Is LongLLMLingua superseded?

LongLLMLingua (Retrieval-augmented generation): superseded — cited as a baseline and beaten by newer methods. 3 paper(s) critique it, 12 beat it on benchmarks — #18 of 1179 most-superseded. Sub-problem: cluster led by RECOMP. Newer alternatives in the same sub-problem include Bottleneck Attention Intervention for Recovery (BAIR), QREAM, CoCR-RAG, XProvence, Crucible.

Retrieval-augmented generation

LongLLMLingua

LongLLMLingua: Accelerating and Enhancing LLMs in Long Context Scenarios via Prompt Compression

Superseded baseline#18 of 1,179 most-superseded · first seen Oct 10, 2023

Superseded — cited as a baseline and beaten by newer methods

3 papers critique it · 12 beat it on benchmarks

What papers say

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

While some methods, such as perplexity-based approaches, can process longer texts, they often lack comprehensive document understanding.

— Hierarchical Document Refinement for Long-context Retrieval-augmented Generation

“While these methods improve textual RAG, they do not address the fundamental cross-modal attention collapse that occurs in multimodal architectures.”

— The Cost of Context: Mitigating Textual Bias in Multimodal Retrieval-Augmented Generation

“the pre-processing methods introduce additional computational costs during inference and may lead to the loss of essential information.”

— R^2AG: Incorporating Retrieval Information into Retrieval Augmented Generation

Beaten on benchmarks

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

R²AG beats LongLLMLingua
0.6930 vs 0.3635
NQ-10 Acc · [Frozen LLMs]
R^2AG: Incorporating Retrieval Information into Retrieval Augmented Generation
AttentionRAG beats LongLLMLingua
0.53 vs 0.28
EM · [Qwen2.5-7B-Instruct / HotpotQA]
AttentionRAG: Attention-Guided Context Pruning in Retrieval-Augmented Generation
FlexRAG w. SC. beats LongLLMLingua
29.93 vs 17.82
Average (LMQA) · [8x compression ratio]
Lighter And Better: Towards Flexible Context Adaptation For Retrieval Augmented Generation
AttnComp beats LongLLMLingua
19.6 vs 13.8
Acc · [MuSiQue]
AttnComp: Attention-Guided Adaptive Context Compression for Retrieval-Augmented Generation
SARA beats LongLLMLingua
54.90 vs 39.27
ROUGE-L · [512 tokens]
SARA: Selective and Adaptive Retrieval-augmented Generation with Context Compression
QREAM-FT beats LongLLMLingua
45.6 vs 36.3
Accuracy · [Standard RAG Pipeline with Llama-3-8B-Instruct]
Align Documents to Questions: Question-Oriented Document Rewriting for Retrieval-Augmented Generation
LongRefiner beats LongLLMLingua
50.6 vs 40.4
Accuracy · [2k token limit, Qwen2.5-7B-Instruct]
Hierarchical Document Refinement for Long-context Retrieval-augmented Generation
CORE beats LongLLMLingua
36.72 vs 29.37
EM · [Compression of top 5 documents on 2WikiMultihopQA]
CORE-RAG: Lossless Compression for Retrieval-Augmented LLMs via Reinforcement Learning
Self-Correcting RAG beats LongLLMLingua
37.1 vs 31.6
Exact Match (EM) · [Average across all QA benchmarks]
Self-Correcting RAG: Enhancing Faithfulness via MMKP Context Selection and NLI-Guided MCTS
SKR beats LongLLMLingua
43.61 vs 40.05
Avg. · [Llama-2-7B, Top-5]
Supportiveness-based Knowledge Rewriting for Retrieval-augmented Language Modeling
LongLLMLingua + BAIR beats LongLLMLingua
87.45 vs 84.36
Acc. · [Qwen2.5-VL-3B (FACET)]
The Cost of Context: Mitigating Textual Bias in Multimodal Retrieval-Augmented Generation

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.