Research map / Architectures and efficiency / Attention and transformers
Focused Transformer: Contrastive Training for Context Scaling
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| fine-tune with FoT crossbatch instead of standard long-context fine-tuning (ablation) | OpenLLaMA-3B · TREC few-shot 4K | accuracy: 57.2 -> 60.9 | Helped |
| fine-tune with FoT instead of standard fine-tuning (ablation) | OpenLLaMA-3B · WebQS few-shot 4K | accuracy: 18.7 -> 21.0 | Helped |
| evaluate with context beyond training length (6K, 8K) | FoT 3B fine-tuned on 4K · TREC | accuracy: 60.9 -> 62.5 | Helped |
| extend context 2K to 8K | LongLLaMA 7B · TREC | accuracy: 63.2 -> 75.9 | Helped |
| extend context 2K to 8K | LongLLaMA 3B · WebQS | accuracy: 21.2 -> 22.4 | Helped |
| use differentiable keys/values (FoT d=1) instead of non-differentiable memory (MT) (ablation) | FoT · PG-19 single-doc | perplexity: 14.68 -> 14.18 (512) | Helped |
| increase crossbatch dimension d (more negatives) (ablation) | FoT · PG-19 multi-doc | perplexity | Helped |
| use FoT instead of Memorizing Transformer | 184M transformer fine-tuned on C4 · arXiv 64K | perplexity: 8.60 -> 6.81 | Helped |
| FoT vs Transformer-XL at 2K context | 184M transformer · arXiv 2K | perplexity: 8.21 -> 8.17 | No clear effect |
| increase eval context length 2K to 64K | FoT · arXiv | perplexity: 8.17 -> 6.81 | Helped |
| fine-tune with FoT for long context | LongLLaMA · LM Evaluation Harness short context | average score: 0.53 -> 0.53 (3B), 0.55 -> 0.55 (7B) | No clear effect |
| standard training without crossbatch (distraction issue) (ablation) | standard transformer · attention mass on relevant context | positive attention mass | Hurt |
Closest papers by meaning
- Massive Values in Self-Attention Modules are the Key to Contextual Knowledge Understanding (ICML 2025 · similarity 0.88)
- Exploring Context Window of Large Language Models via Decomposed Positional Vectors (NeurIPS 2024 · similarity 0.88)
- Core Context Aware Transformers for Long Context Language Modeling (ICML 2025 · similarity 0.88)
- NExtLong: Toward Effective Long-Context Training without Long Documents (ICML 2025 · similarity 0.88)
- LongLoRA: Efficient Fine-tuning of Long-Context Large Language Models (ICLR 2024 · similarity 0.88)
- A Training-Free Sub-quadratic Cost Transformer Model Serving Framework with Hierarchically Pruned Attention (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.