Research map / Large language models / Retrieval and knowledge
Atlas: Few-shot Learning with Retrieval Augmented Language Models
Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane Dwivedi-Yu, Armand Joulin, Sebastian Riedel, Edouard Grave
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 |
|---|---|---|---|
| add joint pre-training of retriever and reader (ablation) | Atlas · KILT 64-shot (NQ/WoW/FEVER) | average: 30.0 -> 42.2 | Improved |
| train the retriever during pre-training (Perplexity Distillation) instead of fixed (ablation) | Atlas · KILT 64-shot avg | average: 42.2 -> 45.7 | Improved |
| train the retriever during pre-training instead of fixed (ablation) | Atlas · KILT 1024-shot avg | average: 51.1 -> 51.0 | No clear effect |
| use masked language modelling instead of prefix language modelling pretext task (ablation) | Atlas · KILT 64-shot avg | average: 40.1 -> 42.4 | Improved |
| use Wikipedia index instead of Common Crawl index with Wikipedia training data (ablation) | Atlas · KILT 64-shot avg | average: 39.8 -> 42.4 | Improved |
| fine-tune retriever query-side only instead of fixed retriever (ablation) | Atlas · KILT 64-shot avg | average: 41.1 -> 45.7 | Improved |
| full fine-tuning instead of query-side fine-tuning (ablation) | Atlas · KILT 1024-shot NQ | exact match: 44.9 -> 47.0 | Improved |
| top-100 re-ranking instead of full index update (ablation) | Atlas · KILT 64-shot avg | average: 44.1 -> 44.7 | No clear effect |
| use de-biased cyclic-permutation inference (ablation) | Atlas-11B · MMLU zero-shot | accuracy: 36.8 -> 47.1 | Improved |
| use de-biased cyclic-permutation inference (ablation) | Atlas-11B · MMLU 5-shot (multi-task) | accuracy: 56.4 -> 56.6 | No clear effect |
| use retrieval (Atlas) instead of closed-book T5 | T5 · MMLU 5-shot, 11B | accuracy: 36.1 -> 43.4 | Improved |
| use temporally matched Wikipedia index (ablation) | Atlas · NaturalQuestions 64-shot | exact match: 41.3 -> 45.1 | Improved |
Later papers that compared with this method
3 later accepted papers report Atlas base as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- BTR: Binary Token Representations for Efficient Retrieval Augmented Language Models (ICLR 2024 · better on 5 of 10 shared results)
- RA-DIT: Retrieval-Augmented Dual Instruction Tuning (ICLR 2024 · better on 1 of 1 shared result)
- Knowledge-in-Context: Towards Knowledgeable Semi-Parametric Language Models (ICLR 2023 · better on 0 of 1 shared result)
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
- Making Retrieval-Augmented Language Models Robust to Irrelevant Context (ICLR 2024 · similarity 0.87)
- Few-Shot Recognition via Stage-Wise Retrieval-Augmented Finetuning (CVPR 2025 · similarity 0.87)
- Retrieval meets Long Context Large Language Models (ICLR 2024 · similarity 0.86)
- The Surprising Effectiveness of Test-Time Training for Few-Shot Learning (ICML 2025 · similarity 0.86)
- Physics of Language Models: Part 3.1, Knowledge Storage and Extraction (ICML 2024 · similarity 0.85)
- Supervised Knowledge Makes Large Language Models Better In-context Learners (ICLR 2024 · similarity 0.85)
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.
