MicroVerse: A Preliminary Exploration Toward a Micro-World Simulation

Rongsheng Wang; Minghao Wu; Hongru Zhou; Zhihan Yu; Zhenyang Cai; Junying Chen; Benyou Wang

cs.AIFeb 28, 2026

MicroVerse: A Preliminary Exploration Toward a Micro-World Simulation

Authors: Rongsheng Wang, Minghao Wu, Hongru Zhou, Zhihan Yu, Zhenyang Cai, Junying Chen, Benyou Wang

Organizations: The Chinese University of Hong Kong, Shenzhen · Shenzhen Loop Area Institute · Peking Union Medical College Hospital

Abstract

Recent advances in video generation have opened new avenues for macroscopic simulation of complex dynamic systems, but their application to microscopic phenomena remains largely unexplored. Microscale simulation holds great promise for biomedical applications such as drug discovery, organ-on-chip systems, and disease mechanism studies, while also showing potential in education and interactive visualization. In this work, we introduce MicroWorldBench, a multi-level rubric-based benchmark for microscale simulation tasks. MicroWorldBench enables systematic, rubric-based evaluation through 459 unique expert-annotated criteria spanning multiple microscale simulation task (e.g., organ-level processes, cellular dynamics, and subcellular molecular interactions) and evaluation dimensions (e.g., scientific fidelity, visual quality, instruction following). MicroWorldBench reveals that current SOTA video generation models fail in microscale simulation, showing violations of physical laws, temporal inconsistency, and misalignment with expert criteria. To address these limitations, we construct MicroSim-10K, a high-quality, expert-verified simulation dataset. Leveraging this dataset, we train MicroVerse, a video generation model tailored for microscale simulation. MicroVerse can accurately reproduce complex microscale mechanism. Our work first introduce the concept of Micro-World Simulation and present a proof of concept, paving the way for applications in biology, education, and scientific visualization. Our work demonstrates the potential of educational microscale simulations of biological mechanisms. Our data and code are publicly available at https://github.com/FreedomIntelligence/MicroVerse

Figures & tables

Figure 1: Failure cases of Sora and Veo3 on Microscale Simulation. Although Sora and Veo3 generate results that appear visually correct, their violations of physical laws are particularly evident.

Figure 2: Illustration of MicroWorldBench Evaluation Process. A MicroWorldBench example consists of a generated microscopic video and a set of task-specific evaluation criteria written by experts. A MLLM-based scoring system rates responses according to each criterion.

Open-Source Video Generation Models
Model	Average $\uparrow$	Organ-level $\uparrow$	Cellular-level $\uparrow$	Subcellular-level $\uparrow$
HunyuanVideo	23.2	23.1	23.8	19.4
CogVideoX-5B	43.5	39.9	47.0	38.6
Wan2.1-T2V-1.3B	49.4	45.9	51.7	52.4
Wan2.2-TI2V-5B	51.6	46.6	53.9	49.5
Wan2.1-T2V-14B	54.8	55.7	54.4	52.8

Table 1: Performance comparison of different video generation models on MicroWorldBench. Bold indicates the best performance.

Open-Source Video Generation Models
Model	Average $\uparrow$	Scientific Fidelity $\uparrow$	Visual Quality $\uparrow$	Instruction Following $\uparrow$
HunyuanVideo	23.2	15.6	48.2	23.4
CogVideoX-5B	43.5	37.4	64.1	38.6
Wan2.1-T2V-1.3B	49.4	40.3	71.8	50.1
Wan2.2-TI2V-5B	51.6	40.7	82.7	47.0
Wan2.1-T2V-14B	54.8	42.7	86.0	53.8

Table 2: Performance comparison of different video generation models on MicroWorldBench (dimension-wise scores). Bold indicates the best performance.

Figure 3: Overview of our data filtering pipeline. Each stage applies specific filters and shows the volume of data removed and retained.

Panel of Figure 4 — Figure 4: Distributions of fundamental video attributes in the MicroSim-10K.

Panel of Figure 5 — Figure 5: Distributions of video popularity indicators in the MicroSim-10K.

