Update dataset card and taxonomy graphic

README.md

@@ -5,16 +5,17 @@ language:
 task_categories:
   - visual-question-answering
 tags:
-  - vision-language
+  - vision-language-models
   - multimodal-reasoning
+  - visual-reasoning
   - geometric-reasoning
   - spatial-reasoning
-  - visual-reasoning
   - physical-reasoning
+  - sycophancy
+  - robustness
   - procedural-generation
   - synthetic
   - exact-ground-truth
-  - robustness
   - benchmark
 pretty_name: GRIP
 size_categories:
@@ -34,59 +35,106 @@ configs:
         path: combined/all_annotations_combined-*.parquet
 ---
 
-# GRIP
+# GRIP: Geometric Reasoning and Intuitive Physics
+
+**A controlled benchmark that measures whether a vision-language model keeps a correct answer when it is challenged.**
+
+| | |
+|---|---|
+| Paper | *Geometric Parrots: Rule Absence vs. Rule Fragility in Spatial and Physical Reasoning for Vision-Language Models* |
+| Project page | [GRIP Benchmark](https://bilaljawaid980.github.io/grip-benchmark/) |
+| Code | [GitHub](https://github.com/bilaljawaid980/Geomatric-Reasoning-Benchmark-Dataset) |
+| Scale | 100,000 images · 600,000 questions · 34 domains · 9 reasoning families |
+
+![GRIP taxonomy and Dual-Loop Evaluation overview](assets/grip-taxonomy-burst.png)
+
+## What GRIP is for
+
+Vision-language models can answer a geometric or physical reasoning question correctly and then abandon that answer when challenged. Single-turn accuracy cannot distinguish rule absence from rule fragility because it scores each response in isolation.
+
+GRIP separates them. Every scene is produced by deterministic code from a seeded generator, and every answer is computed from the same latent scene state used by the renderer. This makes it possible to assign direction to an answer change:
+
+- **Correction:** wrong → right
+- **Capitulation:** right → wrong
 
-**Geometric Parrots: Rule Absence vs. Rule Fragility in Spatial and Physical Reasoning for Vision-Language Models**
+Correction and capitulation are measured over different populations and are reported separately.
 
-GRIP is a procedurally generated benchmark for evaluating geometric, spatial, visual, and physical reasoning in vision-language models. It pairs deterministic scene generation with exact answers computed from the same latent scene state, avoiding human-label ambiguity and enabling controlled analysis of both reasoning ability and reasoning stability.
+## Benchmark construction
 
-## Dataset summary
+Generation rejects scenes near decision boundaries, rounding ties, or classification thresholds and records the distance from each accepted scene to its nearest boundary. Difficulty is therefore controlled during generation rather than introduced through post-hoc annotation.
 
-| Property | Value |
-|---|---:|
-| Images | 100,000 |
-| Reasoning domains | 34 |
-| Reasoning families | 9 |
-| Closed-loop questions | 500,000 |
-| Open-response questions | 100,000 |
-| Total questions | 600,000 |
-| Language | English |
-| License | MIT |
+Thirty-three domains contain 3,000 images each, while Projectile Motion contains 1,000, giving 100,000 images in total. Each image has five closed-loop questions and one open-response question, for 600,000 questions.
 
-The nine reasoning families are **plane, solid, transformational, physical, topological, projective, analytic, optical, and inductive reasoning**.
+### Reasoning families
 
-## Benchmark design
+| Family | Domains | Included domains |
+|---|:--:|---|
+| Plane | 5 | Nested Squares, Nested Triangles, Nested Hexagons, Line Intersection, Angle Estimation |
+| Solid | 6 | Cube Nets, Cube Structure, 3D Combination, Orthographic, Polyhedra, Depth & Height |
+| Transformational | 6 | Rotation Matching, Symmetry Patterns, Fold & Punch, 2D Combination, Embedded Figures, Overlapping Circles |
+| Physical | 7 | Physical Stability, Gear Trains, Free-Body Diagrams, Projectile Motion, Laser & Mirrors, Clock Reading, Gauge Reading |
+| Topological | 3 | Surface Topology, Route Puzzles, Hex Pathfinding |
+| Projective | 2 | Occluded Patterns, Shadow Inference |
+| Analytic | 2 | Coordinate Geometry, Compass Bearings |
+| Optical | 2 | Optical Illusions, Impossible Objects |
+| Inductive | 1 | Progressive Matrices |
+| **Total** | **34** | |
 
