Add dataset card

README.md

@@ -1,3 +1,130 @@
 ---
 license: apache-2.0
+pretty_name: MedImageOSWorld
+task_categories:
+  - image-text-to-text
+tags:
+  - gui-agents
+  - computer-use
+  - benchmark
+  - medical-imaging
+  - vision-language-models
+  - ct
+  - mri
+  - ultrasound
+  - simulation
 ---
+
+# MedImageOSWorld: Benchmarking GUI Agents for Medical Image Consoles
+
+> Anonymous release accompanying a paper under double-blind review.
+
+<video controls autoplay muted loop playsinline width="100%" src="https://huggingface.co/datasets/neutral-123/MedImageOSWorld/resolve/main/MedImageOSWorld_30s_v4.mp4"></video>
+
+MedImageOSWorld is a benchmark for evaluating whether general-purpose GUI agents can operate simulated **CT**, **MR**, and **ultrasound** acquisition consoles — and whether their actions actually acquire the intended anatomy.
+
+Agents see only screenshots and act through mouse and keyboard. They configure protocols, plan acquisitions, inspect the resulting images, and make corrective adjustments. A hidden oracle then scores the final acquisition against anatomical ground truth in the digital subject's coordinate frame.
+
+## Why it is different: action-conditioned acquisition
+
+In most computer-use and clinical-agent benchmarks, the data the agent looks at is fixed. In MedImageOSWorld, **agent actions change the physical acquisition state**, which changes the image evidence produced at later steps:
+
+| Console | What the agent plans | What changes in the next image |
+| --- | --- | --- |
+| CT | scan range, isocentre, monitoring ROI, timing on the scout | coverage of the acquired stack |
+| MR | plane position, orientation, FOV, slab/slices on localizers | resliced anatomy at the committed geometry |
+| Ultrasound | probe position, orientation, contact, depth/gain/focus | live image rendered from the probe pose |
+
+Display-only actions (e.g. window/level) change the interface but not the acquisition. An agent can complete the workflow and still acquire the wrong anatomy; the benchmark measures both separately.
+
+## Environment
+
+- **Observation:** a PNG screenshot of the console at a fixed viewport. No DOM, accessibility tree, simulator state, or scanner-control API.
+- **Action:** constrained mouse events and an allowlisted set of keyboard events, plus explicit wait. Simulated time advances only through executed actions.
+- **Subjects:** XCAT-based digital human anatomy phantoms with registered anatomical structures in a shared subject coordinate frame. No human or animal scans are acquired or used.
+- **Action receipts:** every step logs the raw model output, any normalization/repair, whether it was executed or rejected, and feedback returned to the agent.
+
+## Tasks and capability ladder
+
+The benchmark contains **49 task definitions** — 8 CT, 34 MR, 7 ultrasound — designed by three medical imaging experts (assisted by a frontier AI model), cross-reviewed, and finalized against a shared capability ladder:
+
+| Level | Capability |
+| --- | --- |
+| L0 | Operate / read a value |
+| L1 | Follow a workflow |
+| L2 | Judge view / adequacy |
+| L3 | Plan geometry (mm, °) |
+| L4 | Set parameters |
+| L5 | Control with feedback |
+| L6 | Detect & recover |
+
+Task families cover thoracic, renal, hepatic, and cardiac acquisition: CT lung coverage, contrast timing, dose-index constraints, and undercoverage recovery; MR cardiac cine planes, T1/T2 mapping, and bilateral renal coverage; ultrasound kidney search, optimization, and fault recovery. Each task condition targets 30 digital subjects × 3 seeds.
+
+## Evaluation
+
+Each episode receives `PASS`, `PARTIAL`, `FAIL`, or an undetermined outcome; environment or scorer faults are `INVALID` and excluded from denominators. Three 0–100 metrics are reported side by side and never merged:
+
+- **SR — Success rate** (primary): share of slots meeting the task's full success criterion.
