Is LLMLingua superseded?

LLMLingua (Retrieval-augmented generation): superseded — cited as a baseline and beaten by newer methods. 4 paper(s) critique it, 6 beat it on benchmarks — #25 of 1179 most-superseded.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.

Method Drift›Retrieval-augmented generation

Superseded baseline#25 of 1,179 most-superseded

LLMLingua

LLMLingua: Compressing Prompts for Accelerated Inference of Large Language Models

Retrieval-augmented generation · first seen Oct 9, 2023

superseded — cited as a baseline and beaten by newer methods

4 papers critique it · 6 beat it on benchmarks

What papers say

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

“they rely exclusively on post-retrieval context compression without improving initial retrieval quality, creating an inherent performance ceiling”
— MacRAG: Compress, Slice, and Scale-up for Multi-Scale Adaptive Context RAG
“CASC consistently outperforms strong baselines, including standard Top-K RAG, and existing context compression methods like RECOMP fangyuan2024recomp and LLMLingua huiqiang2023llmlin, across various Reader LLM backbones”
— Context-Adaptive Synthesis and Compression for Enhanced Retrieval-Augmented Generation in Complex Domains
“However, these methods lack context-awareness, making it challenging to determine the optimal compression ratio for a given LLM, resulting in context redundancy or over-compression.”
— AttentionRAG: Attention-Guided Context Pruning in Retrieval-Augmented Generation
“While explicit methods are generally agnostic to downstream LLMs and therefore more practical, they may suffer from higher compression loss due to the over-removal of input tokens.”
— Lighter And Better: Towards Flexible Context Adaptation For Retrieval Augmented Generation

Beaten on benchmarks

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

Oreo beats LLMLingua · EM [NaturalQuestions single-hop]
0.4413 vs 0.4125
Oreo: A Plug-in Context Reconstructor to Enhance Retrieval-Augmented Generation
CASC beats LLMLingua · F1-score [Llama-3-70B]
65.15 vs 56.88
Context-Adaptive Synthesis and Compression for Enhanced Retrieval-Augmented Generation in Complex Domains
CASC beats LLMLingua · F1-score [Llama-3-8B]
56.12 vs 54.30
Context-Adaptive Synthesis and Compression for Enhanced Retrieval-Augmented Generation in Complex Domains
CASC beats LLMLingua · F1-score [GPT-4o]
65.80 vs 54.80
Context-Adaptive Synthesis and Compression for Enhanced Retrieval-Augmented Generation in Complex Domains
SARA beats LLMLingua · F1 [512 tokens]
36.23 vs 31.29
SARA: Selective and Adaptive Retrieval-augmented Generation with Context Compression
SARA beats LLMLingua · ROUGE-L [512 tokens]
39.17 vs 32.38
SARA: Selective and Adaptive Retrieval-augmented Generation with Context Compression
SARA beats LLMLingua · F1 [1024 tokens]
40.37 vs 33.18
SARA: Selective and Adaptive Retrieval-augmented Generation with Context Compression
SARA beats LLMLingua · ROUGE-L [1024 tokens]
42.24 vs 32.19
SARA: Selective and Adaptive Retrieval-augmented Generation with Context Compression
K-comp beats LLMLingua · BertScore [MedQuAD, With compressor, Llama-3-8B]
83.79 vs 76.45
K-COMP: Retrieval-Augmented Medical Domain Question Answering With Knowledge-Injected Compressor
K-comp beats LLMLingua · UniEval [MedQuAD, With compressor, Llama-3-8B]
64.05 vs 54.65
K-COMP: Retrieval-Augmented Medical Domain Question Answering With Knowledge-Injected Compressor
K-comp beats LLMLingua · BertScore [MedQuAD, With compressor, Llama-3-70B]
82.06 vs 78.78
K-COMP: Retrieval-Augmented Medical Domain Question Answering With Knowledge-Injected Compressor
K-comp beats LLMLingua · BertScore [MedQuAD, With compressor, Mixtral-8x7B]
80.64 vs 59.79
K-COMP: Retrieval-Augmented Medical Domain Question Answering With Knowledge-Injected Compressor

Newer alternatives

Recent methods in the same sub-problem, not yet superseded in the knowledge base.