-### Procedural scenes and exact ground truth
+### Five-question ladder
 
-Each image is rendered from a deterministic seeded program. The generator uses the same scene parameters to derive the answer, rejects samples near decision boundaries or rounding ties, and stores detailed scene annotations. This makes difficulty and correctness properties of generation rather than post-hoc human annotation.
+Each image is paired with five ordered closed-loop questions. Every question is evaluated in a fresh single-turn conversation containing one image and one question.
+
+| Position | Competence |
+|:--:|---|
+| Q1 | Direct observation, recognition, or counting |
+| Q2 | Spatial relation or classification |
+| Q3 | Measurement or multi-element inference |
+| Q4 | Multi-step integration or derived comparison |
+| Q5 | Counterfactual reasoning after modifying the scene |
 
 ### Dual-Loop Evaluation
 
-- **Closed loop:** five questions are asked in fresh single-turn conversations. They progress from direct observation through relational reasoning, measurement, multi-step integration, and counterfactual reasoning.
-- **Open loop:** a free-response question supports controlled re-elicitation under no-challenge controls, unsupported user assertions, and peer-model responses.
+- **Closed loop:** the five ordered questions are answered independently.
+- **Open loop:** the stored first answer is re-elicited under a no-challenge control, an unsupported assertion, or peer-model responses.
+
+Both loops share images, questions, and exact ground truth, enabling capability and stability to be evaluated on the same benchmark items.
 
-Because every target is known exactly, answer changes can be signed. A **correction** moves from an incorrect answer to the ground truth; a **capitulation** abandons a correct answer. GRIP therefore separates **rule absence** from **rule fragility**.
+## Dataset structure
 
-## Dataset configurations
+GRIP is published with a single `test` split and three configurations:
 
-| Config | Rows | Contents |
+| Configuration | Rows | Description |
 |---|---:|---|
-| `closed_loop` | 500,000 | Five question levels per image, prompts, answer formats, exact targets, and images |
-| `open_loop` | 100,000 | One open-response prompt per image with acceptance sets, tolerances, targets, and structured answers |
-| `scene_annotations` | 100,000 | Images, generation parameters, latent scene state, and domain-specific derived quantities |
+| `closed_loop` | 500,000 | Five closed-loop questions per image, with prompts, exact targets, answer formats, and images |
+| `open_loop` | 100,000 | One open-response item per image, including acceptance sets, tolerances, targets, and structured answer fields |
+| `scene_annotations` | 100,000 | Images, seeds, generation parameters, latent scene state, and domain-specific derived quantities |
+
+### Core fields
 
-All configurations use a single `test` split because GRIP is an evaluation benchmark. The repository also retains combined question-only CSV files and the original per-domain files for transparent inspection and reproducibility.
+`closed_loop` contains:
+
+- `dataset`, `dataset_version`: domain and generator version
+- `question_id`: stable question identifier
+- `task`: question position in the five-level ladder
+- `image_bytes`: embedded image data
+- `image`, `image_path`: image identifier and repository-relative path
+- `prompt`: question text
+- `groundtruth`: exact answer
+- `answer_format`: required response type
 
-## Loading the data
+`open_loop` contains `dataset`, `dataset_version`, `question_id`, `image`, `image_path`, `prompt`, `acceptance_set`, `tolerances`, `targets`, and domain-specific structured target columns.
 
