CycliST: A Video Language Model Benchmark for Reasoning on Cyclical State Transitions

CycliST is a synthetic, diagnostic benchmark for evaluating Video Language Models (VLMs) on their ability to reason over cyclical state transitions; periodic patterns in object motion and visual attributes. Published in the Journal of Data-centric Machine Learning Research (DMLR), 2026.

Dataset Summary

Cyclical patterns are everywhere in the physical world, from traffic lights and orbiting satellites to heartbeats and conveyor belts. Yet existing video-reasoning benchmarks largely capture linear or causal structure and rarely test whether a model can detect, track, and exploit periodic dynamics. CycliST is built to fill this gap.

Inspired by the diagnostic tradition of CLEVR, CycliST renders short, richly annotated video sequences in which objects undergo smooth, periodic changes and always return to each configuration at regular intervals. Each video is paired with template-generated question–answer pairs and complete ground-truth scene metadata, enabling fine-grained evaluation of spatio-temporal reasoning.

• 14,800 Full-HD videos (1920×1080), rendered at 32 fps, 5 seconds / 160 frames each
• ~120k template-generated question–answer pairs
• 5 difficulty tiers varying the number of cyclic objects, scene clutter, and lighting
• Physically based rendering via Blender (Cycles engine)
• Complete per-frame ground truth (positions, scale, rotation, color, spatial relations)

In the accompanying experiments, state-of-the-art open-source and proprietary VLMs (Intern, LLaVA-Video, LLaVA-OneVision, and Gemini families, 7B–78B) struggle to reliably recognize cyclic patterns, count objects in motion, or estimate periodicity, revealing a significant gap in current temporal reasoning capabilities.

Sample Videos

One example per difficulty tier. The full-resolution 1920×1080 renders live under videos//.

L1 – Unicycle One cyclic object, 2–3 clutter objects	L2 – Unicycle-Cluttered One cyclic object, 4–9 clutter objects	L3 – Bicycle Two cyclic objects
L4 – Tricycle Three cyclic objects	L5 – Nightrider Adds a scene-wide periodic light cycle

Supported Tasks

• Video Question Answering (VQA): Free-form answers to template-generated questions, evaluated with an LLM judge (Llama3-70B) for robustness to phrasing.
• Scene Understanding / Captioning: Describing all objects, their attributes, and their cyclic transitions, scored against ground-truth scene graphs (precision / recall / F1).

Dataset Structure

The repository is organized into three top-level folders:

Folder	Contents
videos/	Rendered Full-HD .mp4 video sequences (160 frames @ 32 fps).
scenes/	Per-scene ground-truth metadata (JSON): object attributes and the full temporal evolution of positions, scale, rotation, color, and spatial relations.
questions/	Template-generated question–answer pairs (JSON), organized by question template and tier.

Splits

The dataset is divided into train, validation, and test splits (videos):

Tier	Train	Test	Validation	Total
L1 – Unicycle	1,500	750	750	3,000
L2 – Unicycle-Cluttered	1,500	750	750	3,000
L3 – Bicycle	1,500	750	750	3,000
L4 – Tricycle	1,540	770	770	3,080
L5 – Nightrider	1,360	680	680	2,720
Total	7,400	3,700	3,700	14,800

Difficulty Tiers

Each tier changes one aspect of scene complexity:

• L1 – Unicycle: Exactly one cyclic object with 2–3 clutter objects. Tests fundamental perception of a single periodic motion or attribute change in isolation.
• L2 – Unicycle-Cluttered: One cyclic object amidst 4–9 clutter objects, raising the difficulty of isolating the relevant entity.
• L3 – Bicycle: Two cyclic objects (plus 2–3 clutter objects), introducing interacting cycles and relative-phase reasoning.
• L4 – Tricycle: Three cyclic objects, further escalating spatio-temporal complexity.
• L5 – Nightrider: Adds a scene-wide periodic light cycle (smoothly interpolating between bright and dark) on top of a balanced mix of L1/L3/L4-style scenes.

Cycle Types

Objects evolve via cycle functions that always return to their initial state after one full period:

• Motion cycles (change position): linear (back-and-forth between two points) and orbiting (circular trajectory around a center object, which may itself be moving).
• Attribute cycles (change appearance): size (small ↔ large), color (continuous hue interpolation), and orientation (continuous rotation; omitted for rotation-invariant shapes such as spheres).
• Light cycles (scene-wide): sinusoidal modulation of light intensity (used in the Nightrider tier).

