| { |
| "@context": { |
| "@language": "en", |
| "@vocab": "https://schema.org/", |
| "citeAs": "cr:citeAs", |
| "column": "cr:column", |
| "conformsTo": "dct:conformsTo", |
| "cr": "http://mlcommons.org/croissant/", |
| "rai": "http://mlcommons.org/croissant/RAI/", |
| "data": { |
| "@id": "cr:data", |
| "@type": "@json" |
| }, |
| "dataType": { |
| "@id": "cr:dataType", |
| "@type": "@vocab" |
| }, |
| "dct": "http://purl.org/dc/terms/", |
| "equivalentProperty": "cr:equivalentProperty", |
| "examples": { |
| "@id": "cr:examples", |
| "@type": "@json" |
| }, |
| "extract": "cr:extract", |
| "field": "cr:field", |
| "fileProperty": "cr:fileProperty", |
| "fileObject": "cr:fileObject", |
| "fileSet": "cr:fileSet", |
| "format": "cr:format", |
| "includes": "cr:includes", |
| "isLiveDataset": "cr:isLiveDataset", |
| "jsonPath": "cr:jsonPath", |
| "key": "cr:key", |
| "md5": "cr:md5", |
| "parentField": "cr:parentField", |
| "path": "cr:path", |
| "prov": "http://www.w3.org/ns/prov#", |
| "recordSet": "cr:recordSet", |
| "references": "cr:references", |
| "regex": "cr:regex", |
| "repeated": "cr:repeated", |
| "replace": "cr:replace", |
| "samplingRate": "cr:samplingRate", |
| "sc": "https://schema.org/", |
| "separator": "cr:separator", |
| "source": "cr:source", |
| "subField": "cr:subField", |
| "transform": "cr:transform" |
| }, |
| "@type": "sc:Dataset", |
| "name": "ActiveVision", |
| "description": "ActiveVision is a benchmark for evaluating active visual reasoning in multimodal large language models. It contains 85 instances across 17 tasks organised into three families — Distributed Scanning (5 tasks), Sequential Traversal (5 tasks), and Visual Attribute Transfer (7 tasks) — each requiring iterative inspection of an image to accumulate evidence and answer. All instances are programmatically generated; the released images are photorealistic re-renders of matplotlib structural drafts produced via OpenAI gpt-image-2.", |
| "conformsTo": "http://mlcommons.org/croissant/1.0", |
| "dct:conformsTo": "http://mlcommons.org/croissant/RAI/1.0", |
| "license": "https://creativecommons.org/licenses/by/4.0/", |
| "url": "https://huggingface.co/datasets/activevision/hpXgvFBl7ZxO", |
| "version": "0.4.0", |
| "datePublished": "2026-05-06", |
| "creator": { |
| "@type": "sc:Organization", |
| "name": "Anonymous (NeurIPS 2026 double-blind submission)" |
| }, |
| "publisher": { |
| "@type": "sc:Organization", |
| "name": "Anonymous (NeurIPS 2026 double-blind submission)" |
| }, |
| "keywords": [ |
| "multimodal", |
| "vision-language", |
| "benchmark", |
| "active perception", |
| "active vision", |
| "visual reasoning", |
| "synthetic", |
| "MLLM evaluation" |
| ], |
| "citeAs": "Anonymous. ActiveVision: An Exam for Active Observers. NeurIPS 2026 Datasets and Benchmarks Track (under review).", |
|
|
| "rai:dataCollection": "All 85 instances are programmatically generated. For each task, a Python generator (creation.py, seedable, deterministic) lays out a structural specification — region partition, arrow geometry, maze graph, brush-stroke field, etc. — and computes the ground-truth answer in closed form from the same seed. The structural specification is rendered in two stages: (1) a plain matplotlib draft used as the geometric reference, and (2) a photorealistic re-render produced via a single OpenAI gpt-image-2 image-edit call using a per-task prompt that preserves silhouettes, positions, counts, and labels but changes only the surface material (e.g. stones on sand, hedge maze from above, starfield). The released benchmark contains the photorealistic re-renders only.", |
| "rai:dataCollectionType": [ |
| "Software Collection", |
| "Experiments" |
| ], |
| "rai:dataCollectionTimeframe": [ |
| "2026-04-30", |
| "2026-05-06" |
| ], |
| "rai:dataCollectionRawData": "Per-task generators are included under code/<category>/<task>/creation.py; the matplotlib structural drafts can be regenerated from seed at any difficulty. The per-task gpt-image-2 image-edit prompts are documented in code/gpt_image_prompts.json.", |
| "rai:dataPreprocessingProtocol": [ |
| "Stage 1 (geometry): matplotlib + numpy + scipy generate the structural draft and ground-truth answer deterministically from a fixed seed (creation.py --seed --difficulty 4).", |
| "Stage 2 (re-render): the matplotlib draft is sent to OpenAI gpt-image-2 via the image-edit endpoint, with a per-task prompt that keeps silhouettes, positions, counts, and labels intact while replacing surface material with a photorealistic style. No human-in-the-loop curation occurs between Stage 1 and Stage 2." |
| ], |
