BM25 (Retrieval-augmented generation): superseded — cited as a baseline and beaten by newer methods. 6 paper(s) critique it, 16 beat it on benchmarks — #10 of 1179 most-superseded. Sub-problem: cluster led by BM25. Newer alternatives in the same sub-problem include Beyond Topical Similarity, Experience-RAG Skill, LFRAG, Don't Retrieve, Navigate, MIGRASCOPE.

Is BM25 superseded? Critiques, benchmarks & alternatives

What papers say

Verbatim critique sentences, each from a paper that cites BM25 as a baseline.

This paradigm inevitably discards critical visual and structural information, including the row-column relationships of tables, data trends of charts, and layout logic between text and figures.

— LFRAG: Layout-oriented Fine-grained Retrieval-Augmented Generation on Multimodal Document Understanding

“The semantic and lexical gaps limit the effectiveness of sparse models like BM25 and TF-IDF, which operate on keyword matching”

— ModernBERT + ColBERT: Enhancing biomedical RAG through an advanced re-ranking retriever

“Despite their widespread use, many RAG systems rely on static, off-the-shelf retrieval modules — e.g., BM25 ... that are minimally adapted to the downstream task or domain.”

— Test-time Corpus Feedback: From Retrieval to RAG

“We prioritized a dense retrieval approach over sparse methods (such as BM25) because layperson queries often lack the precise legal terminology found in regulatory documents.”

— Benchmarking Large Language Models for Quebec Insurance: From Closed-Book to Retrieval-Augmented Generation

“They offer computational efficiency but lack deeper semantic comprehension.”

— MegaRAG: Multimodal Knowledge Graph-Based Retrieval Augmented Generation

“However, many practical systems still assume that one fixed retrieval strategy is sufficient across all tasks. This assumption is problematic in realistic agent settings.”

— An Agent-Oriented Pluggable Experience-RAG Skill for Experience-Driven Retrieval Strategy Orchestration

Beaten on benchmarks

Head-to-head results where a newer method reports beating BM25. Values are copied from the source paper's tables — verify against the cited paper.

ROGRAG beats BM25
0.38 vs 0.03
Rouge · [SeedBench QA-4 Rouge]
ROGRAG: A Robustly Optimized GraphRAG Framework
GraphRAFT beats BM25
63.71 vs 12.75
Hit@1 · [STARK-PRIME]
GraphRAFT: Retrieval Augmented Fine-Tuning for Knowledge Graphs on Graph Databases
KG-Retriever beats BM25
0.233 vs 0.069
BLEU · [CRUD-QA2]
KG-Retriever: Efficient Knowledge Indexing for Retrieval-Augmented Large Language Models
BookRAG beats BM25
43.8 vs 18.3
Exact Match · [MMLongBench]
BookRAG: A Hierarchical Structure-aware Index-based Approach for Retrieval-Augmented Generation on Complex Documents
G-threshold beats BM25
0.944 vs 0.420
Recall@1 · [TriviaQA dataset]
Revisiting RAG Retrievers: An Information Theoretic Benchmark
HippoRAG beats BM25
0.833 vs 0.415
Recall@1 · [MuSiQue dataset]
Revisiting RAG Retrievers: An Information Theoretic Benchmark
G-naive beats BM25
0.957 vs 0.486
Recall@1 · [2Wiki dataset]
Revisiting RAG Retrievers: An Information Theoretic Benchmark
FairFilter-BM25 beats BM25
0.056 vs 0.407
Overall Fairness · [Llama3.2-1b with BM25]
The Other Side of the Coin: Exploring Fairness in Retrieval-Augmented Generation
Corpus2Skill beats BM25
0.708 vs 0.406
Context Recall · [WixQA benchmark]
Don't Retrieve, Navigate: Distilling Enterprise Knowledge into Navigable Agent Skills for QA and RAG
G-window beats BM25
0.110 vs 0.252
Divergence · [2Wiki dataset]
Revisiting RAG Retrievers: An Information Theoretic Benchmark
ft-DPR beats BM25
63.1 vs 48.3
R@1K · [512 tokens context (or N/A for sparse)]
CLERC: A Dataset for Legal Case Retrieval and Retrieval-Augmented Analysis Generation
QMKGF beats BM25
47.42 vs 38.53
R-1 · [MuSiQue]
A Query-Aware Multi-Path Knowledge Graph Fusion Approach for Enhancing Retrieval-Augmented Generation in Large Language Models

What to use instead

Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.