CorpusQA
A research paper examining the deep difficulties that conventional RAG architectures face when the evidence required to answer a query is scattered across an entire document collection rather than concentrated in a few passages.
Timeline (last 90 days)
CorpusQA is a 2026 study (the paper "CorpusQA: A 10 Million Token Benchmark for Corpus-Level Analysis and Reasoning"), published by Zhiyuan Lu and other authors on January 21, 2026 and revised on April 25, 2026, containing finance, education, and real estate document tasks.
In the CorpusQA evaluation, Gemini 2.5 Pro achieved an overall accuracy of 80.19% at a 128K context length and 50.73% at a 1M context length, a drop of 29.46 percentage points.
In the 1M-token CorpusQA evaluation, standard RAG achieved an accuracy of 17.38%, while the Memory Agent approach using Gemini 2.5 Pro reached 44.95%.
All Facts (3)
CorpusQA is a 2026 study (the paper "CorpusQA: A 10 Million Token Benchmark for Corpus-Level Analysis and Reasoning"), published by Zhiyuan Lu and other authors on January 21, 2026 and revised on April 25, 2026, containing finance, education, and real estate document tasks.
50%UnverifiedIn the CorpusQA evaluation, Gemini 2.5 Pro achieved an overall accuracy of 80.19% at a 128K context length and 50.73% at a 1M context length, a drop of 29.46 percentage points.
50%UnverifiedIn the 1M-token CorpusQA evaluation, standard RAG achieved an accuracy of 17.38%, while the Memory Agent approach using Gemini 2.5 Pro reached 44.95%.
50%