption and mitigate the "lost in the middle" problem in LLMs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.