Research map / Continual learning / Continual fine-tuning of LLMs
Sparse Spectral LoRA: Routed Experts for Medical VLMs
Omid Nejatimanzari, Hojat Asgariandehkordi, Taha Koleilat, Yiming Xiao, Hassan Rivaz
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 SVD-segment experts in MoE vs single LoRA with SVD fragments (ablation) | MedQwen · medical VQA average | avg accuracy: 60.1 -> 68.2 | Improved |
| add SVD-segment initialization to MoE with scaling (ablation) | MedQwen · medical VQA average | avg accuracy: 67.3 -> 68.2 | Improved |
| add MoE scaling to SVD-initialized MoE (ablation) | MedQwen · medical VQA average | avg accuracy: 67.3 -> 68.2 | Improved |
| replace MoELoRA with SVD-structured MoE (ablation) | MedQwen · PathVQA close-ended | accuracy: 78.4 -> 84.2 | Improved |
| replace standard LoRA with MedQwen in sequential training (Harvard-FairVLMed to PathVQA) (ablation) | Qwen2.5-VL 7B · Harvard-FairVLMed after PathVQA | accuracy retention: over 50% drop -> 5% drop | Improved |
| increase LoRA rank from 64 to 128 (ablation) | MedQwen · radiology zero-shot classification | accuracy: +0.3% | No clear effect |
| activate more experts than 2-of-8 (ablation) | MedQwen · medical VQA | accuracy | Hurt |
| increase total experts (2-of-8 to 4-of-16 / 8-of-32) (ablation) | MedQwen · medical VQA | accuracy | Mixed |
| use MedQwen instead of single rank-32 LoRA | BiomedCLIP ViT-B/16 · zero-shot radiology classification | avg accuracy: 55.45 -> 58.83 | Improved |
| use MedQwen instead of PiSSA | BiomedCLIP ViT-B/16 · zero-shot radiology classification | avg accuracy: 50.08 -> 58.83 | Improved |
| use MedQwen instead of general Qwen-2.5-VL | Qwen2.5-VL 7B · medical VQA average | avg accuracy: 49.3 -> 68.2 | Improved |
Closest papers by meaning
- VILA-M3: Enhancing Vision-Language Models with Medical Expert Knowledge (CVPR 2025 · similarity 0.89)
- Uni-Med: A Unified Medical Generalist Foundation Model For Multi-Task Learning Via Connector-MoE (NeurIPS 2024 · similarity 0.87)
- OmniMedVQA: A New Large-Scale Comprehensive Evaluation Benchmark for Medical LVLM (CVPR 2024 · similarity 0.87)
- Toward a Vision-Language Foundation Model for Medical Data: Multimodal Dataset and Benchmarks for Vietnamese PET/CT Report Generation (NeurIPS 2025 · similarity 0.87)
- OpenMedReason: Scientific Reasoning Supervision for Medical Vision–Language Models (NeurIPS 2026 · similarity 0.87)
- Towards a Multimodal Large Language Model with Pixel-Level Insight for Biomedicine (AAAI 2025 · similarity 0.87)
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.
