Research map / Vision-language and multimodal / Multimodal LLMs
Is Visual in-Context Learning for Compositional Medical Tasks within Reach?
Simon Reiß, Zdravko Marinov, Alexander Jaus, Constantin Seibold, M. Saquib Sarfraz, Erik Rodner, Rainer Stiefelhagen
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 |
|---|---|---|---|
| use sequence-token masking instead of random token masking (ablation) | GPT2 in-context learner · MedSAM compositional segmentation | IoU: 26.49% -> 52.60% | Improved |
| use image-token masking instead of random token masking (ablation) | GPT2 in-context learner · MedSAM compositional segmentation | IoU: 26.49% -> 44.83% | Improved |
| use sequence-token masking instead of image-token masking (ablation) | GPT2 in-context learner · MedSAM compositional denoising | PSNR: 16.851 -> 20.131 | Improved |
| mix token, image and sequence masking (ablation) | GPT2 in-context learner · MedSAM compositional segmentation | IoU: 52.60% -> 50.39% | Hurt |
| add 2 END special tokens (ablation) | sequence-token masking · MedSAM compositional segmentation | IoU: 52.60% -> 52.82% | Improved |
| increase END tokens from 2 to 4 (ablation) | sequence-token masking · MedSAM compositional segmentation | IoU: 52.82% -> 52.78% | No clear effect |
| train longer (52 epochs instead of 8) (ablation) | sequence-token masking k=2 · MedSAM compositional segmentation | IoU: 52.82% -> 61.21% | Improved |
| fine-tune codebook on task outputs (ablation) | VQ-GAN · MedSAM codebook upper bound, fixed colors | segmentation IoU: 59.89% -> 86.19% | Improved |
| add task balancing to codebook training (ablation) | VQ-GAN task training · MedSAM codebook upper bound, fixed colors | segmentation IoU: 86.19% -> 88.07% | Improved |
| add color remapping augmentation (ablation) | VQ-GAN task balance · MedSAM codebook upper bound, fixed colors | segmentation IoU: 88.07% -> 84.07% | Hurt |
| add color remapping augmentation (ablation) | VQ-GAN task balance · MedSAM codebook upper bound, random colors | segmentation IoU: 69.21% -> 83.00% | Improved |
| add color remapping augmentation (ablation) | VQ-GAN task and data balance · MedSAM codebook upper bound, random colors | segmentation IoU: 71.15% -> 83.19% | Improved |
Closest papers by meaning
- Test-Time Visual In-Context Tuning (CVPR 2025 · similarity 0.87)
- Unveiling the Compositional Ability Gap in Vision-Language Reasoning Model (NeurIPS 2025 · similarity 0.86)
- Exploring Task-Level Optimal Prompts for Visual In-Context Learning (AAAI 2025 · similarity 0.85)
- Is attention required for ICL? Exploring the Relationship Between Model Architecture and In-Context Learning Ability (ICLR 2024 · similarity 0.85)
- Embracing Collaboration Over Competition: Condensing Multiple Prompts for Visual In-Context Learning (CVPR 2025 · similarity 0.85)
- SMMILE: An expert-driven benchmark for multimodal medical in-context learning (NeurIPS 2025 · 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.