+- **PR — Progress rate:** mean fraction of workflow stages completed.
+- **AQ — Acquisition quality:** continuous, normalized quality of the acquired result from hidden ground truth (geometry in mm/degrees, anatomical coverage fractions).
+
+The overall score weights capability levels by `w_i = i + 1`:
+
+$$
+\text{Score} = 100\,\frac{\sum_{i=0}^{6}(i+1)\sum_{j=1}^{n_i} s_{ij}}{\sum_{i=0}^{6}(i+1)\,n_i}
+$$
+
+where `s_ij` is the SR proportion for task `j` at level `i` (weighted denominator 181).
+
+## Results
+
+11 open-weight models (greedy decoding) and 2 API-served proprietary models, over **31,197 valid scored episodes** (30,513 open-weight, 684 proprietary; proprietary agents use the first five assets per task).
+
+| Agent | Access | Overall SR |
+| --- | --- | ---: |
+| GPT-6 Sol | API | 34.5 |
+| Claude Opus 5.5 | API | 34.1 |
+| Qwen3.8-27B | open-weight | 25.0 |
+| Muse-Glimmer-30B | open-weight | 20.8 |
+| Gemma-4-31B | open-weight | 14.6 |
+| UI-Mate-27B | open-weight | 10.7 |
+| Qwen3-VL-30B-A3B-Thinking | open-weight | 7.7 |
+| GUI-Owl-1.5-32B-Thinking | open-weight | 6.5 |
+| GUI-Owl-1.5-32B-Instruct | open-weight | 5.5 |
+| Qwen2.5-VL-72B-Instruct | open-weight | 5.2 |
+| Qwen3-VL-30B-A3B-Instruct | open-weight | 3.7 |
+| OpenCUA-32B | open-weight | 3.3 |
+| UI-TARS-1.5-7B | open-weight | 3.0 |
+
+The API-served agents exceed Qwen3.8-27B in workflow progress by only 9–11 points (PR 85.5 and 83.4 vs. 74.5), but lead in acquisition quality by about 19 points (AQ 61.1 and 61.3 vs. 42.0).
+
+Key findings:
+
+- **Agents complete workflows but rarely acquire the intended anatomy.** Open-weight SR ranges 3.0–25.0 while PR reaches 21.5–74.5.
+- **Success collapses between perception and planning.** The best open-weight agent reaches 79–90% at L0–L2, at most 12% at L3, and every open-weight agent scores 0% at L5.
+- **CT is nearly unsolved** for open-weight agents (SR ≤ 1.5); proprietary agents reach 10.3 and 1.7.
+- **Agents observe but do not correct.** Review gates are almost always passed, but agents change the responsible acquisition parameter in only 14% of renal MR coverage-correction episodes, and in CT recovery tasks no agent corrected a clipped range.
+- **Budget exhaustion dominates failure:** 64.6% of scored open-weight episodes end with an exhausted action budget.
+
+## Release status
+
+The simulator and data will be made available through this repository. File checksums, access instructions, and the handling of licensed assets and restricted scoring components are still being finalized. Hidden oracles, ground-truth masks, and scoring tolerances are never exposed to evaluated agents.
+
+## Scope and limitations
+
+- Simulated consoles and task-specific physical abstractions — **not** a scanner, not a medical device, and not evidence of scanner equivalence or clinical validation.
+- Geometry and coverage measurements characterize acquired anatomy under these models; clinical image quality and diagnostic utility require separate evaluation.
+- The capability ladder organizes task requirements; it is not a psychometrically validated difficulty scale.
+- Results depend on the subject cohort, task versions, and agent interface.
+- XCAT phantoms are licensed separately and are not redistributed under this repository's license.
+
+## Citation
+
+```bibtex
+@misc{medimageosworld2026,
+  title  = {MedImageOSWorld: Benchmarking GUI Agents for Medical Image Consoles},
+  author = {Anonymous},
+  year   = {2026},
+  note   = {Under review}
+}
+```