Retrieval-augmented generation

RECOMP

RECOMP: Improving Retrieval-Augmented LMs with Compression and Selective Augmentation

Superseded baseline#16 of 1,179 most-superseded · first seen Oct 6, 2023

Superseded — cited as a baseline and beaten by newer methods

6 papers critique it · 12 beat it on benchmarks

What papers say

Verbatim critique sentences, each from a paper that cites RECOMP 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
A critical limitation of these methods is their dependence on fixed compression ratios.
AttnComp: Attention-Guided Adaptive Context Compression for Retrieval-Augmented Generation
their approach still requires additional models to summarise the retrieved documents
Eliciting Critical Reasoning in Retrieval-Augmented Language Models via Contrastive Explanations
Prior context pruning approaches such as RECOMP~xu2023recomp or DSLR~hwang2024dslr encode sentences in a passage independently of each other. In contrast, Provence encodes all the sentences in a retrieved passage together with a query, in a single reranker forward pass.
XProvence: Zero-Cost Multilingual Context Pruning for Retrieval-Augmented Generation
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
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 RECOMP. Values are copied from the source paper's tables — verify against the cited paper.

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Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.