Research map / Vision-language and multimodal / Multimodal LLMs
MedSynapse-V: Bridging Visual Perception and Clinical Intuition via Latent Memory Evolution
Chunzheng Zhu, Jiaqi Zeng, Junyu Jiang, Jianxin Lin, Yijun Wang
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 CCR reinforcement learning instead of SFT (ablation) | MedSynapse-V · 5 VQA benchmarks average (8B) | average accuracy: 59.2 -> 67.7 | Improved |
| add CCR reinforcement learning after MQPM (ablation) | MedSynapse-V · 5 VQA benchmarks average (8B) | average accuracy: 55.4 -> 67.7 | Improved |
| add MQPM warmup before RL (ablation) | MedSynapse-V · 5 VQA benchmarks average (8B) | average accuracy: 52.9 -> 67.7 | Improved |
| add causal counterfactual reward (ablation) | MedSynapse-V CCR · 5 VQA benchmarks average (8B) | average accuracy: 63.6 -> 67.7 | Improved |
| keep anatomical encoder at inference (no IMT removal) (ablation) | MedSynapse-V · 5 VQA benchmarks average (8B) | average accuracy: 67.7 -> 69.1 | Mixed |
| remove encoder via IMT distillation (ablation) | MedSynapse-V · Inference efficiency | ms/token: 168 -> 102 | Improved |
| use MedSAM3 instead of SAM-Med2D as anatomical encoder (ablation) | MedSynapse-V · 5 VQA benchmarks average (8B) | average accuracy: 63.3 -> 67.7 | Improved |
| use pretrained anatomical encoder instead of random init (ablation) | MedSynapse-V · 5 VQA benchmarks average (8B) | average accuracy: 52.0 -> 67.7 | Improved |
| increase diagnostic probe count N beyond 16 (ablation) | MedSynapse-V · benchmarks | accuracy | Hurt |
| add MedSynapse-V (IMT) to Qwen3-VL-8B | Qwen3-VL-8B · 7 medical benchmarks (average) | accuracy: 48.6 -> 59.6 | Improved |
| adapt Coconut latent reasoning to Qwen3-VL-8B | Qwen3-VL-8B · 7 medical benchmarks (average) | accuracy: 48.6 -> 44.5 | Hurt |
| adapt IVT-LR latent reasoning to Qwen3-VL-8B | Qwen3-VL-8B · 7 medical benchmarks (average) | accuracy: 48.6 -> 50.5 | No clear effect |
Closest papers by meaning
- MedEyes: Learning Dynamic Visual Focus for Medical Progressive Diagnosis (AAAI 2026 · similarity 0.88)
- DDx-TRACE: A Benchmark for Medical Diagnostic Trajectories in VLMs (NeurIPS 2026 · similarity 0.86)
- MedVR: Annotation-Free Medical Visual Reasoning via Agentic Reinforcement Learning (ICLR 2026 · similarity 0.86)
- CRAFT: Causal Responsibility and Failure Tracing in Medical Vision Language Models (NeurIPS 2026 · similarity 0.86)
- MedAgent-Pro: Towards Evidence-based Multi-modal Medical Diagnosis via Reasoning Agentic Workflow (ICLR 2026 · similarity 0.86)
- Beyond Classification Accuracy: Neural-MedBench and the Need for Deeper Reasoning Benchmarks (ICLR 2026 · similarity 0.86)
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.
