Research map / Vision-language and multimodal / Multimodal LLMs
ORBIT: A Prognostic World Model for Ocular Reasoning Based on Imagined Trajectories
Jiangtao Yan, Yanlin Qu, Yansheng Qiu, Shujian Gao, Wei Yu, Zheng Wang, Xiaodong Sun, Huixun Jia, Diping Song
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ORBIT, a prognostic world model for longitudinal ophthalmic management of blinding fundus diseases, framed as a POMDP that reasons over imagined future trajectories.
Problem
Longitudinal management of blinding fundus diseases is a POMDP requiring a precision-risk trade-off between intervention and over-treatment, since true pathology is often obscured in static observations; existing vision models are uninterpretable and memoryless, and vision-language models are noted
Approach
Frames ocular reasoning as prognostic world modeling: the model imagines future disease trajectories to support intervention decisions under partial observability. Details beyond the abstract are not available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add closed-loop visual foresight with Rectify step to open-loop SFT+RL policy (ablation) | ORBIT · ORBIT-LD Test, Decision (overall) | AUC: 86.73 -> 88.96 | Improved |
| add closed-loop visual foresight (ablation) | ORBIT · ORBIT-LD Test, Decision (overall) | Specificity: 84.92 -> 95.63 | Improved |
| add closed-loop visual foresight (ablation) | ORBIT · ORBIT-LD Test, Decision (overall) | Sensitivity: 88.54 -> 82.29 | Hurt |
| combine SFT and RL versus RL only in open-loop setting (ablation) | ORBIT policy agent · ORBIT-LD Test, Decision (overall) | Accuracy: 87.96 (RL only) vs 86.85 (SFT+RL) | No clear effect |
| domain-specific optimization of Qwen3-VL-8B (SFT/RL) versus generalist baseline (ablation) | Qwen3-VL-8B-Instruct · ORBIT-LD Test, Decision (overall) | AUC: 72.32 -> 86.41 (SFT only) | Improved |
| raise Rectify threshold tau_risk from 0.7 to 0.9 (ablation) | ORBIT · ORBIT-LD Test, Decision (overall) | Sensitivity: 82.29 -> 59.72 | Hurt |
| vary Rectify threshold within 0.6 to 0.8 (ablation) | ORBIT · ORBIT-LD Test, Decision (overall) | Accuracy: 88.33 / 88.52 / 88.15 | No clear effect |
| fine-tune Qwen-Image-Edit as Visual Simulator on longitudinal OCT pairs (ablation) | Qwen-Image-Edit · Real longitudinal OCT pairs | DINOv3-FID: 37.62 -> 2.761 | Improved |
| use five central B-scans instead of one or eight (ablation) | ORBIT · ORBIT-LD Test | Accuracy: 0.741 -> 0.778 (1 vs 5); 0.719 at 8 | Improved |
| fine-tune with ORBIT pipeline versus larger general VLMs on perception | Qwen3-VL-8B-Instruct · ORBIT-LD Test, Perception | Avg Acc: 64.66 -> 78.12 | Improved |
Closest papers by meaning
- CRAFT: Causal Responsibility and Failure Tracing in Medical Vision Language Models (NeurIPS 2026 · similarity 0.86)
- MedEyes: Learning Dynamic Visual Focus for Medical Progressive Diagnosis (AAAI 2026 · similarity 0.86)
- A World Model of Radiologist Reading for Medical Image Representation Learning (NeurIPS 2026 · similarity 0.85)
- Unscrambling disease progression at scale: fast inference of event permutations with optimal transport (NeurIPS 2024 · similarity 0.84)
- DDx-TRACE: A Benchmark for Medical Diagnostic Trajectories in VLMs (NeurIPS 2026 · similarity 0.84)
- MedSIGHT: Towards Grounded Visual Comprehension in Medical Large Vision-Language Models (ICML 2026 · similarity 0.83)
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.