| "rai:dataAnnotationProtocol": "Ground-truth answers are computed programmatically inside creation.py from the same random seed used to render the image. Per-instance metadata (region adjacency, arrow chains, traversal paths, blob Hausdorff distances, etc.) is recorded in data/annotations/<task>.jsonl alongside the rendered image. No human annotators were used.", |
| "rai:annotationsPerItem": "1 (deterministic programmatic ground truth, derived from the generation seed)", |
| "rai:machineAnnotationTools": [ |
| "Python 3.11", |
| "matplotlib", |
| "numpy", |
| "scipy", |
| "OpenCV (cv2)", |
| "Pillow (PIL)", |
| "OpenAI gpt-image-2 (image-edit endpoint)" |
| ], |
| "rai:dataReleaseMaintenancePlan": "v0.4 is a frozen released snapshot at difficulty 4 with 5 instances per task (85 total). The maintained artifact is the generation pipeline (creation.py + gpt_image_prompts.json), which supports regeneration of additional instances at arbitrary difficulty and held-out splits with unpublished seeds. Future versions will increase per-task instance counts. Versioning follows semantic versioning by Git tag (v0.4.0, v0.5.0, ...).", |
| "rai:dataUseCases": [ |
| "Testing", |
| "Validation" |
| ], |
| "rai:dataLimitations": [ |
| "Adversarial design: tasks are deliberately constructed to defeat one-shot symbolic extraction, off-the-shelf computer-vision pipelines (OCR/edge/blob), prior leakage, and gestalt heuristics. Performance on this benchmark is therefore not predictive of general-purpose vision-language ability — it specifically diagnoses active visual observation.", |
| "Small released set: 5 instances per task (85 total), intended for evaluation only and not for training. Larger evaluation splits and held-out seeds can be regenerated from the released pipeline.", |
| "Re-render dependency: photorealistic images depend on OpenAI gpt-image-2 (a closed-source service); exact reproduction of these specific PNGs requires API access. The underlying matplotlib drafts are fully reproducible from seed.", |
| "Stylistic coverage: the gpt-image-2 re-render templates currently span a Western-centric set of natural surfaces (beach, forest floor, kitchen counter, hedge maze, starfield, etc.); broader cultural-aesthetic coverage is left to future versions." |
| ], |
| "rai:dataBiases": [ |
| "Both image families (matplotlib drafts and gpt-image-2 re-renders) inherit aesthetic and cultural priors from their generators.", |
| "Question text is in English only.", |
| "No human subjects, no demographic data, no scenes depicting real persons, locations, or events." |
| ], |
| "rai:dataSocialImpact": "Intended for diagnostic evaluation of multimodal LLM perception. Negative impacts are limited because the dataset contains no humans, no real-world events, and no personal data. Positive impact: surfaces gaps in active visual reasoning that are hidden by current static-perception benchmarks, motivating architectural research toward iterative perception-driven reasoning.", |
| "rai:personalSensitiveInformation": "None. The dataset contains only synthetic images of geometric or natural-material patterns (regions, arrows, mazes, contours, brush strokes, starfields). No human subjects, no faces, no PII, no health, location, age, gender, ethnicity, or socio-economic data.", |
| "rai:hasSyntheticData": true, |
| "prov:wasGeneratedBy": "code/<category>/<task>/creation.py — deterministic seeded generators (Python 3.11). See code/scope.md for the full pipeline specification and the design rationale (the generation pipeline is the artifact, per the paper).", |
|
|
| "distribution": [ |
| { |
| "@type": "cr:FileObject", |
| "@id": "manifest", |
| "name": "manifest.json", |
| "description": "Canonical index of all 85 benchmark instances with id, task, category, image path, question, ground-truth answer, and image SHA-256.", |
| "contentUrl": "data/manifest.json", |
| "encodingFormat": "application/json", |
| "sha256": "2caa9eb71ca1687ab2831058a0f64f81619df9ac726a79d519e4ff015ffbba4c" |
| }, |
| { |
| "@type": "cr:FileSet", |
| "@id": "images", |
| "name": "images", |
| "description": "85 photorealistic benchmark images (one PNG per instance), produced by re-rendering matplotlib structural drafts through OpenAI gpt-image-2.", |
| "encodingFormat": "image/png", |
| "includes": "data/images/*.png" |
| }, |
| { |
| "@type": "cr:FileSet", |
| "@id": "annotations", |
| "name": "annotations", |
| "description": "Per-instance verification metadata in JSON Lines format, one file per task. Records include region adjacency, arrow chains, traversal paths, Hausdorff distances, etc., derived from the same generator seed as the image.", |