-Install the core dependencies:
+`scene_annotations` contains the common domain, version, image, and seed identifiers together with domain-specific latent parameters and derived quantities. Fields that do not apply to a domain are null.
+
+## Loading
+
+Install the Hugging Face Datasets library:
 
 ```bash
-python -m pip install -r requirements.txt
+pip install datasets
 ```
 
-Load a configuration from the Hugging Face Hub:
+Load any configuration from the Hub:
 
 ```python
 from datasets import load_dataset
@@ -102,76 +150,48 @@ open_loop = load_dataset(
     "open_loop",
     split="test",
 )
-```
 
-For low-memory inspection, add `streaming=True`. A full repository clone requires Git LFS:
-
-```bash
-git lfs install
-git clone <repository-url>
+scene_annotations = load_dataset(
+    "bilaljawaid980/GRIP-Benchmark",
+    "scene_annotations",
+    split="test",
+)
 ```
 
-## Key fields
-
-### `closed_loop`
-
-- `dataset`, `dataset_version`: domain and generator version
-- `question_id`: stable question identifier
-- `task`: position in the five-level reasoning ladder
-- `image_bytes`: embedded image data
-- `image`, `image_path`: image identifiers and repository-relative paths
-- `prompt`: question text
-- `groundtruth`: exact answer
-- `answer_format`: required response type
+For low-memory inspection, pass `streaming=True`.
 
-### `open_loop`
+> Do not send answer-key fields to a model under evaluation. Ground-truth, answer-format, tolerance, and annotation fields are provided for local scoring and analysis.
 
-- `question_id`, `image`, `image_path`, `prompt`
-- `acceptance_set`: accepted structured answer representation
-- `tolerances`: explicit numeric tolerances where applicable
-- `targets`: visual entities or quantities that must be recovered
-- domain-specific structured target columns
+## Results at a glance
 
-### `scene_annotations`
+Sixteen models were evaluated: ten hosted models and six open-weight models.
 
-The common fields identify the domain, version, scene, image, and seed. Remaining fields store domain-specific latent parameters and derived quantities. Unused domain-specific columns are null.
+- **Capability:** hosted closed-loop accuracy ranges from 68.9% to 41.3%. After adjustment for constant-answer baselines, the range is 50.7% to 4.8%.
+- **Noise robustness:** additive monochrome Gaussian grain at σ ∈ {15, 25, 40} changes accuracy by no more than 6.1 percentage points for any hosted model.
+- **Challenge sensitivity:** across 32,270 challenges, control-adjusted capitulation ranges from 6.18 to 72.75 percentage points.
+- **Peer effects:** adjusted correction is positive across all ten hosted models in the natural-peer experiment, while constructed consensus produces substantially larger capitulation effects.
+- **Family variation:** hosted family means range from 62.8% for Physical reasoning to 39.5% for Topological reasoning.
 
-## Evaluation protocols and results
+Detailed results are available in the repository:
 
-| Test | Purpose | Results |
+| Test | Evaluation | Results |
 |---|---|---|
-| Test 1 | Closed-loop accuracy by question level and reasoning family | [TEST_1_RESULTS.md](tests/test_1_closed_loop/results/TEST_1_RESULTS.md) |
-| Test 2 | Robustness to controlled visual noise | [TEST_2_RESULTS.md](tests/test_2_grain_robustness/results/TEST_2_RESULTS.md) |
-| Test 3 | Corrections and capitulations under unsupported challenge | [TEST_3_RESULTS.md](tests/test_3_sycophancy/results/TEST_3_RESULTS.md) |
-| Test 4 | Natural-peer, constructed-consensus, control, and confidence effects | [TEST_4_RESULTS.md](tests/test_4_crossexam/results/TEST_4_RESULTS.md) |
-
-The evaluation implementation is in [`grip_eval/`](grip_eval/), with experiment-specific scripts and archived outputs under [`tests/`](tests/).
-
-## Repository structure
-
-```text
-Dataset/       Per-domain images, annotations, questions, validators, and generators
-combined/      Hub-ready closed-loop, open-loop, and scene-annotation tables
-grip_eval/     Shared parsing and evaluation utilities
-tests/         Protocols, scripts, results, and archived model responses
-analysis/      Aggregate analyses and audit outputs
-figures/       Paper figures and selected qualitative samples
-scripts/       Repository-level validation and maintenance utilities
-```
-
-## Intended use and limitations
+| Test 1 | Closed-loop accuracy by level and reasoning family | [Results](tests/test_1_closed_loop/results/TEST_1_RESULTS.md) |
+| Test 2 | Robustness to controlled visual noise | [Results](tests/test_2_grain_robustness/results/TEST_2_RESULTS.md) |
+| Test 3 | Correction and capitulation under unsupported challenge | [Results](tests/test_3_sycophancy/results/TEST_3_RESULTS.md) |
+| Test 4 | Natural-peer, constructed-consensus, control, and confidence effects | [Results](tests/test_4_crossexam/results/TEST_4_RESULTS.md) |
 
