Research map / Label-efficient and robust learning / Recognition and long-tail
Unsupervised Cross-Task Generalization via Retrieval Augmentation
Bill Yuchen Lin, Kangmin Tan, Chris Scott Miller, Beiwen Tian, Xiang Ren
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 |
|---|---|---|---|
| retrieve with upstream-model hidden representations (ReCross no rerank) instead of SBERT | ReCross · unseen tasks mean | SoftEM: 40.24 -> 43.57 | Improved |
| add re-ranking trained by distant supervision (ablation) | ReCross · unseen tasks mean | SoftEM: 43.57 -> 44.53 | Improved |
| ReCross retrieval augmentation over BART0 | BART0 · unseen tasks mean | SoftEM: 40.38 -> 44.53 | Improved |
| ReCross on BIG-bench subset | BART0 · BIG-bench subset | SoftEM: 28.07 -> 35.23 | Improved |
| random retrieval instead of no retrieval | BART0 · unseen tasks mean | SoftEM: 40.38 -> 39.13 | Hurt |
| increase query size from 16 to 32 (ablation) | ReCross · unseen tasks mean | SoftEM: 44.53 -> 42.52 | Hurt |
| reduce query size to 1 (ablation) | ReCross · unseen tasks mean | SoftEM: 44.53 -> 43.20 | Hurt |
| reduce retrieved size from 512 to 256 (ablation) | ReCross · unseen tasks mean | SoftEM: 44.53 -> 40.80 | Hurt |
| increase retrieved size from 512 to 1024 (ablation) | ReCross · unseen tasks mean | SoftEM: 44.53 -> 44.02 | No clear effect |
| increase upsampling ratio from 2 to 3 or 4 (ablation) | ReCross · unseen tasks mean | SoftEM: 44.53 -> 43.92 (mu=3), 43.91 (mu=4) | Hurt |
| perturb template instructions by concatenation (ablation) | ReCross · unseen tasks mean | SoftEM: 44.53 -> 40.52 | Hurt |
| perturb template instructions by random word change (ablation) | ReCross · unseen tasks mean | SoftEM: 44.53 -> 42.07 | Hurt |
Closest papers by meaning
- Understanding Retrieval-Augmented Task Adaptation for Vision-Language Models (ICML 2024 · similarity 0.85)
- Contrastive Retrieval Heads for Improved Attention-Based Reranking (NeurIPS 2026 · similarity 0.85)
- From RAG to Memory: Non-Parametric Continual Learning for Large Language Models (ICML 2025 · similarity 0.85)
- From Instance Training to Instruction Learning: Task Adapters Generation from Instructions (NeurIPS 2024 · similarity 0.84)
- Efficient Length-Generalizable Attention via Causal Retrieval for Long-Context Language Modeling (ICML 2025 · similarity 0.84)
- Attention in Large Language Models Yields Efficient Zero-Shot Re-Rankers (ICLR 2025 · similarity 0.84)
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.
