5 papers

Retrieval & RAG

Grounding generation in external knowledge, from classic RAG to reasoning retrieval.

Papers
5
Read time
1.4h
Avg length
17m
01
Retrieval & RAG

DeepRAG: Thinking to Retrieve Step by Step for Large Language Models

DeepRAG teaches an LLM to break a hard question into sub-questions and, for each one, decide *on its own* whether to hit a retriever or answer from memory — by treating the whole thing as a Markov Decision Process and training the model with tree search + imitation learning + preference calibration, getting ~21-26% more accurate answers while retrieving *less*.

GUAN, ZENG, MENG, XIN, LU, LIN, HAN, SUN, ZHOU (CAS & TENCENT WECHAT AI)
Read
18 min
Source
arXiv:2502.01142
02
Retrieval & RAG

Hybrid Retrieval-Augmented Generation (RAG) Systems with Embedding Vector Databases

Don't make your RAG system choose between semantic vector search and old-school keyword search — run both, fuse the rankings with query-aware weighting, and you cut hallucinations while gaining 15-35% retrieval recall over either method alone.

SARAT KIRAN (UTAH STATE UNIVERSITY) · IJSRCSEIT V11(2), MARCH 2025 · DOI:10.32628/CSEIT25112702
Read
16 min
03
Retrieval & RAG

Modular RAG: Transforming RAG Systems into LEGO-like Reconfigurable Frameworks

Stop thinking of RAG as "retrieve then generate" and start thinking of it as a graph of swappable LEGO bricks — modules, sub-modules, and operators — that you wire together into routing, branching, and looping flows to fit each use case.

YUNFAN GAO, YUN XIONG, MENG WANG, HAOFEN WANG (TONGJI & FUDAN UNIVERSITY) · ARXIV 2024 · 2407.21059V1
Read
22 min
04
Retrieval & RAG

MoR: Better Handling Diverse Queries with a Mixture of Sparse, Dense, and Human Retrievers

Instead of picking one retriever for your RAG system, run many small cheap ones and let zero-shot geometry signals decide, per query, how much to trust each — and eight tiny models combined beat a 7B retriever while costing a fraction of the params.

JUSHAAN KALRA, XINRAN ZHAO, TO EUN KIM, FENGYU CAI, FERNANDO DIAZ, TONGSHUANG WU · CARNEGIE MELLON UNIVERSITY & TU DARMSTADT
Read
16 min
Source
arXiv:2506.15862
05
Retrieval & RAG

RankCoT: Refining Knowledge for Retrieval-Augmented Generation through Ranking Chain-of-Thoughts

Instead of separately reranking documents *or* summarizing them, RankCoT trains a single LLM to write a short query-focused "reasoning note" that implicitly ranks and distills the retrieved pile — and it learns to do this from preference pairs it generates and grades against itself.

MINGYAN WU, ZHENGHAO LIU, YUKUN YAN, XINZE LI, SHI YU, ZHENI ZENG, YU GU, GE YU (NORTHEASTERN UNIVERSITY, TSINGHUA UNIVERSITY) · ARXIV 2025 · 2502.17888V1
Read
14 min