| "encodingFormat": "application/jsonlines", |
| "includes": "data/annotations/*.jsonl" |
| }, |
| { |
| "@type": "cr:FileSet", |
| "@id": "code", |
| "name": "code", |
| "description": "Generation pipeline: per-task creation.py generators, per-task creation.md specs, gpt-image-2 re-render prompts (gpt_image_prompts.json), and the overall design specification (scope.md). The pipeline is the artifact: any number of fresh instances at any difficulty can be regenerated from this code.", |
| "encodingFormat": "text/x-python", |
| "includes": "code/**/*" |
| } |
| ], |
|
|
| "recordSet": [ |
| { |
| "@type": "cr:RecordSet", |
| "@id": "instances", |
| "name": "instances", |
| "description": "One record per benchmark instance. 85 records total, balanced across 17 tasks (5 each) in 3 categories (distributed_scanning: 25, sequential_traversal: 25, visual_attribute_transfer: 35). Records are joined by id: each PNG in the images FileSet is matched to a row of manifest.json by stripping the 'images/' prefix and '.png' extension from the file path and matching the resulting basename to manifest.json's id field.", |
| "key": {"@id": "instances/id"}, |
| "field": [ |
| { |
| "@type": "cr:Field", |
| "@id": "instances/id", |
| "name": "id", |
| "description": "Unique instance identifier formed from the task initials and a zero-based index (e.g. CR-0 for counting_regions instance 0). This field is the join key between the images FileSet (filename without extension) and manifest.json's id column.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileSet": {"@id": "images"}, |
| "extract": {"fileProperty": "filename"}, |
| "transform": {"regex": "([^/]+)\\.png"} |
| }, |
| "references": { |
| "fileObject": {"@id": "manifest"}, |
| "extract": {"jsonPath": "$[*].id"} |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "instances/task", |
| "name": "task", |
| "description": "Task name. One of: attribute_group_counting, bounded_faces_counting, counting_connected_components, counting_regions, tangled_loops, arrow_chain, color_zone_sequence, line_intersections, maze, traverse_ordering, constellation_match_count, contour_silhouette_count, spot_the_contour_diff, spot_the_field_diff, spot_the_signal_diff, spot_the_stroke_diff, stroke_gesture_count.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": {"@id": "manifest"}, |
| "extract": {"jsonPath": "$[*].task"} |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "instances/category", |
| "name": "category", |
| "description": "Task family. One of: distributed_scanning, sequential_traversal, visual_attribute_transfer.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": {"@id": "manifest"}, |
| "extract": {"jsonPath": "$[*].category"} |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "instances/image_path", |
| "name": "image_path", |
| "description": "Relative path to the photorealistic image file inside the dataset distribution.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": {"@id": "manifest"}, |
| "extract": {"jsonPath": "$[*].image"} |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "instances/image_sha256", |
| "name": "image_sha256", |
| "description": "SHA-256 hex digest of the photorealistic image PNG bytes.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": {"@id": "manifest"}, |
| "extract": {"jsonPath": "$[*].image_sha256"} |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "instances/image", |
| "name": "image", |
| "description": "The photorealistic benchmark image as a binary PNG, sourced from the images FileSet. Joined to the rest of the record by instances/id (which is the regex-stripped filename).", |
| "dataType": "sc:ImageObject", |
| "source": { |
| "fileSet": {"@id": "images"}, |
| "extract": {"fileProperty": "content"} |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "instances/question", |
| "name": "question", |
| "description": "The natural-language task question presented to the model. Includes the answer-format instruction (final answer enclosed in <answer>...</answer> tags).", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": {"@id": "manifest"}, |
| "extract": {"jsonPath": "$[*].question"} |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "instances/answer", |
| "name": "answer", |
| "description": "Ground-truth answer. Type varies by task: integer count (e.g. 10), single letter (e.g. 'H'), letter pair (e.g. 'E-F'), or comma-separated letter sequence (e.g. 'L, G, H, F, C, K').", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": {"@id": "manifest"}, |
| "extract": {"jsonPath": "$[*].answer"} |
| } |
| } |
| ] |
| } |
| ] |
| } |
|
|
