Research map / Large language models / Instruction and fine-tuning
Context Tuning for In-Context Optimization
Jack Lu, Ryan Teehan, Zhenbang Yang, Mengye 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 |
|---|---|---|---|
| initialize prefix from demonstration KV cache (CT-KV) instead of random-token Prefix Tuning | Prefix Tuning · BBH (Llama3-8B) | accuracy (%): 52.7 -> 57.9 | Improved |
| initialize soft prompt from demonstrations (CT-Prompt) vs Prompt Tuning | Prompt Tuning · NLP-LR (GPT-2) | accuracy (%): 41.4 -> 43.2 | Improved |
| increase Prompt/Prefix Tuning tokens m to match demonstration tokens (ablation) | Prompt Tuning · BBH (Llama3-8B) | accuracy (%): 50.8 -> 47.5 | Hurt |
| use KV-prefix CT-KV instead of CT-Prompt | Context Tuning · NLP-LR (GPT-2) | accuracy (%): 43.2 -> 44.2 | Improved |
| apply CT-KV after TTT | TTT · NLP-LR (GPT-2) | accuracy (%): 44.1 -> 47.6 | Improved |
| add Leave-One-Out Masking (ablation) | CT-KV · BBH (Llama3-8B) | accuracy (%): 54.4 -> 57.9 | Improved |
| add Leave-One-Out Masking (ablation) | CT-KV · ARC (Llama3.2-1B) | accuracy (%): 23.8 -> 22.5 | Hurt |
| add Token Dropout (ablation) | CT-KV · BBH (Llama3-8B) | accuracy (%): 55.3 -> 57.9 | Improved |
| use CT-KV instead of ICL on large models | ICL · BBH (Qwen3-32B) | accuracy (%): 63.1 -> 69.1 | Improved |
| use CT-KV instead of ICL with 100 demonstrations (many-shot) | ICL · BBH 8 tasks (Llama3-8B-Instruct) | accuracy (%): 46.5 -> 62.6 | Improved |
| train longer (64 vs 32 epochs) (ablation) | Prefix Tuning · BBH (Llama3-8B) | accuracy (%): 52.73 -> 50.49 | Hurt |
| train longer (16 to 64 epochs) (ablation) | CT-KV · BBH (Llama3-8B) | accuracy (%): 57.91 -> 57.88 | No clear effect |
Closest papers by meaning
- Optimization Inspired Few-Shot Adaptation for Large Language Models (NeurIPS 2025 · similarity 0.91)
- Generative Adapter: Contextualizing Language Models in Parameters with A Single Forward Pass (ICLR 2025 · similarity 0.89)
- The Surprising Effectiveness of Test-Time Training for Few-Shot Learning (ICML 2025 · similarity 0.89)
- In-context Vectors: Making In Context Learning More Effective and Controllable Through Latent Space Steering (ICML 2024 · similarity 0.89)
- All You Need is One: Capsule Prompt Tuning with a Single Vector (NeurIPS 2025 · similarity 0.88)
- Many-Shot In-Context Learning (NeurIPS 2024 · 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.
