Research map / Large language models / Retrieval and knowledge
OSCAR: Online Soft Compression for RAG
Maxime Louis, Thibault Formal, Hervé Déjean, Stéphane Clinchant
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| make compression query-dependent (ablation) | OSCAR-llama (x16) · RAG QA (5 QA sets, LLM eval) | average LLM-eval score: 0.73 -> 0.77 | Improved |
| make compression query-dependent (ablation) | OSCAR-llama (x128) · RAG QA (5 QA sets, LLM eval) | average LLM-eval score: 0.71 -> 0.75 | Improved |
| add compressor pretraining (ablation) | OSCAR-llama · RAG QA (5 QA sets, LLM eval) | average LLM-eval score: 0.70 -> 0.77 | Improved |
| train the generator jointly instead of freezing it (ablation) | OSCAR-llama · RAG QA (5 QA sets, LLM eval) | average LLM-eval score: 0.71 -> 0.77 | Improved |
| increase compression rate from x16 to x128 | OSCAR-llama · RAG QA (5 QA sets, LLM eval) | average LLM-eval score: 0.77 -> 0.75 | Hurt |
| use Llama-1B compressor instead of DeBERTa-v3 (ablation) | OSCAR · RAG QA (5 QA sets, LLM eval) | average LLM-eval score: 0.73 -> 0.77 | Improved |
| use Llama-1B compressor instead of Modern-bert-large (ablation) | OSCAR · RAG QA (5 QA sets, LLM eval) | average LLM-eval score: 0.76 -> 0.77 | Improved |
| switch retrieval from SPLADE+DeBERTa to BM25 | OSCAR-llama vs no compression · RAG QA (5 QA sets, LLM eval) | average LLM-eval score: 0.58 vs 0.56 | No clear effect |
| use OSCAR-llama instead of no compression | Mistral-24B · RAG QA, 10 docs | average accuracy: 0.68 -> 0.69 | Improved |
| increase compressor depth from 8 to 10 layers (ablation) | OSCAR-N-Layers · BEIR | nDCG@10: 52.5 -> 54.3 | Improved |
| increase compressor depth from 10 to 16 layers (ablation) | OSCAR-N-Layers · BEIR | nDCG@10: 54.3 -> 55.3 | Improved |
| use e2e reranking by OSCAR itself instead of DeBERTa reranker | OSCAR-llama · RAG QA (6 QA sets, LLM eval) | average LLM-eval score: 0.77 -> 0.77 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Provence — Provence: efficient and robust context pruning for retrieval-augmented generation (ICLR 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Accelerating Inference of Retrieval-Augmented Generation via Sparse Context Selection (ICLR 2025 · similarity 0.89)
- Long-Context LLMs Meet RAG: Overcoming Challenges for Long Inputs in RAG (ICLR 2025 · similarity 0.89)
- Benchmarking Large Language Models in Retrieval-Augmented Generation (AAAI 2024 · similarity 0.88)
- Speculative RAG: Enhancing Retrieval Augmented Generation through Drafting (ICLR 2025 · similarity 0.88)
- RAGGED: Towards Informed Design of Scalable and Stable RAG Systems (ICML 2025 · similarity 0.88)
- Inference Scaling for Long-Context Retrieval Augmented Generation (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
