7 papers I've read closely, with my one-line take. Credit for every idea goes to the authors.
Wan, Jünger, Du, Liu, Krishnan, Weinberger · 2026
Drop the retriever and let a decision model walk the corpus tree. The cleanest argument yet that RAG's weak link is the first stage.
read my post →Günther, Mohr, Williams, Wang, Xiao · 2024
Embed the whole document first, then pool per chunk. A tiny change that fixes chunks losing their pronouns.
read my post →Packer, Wooders, Lin, Fang, Patil, Stoica, Gonzalez · 2023
Treat the context window like RAM and let the agent page memory in and out. The framing most agent memory work still borrows.
read my post →Park, O'Brien, Cai, Morris, Liang, Bernstein · 2023
Recency, importance and relevance as a retrieval score, plus reflection. Still the best memory baseline to argue against.
read my post →Zheng, Yin, Xie, Sun, Huang, Yu, Cao, Kozyrakis, Stoica, Gonzalez, Barrett, Sheng · 2023
RadixAttention makes prefix reuse automatic. Read it to see why a stable prompt prefix is worth designing for.
read my post →Islam, Kannappan, Kiela, Qian, Scherrer, Vidgen · 2023
Real 10-K questions where vector RAG struggles. The benchmark behind most PageIndex claims.
read my post →Khattab, Zaharia · 2020
One vector per token, matched late. The other meaning of "late" in retrieval, and worth knowing next to late chunking.
read my post →