-GRIP is intended for research on vision-language reasoning, robustness, calibration, and interaction-dependent answer stability. It is suitable for evaluation, error analysis, and controlled comparison of multimodal systems.
+## Intended use
 
-The benchmark is synthetic and should not be treated as evidence of real-world physical competence. Performance may depend on prompting, image preprocessing, decoding settings, and answer parsing. The domains test selected geometric and physical abstractions; they do not cover every form of perception, commonsense reasoning, or embodied interaction. Researchers should report model versions, prompts, image transforms, decoding parameters, and scoring rules.
+GRIP is an evaluation benchmark for research on vision-language reasoning, robustness, and interaction-dependent answer stability. It supports controlled model comparison, error analysis, and reproducible evaluation across geometric, spatial, and physical reasoning tasks.
 
-All scenes are procedurally generated and contain no people, personal information, or scraped user content.
+GRIP uses synthetic diagrams and evaluates the abstractions represented by its 34 domains. Results should be reported with the model version, prompts, image preprocessing, decoding settings, and scoring procedure.
 
 ## Reproducibility
 
-Per-domain directories include generator versions, build manifests, validation reports, and exact answer keys. Closed-loop question construction is documented in [`UNIFIED_5_LEVEL_GENERATION_TEMPLATE.md`](UNIFIED_5_LEVEL_GENERATION_TEMPLATE.md); open-response construction is documented in [`OPEN_QUESTION_SPEC.md`](OPEN_QUESTION_SPEC.md).
+Per-domain directories include generator versions, build manifests, validation reports, and exact answer keys. Closed-loop question construction is documented in [`UNIFIED_5_LEVEL_GENERATION_TEMPLATE.md`](UNIFIED_5_LEVEL_GENERATION_TEMPLATE.md), and open-response construction is documented in [`OPEN_QUESTION_SPEC.md`](OPEN_QUESTION_SPEC.md).
 
-Use immutable commit revisions when reporting results. Do not expose answer keys to a model during evaluation.
+Use immutable repository revisions when reporting results.
 
 ## License
 
@@ -179,14 +199,16 @@ GRIP, including its procedurally generated data and accompanying code, is releas
 
 ## Citation
 
-If you use GRIP, cite the accompanying preprint:
-
 ```bibtex
 @misc{jawaid2026geometric,
-  author = {Bilal Jawaid and Hamza Hanif and Muhammad Maaz Uddin and Anas Zafar and Inad Alqurashi and Aizan Zafar and Osama Farooqui and Sheheryar Khan and Ferhat Sadak and Shaina Raza and Rizwan Qureshi},
-  title = {Geometric Parrots: Rule Absence vs. Rule Fragility in Spatial and Physical Reasoning for Vision-Language Models},
-  year = {2026},
-  note = {arXiv preprint}
+  author = {Bilal Jawaid and Hamza Hanif and Muhammad Maaz Uddin and
+            Anas Zafar and Inad Alqurashi and Aizan Zafar and
+            Osama Farooqui and Sheheryar Khan and Ferhat Sadak and
+            Shaina Raza and Rizwan Qureshi},
+  title  = {Geometric Parrots: Rule Absence vs. Rule Fragility in Spatial
+            and Physical Reasoning for Vision-Language Models},
+  year   = {2026},
+  note   = {Preprint}
 }
 ```
 
@@ -204,6 +226,3 @@ If you use GRIP, cite the accompanying preprint:
 10. Shaina Raza — Vector Institute
 11. Rizwan Qureshi — Massachusetts General Hospital
 
-## Contributing and security
-
-See [CONTRIBUTING.md](CONTRIBUTING.md) for contribution and validation requirements, [CODE_OF_CONDUCT.md](CODE_OF_CONDUCT.md) for community standards, and [SECURITY.md](SECURITY.md) for private vulnerability reporting guidance.