Question Categories

Questions are produced with a template-based functional program executed over each scene's ground-truth graph, following the CLEVR/CLEVRER tradition and extended with novel temporal and cyclical operators.

Temporal Descriptive (balanced yes/no, random baseline ≈ 50%), each with an existential (∃, "ever true") and a universal (∀, "always true") quantifier:

• Query – probe an attribute of a single object over time.
• Compare – compare an attribute between two objects.
• Relate – test a spatial relationship between two objects.

Scene Representative (random baseline ≈ 30% where multi-valued):

• Cyclic – orbit (which object is the orbit center), orbit direction (clockwise / counterclockwise), and initial / transition attribute values.
• Numeric – counting (number of cyclic objects), periodicity (frames per cycle), and occurrence (number of completed cycles).

The full set of question templates, placeholders, and functional operators is documented in the appendix of the paper.

Usage

The data is distributed as raw video, scene, and question files. Because the splits span multiple per-template JSON files, the easiest way to start is to clone the repository and read the files directly:

# Requires git-lfs (the videos are large: ~15.6 GB total)
git lfs install
git clone https://huggingface.co/datasets/AIML-TUDA/CycliST

Or download a subset with the Hub API:

from huggingface_hub import snapshot_download

# Download only the question and scene metadata (skip the large video files)
snapshot_download(
    repo_id="AIML-TUDA/CycliST",
    repo_type="dataset",
    allow_patterns=["questions/*", "scenes/*"],
)

Refer to the GitHub repository for the render pipeline, question-generation scripts, dataset-split definitions, and the LLM-judge calibration scripts used in the paper.

Data Generation

Scenes are generated procedurally and validated incrementally with a backtracking placement mechanism that enforces margins to scene boundaries and between objects across all frames. Valid scenes are rendered in Blender using the Cycles engine for photorealistic, physically based rendering, with keyframed transformations and built-in interpolation. Each scene yields a 1920×1080 video at 32 fps plus a JSON file recording the complete temporal metadata (position, scale, rotation, color over time) and spatial relationships used to generate the VQA tasks. Lighting and camera setup follow the CLEVR configuration with randomized translations for visual diversity.

Evaluation

Because VLMs produce free-form text, answers are scored with an LLM judge (Llama3-70B), calibrated on 100 questions per pipeline. Reported judge alignment with human annotation is 100% for yes/no and numeric answers, 92.6% for attribute answers, and an F1 of 87.6% for object mapping in scene understanding. Indefinite answers are counted as incorrect for accuracy and excluded from Mean Absolute Error (MAE). The paper reports temporal-descriptive accuracy, orbit/transition accuracy, cycle counting (accuracy and MAE), periodicity estimation (MAE), and scene/cycle captioning (precision, recall, F1).

Limitations

• The dataset is entirely synthetic and does not capture the full nuance of real-world scenes.
• Cycles use stationary frequencies; real-world cyclicity is often non-stationary (e.g., varying day–night intervals).
• Object geometry and material are fixed over time, and causal events between objects (e.g., a traffic light affecting vehicle flow) are not modeled.
• The experiments characterize benchmark-level performance gaps rather than reverse-engineering individual models' internal failure modes.

Licensing

Released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Citation

@article{kohaut2026cyclist,
  title   = {CycliST: A Video Language Model Benchmark for Reasoning on Cyclical State Transitions},
  author  = {Simon Kohaut and Daniel Ochs and Shun Zhang and Benedict Flade and Julian Eggert and Kristian Kersting and Devendra Singh Dhami},
  journal = {Journal of Data-centric Machine Learning Research},
  year    = {2026},
  url     = {https://openreview.net/forum?id=l03g53HUL2}
}

Authors

Simon Kohaut¹²*, Daniel Ochs¹*, Shun Zhang¹, Benedict Flade³, Julian Eggert³, Kristian Kersting¹, Devendra Singh Dhami⁴  (*equal contribution)

• ¹ Artificial Intelligence and Machine Learning Lab, TU Darmstadt
• ² Konrad Zuse School of Excellence in Learning and Intelligent Systems (ELIZA)
• ³ Honda Research Institute Europe GmbH
• ⁴ Uncertainty in Artificial Intelligence Group, TU Eindhoven