Data	FVD vs. MicroSim-10K $\downarrow$	FVD vs. 643-Real-Biological-Clips $\downarrow$
MicroSim-10K	0	123.9
643-Real-Biological-Clips	123.9	0
Commercial Models
Veo3	42.6	118.1
Sora	116.9	136.3
Open-Source Models

Table 3: FVD comparison across models (lower is better). The FVD between MicroSim-10K and real biological videos is 123.9, indicating a close distributional alignment.

Base Model: Wan2.1-1.3B
	Data Size	Scientific Fidelity $\uparrow$	Visual Quality $\uparrow$	Instruction Following $\uparrow$
Baseline	0	40.3	71.8	50.1
FT on MicroSim-10K	9,601	43.0	68.5	49.3
FT on General + MicroSim-10K	19,202	44.1 (+3.8)	74.4 (+2.6)	53.8 (+3.7)
Base Model: Wan2.1-14B (Full-Blooded Model)
Baseline	0	42.7	86.0	53.8

Table 4: Impact of Data Scale and Mixed-Domain Training on MicroVerse Performance. Bold indicates the best performance. FT indicates fine-tuning.

Panel of Figure 6 — Figure 6: Human Evaluation and Consistency Results.

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Overview of the data construction process.

Deduplication Type	Method	Metric
Source-level	Video ID matching	# overlapping IDs
Text-level	Caption embedding	# pairs $>0.90$
Vision-level	Frame embedding	# pairs $>0.95$

Appendix

Table 5: Three-level deduplication pipeline results.

Biological Level	Original	Post-Deduplication	Removed
Organ-level	238	233	-5
Cellular-level	189	182	-7
Subcellular-level	32	30	-2
Total	459	445	-14

Appendix

Table 6: Task distribution after deduplication.

Test Set	Average	Scientific Fidelity	Visual Quality	Instruction Following
Original (459 tasks)	50.2	43.0	68.5	49.3
Clean (445 tasks)	50.0	42.7	68.5	49.1

Appendix

Table 7: Performance comparison on original vs. clean test set.

Variant	Data Size	Sci. Fidelity	Vis. Quality	Inst. Following
Baseline Model (Wan2.1-1.3B)	0	40.3	71.8	50.1
FT on raw (uncleaned) data	34,318	42.1	67.4	44.5
FT on MicroSim-10K	9,601	43.0	68.5	49.3

Appendix

Table 8: Ablation study on dataset filtering.

Variant	Data Size	Sci. Fidelity	Vis. Quality	Inst. Following
Baseline Model (Wan2.1-1.3B)	0	40.3	71.8	50.1
FT on MicroSim-10K (50% sample)	4,800	41.9	69.1	48.4
FT on MicroSim-10K	9,601	43.0	68.5	49.3

Appendix

Table 9: Ablation study on dataset size.

Variant	Data Size	Sci. Fidelity	Vis. Quality	Inst. Following
Baseline Model (Wan2.1-1.3B)	0	40.3	71.8	50.1
FT on MicroSim-10K (CFG=5%)	9,601	41.5	70.2	49.8
FT on MicroSim-10K (CFG=10%)	9,601	43.0	68.5	49.3
FT on MicroSim-10K (CFG=15%)	9,601	38.2	63.1	51.4

Appendix

Table 10: Ablation study on CFG rate.

Expert ID	Field of Expertise	Years of Experience
1	Cell Biology/Genetics	12 years
2	Biochemistry / Metabolic Pathway Modeling	10 years
3	Biochemistry / Metabolic Pathway Modeling	8 years

Appendix

Table 11: Background information of the experts involved in the evaluation.

Panel of Figure 8 — Figure 8: Actions Log of Three Independent Experts.

Action	Expert 1	Expert 2	Expert 3
Adjusting weights	180	221	189
Filtering	137	128	146
No Action	94	82	95
Supplementing	48	28	29

Appendix

Table 12: Detailed expert actions during the rubric revision process.

Biological Level	Number of Tasks	Total Expert Actions	Avg. Actions per Task
Organ-level	238	361	1.52
Cellular-level	189	567	3.00
Subcellular-level	32	176	5.50

Appendix

Table 13: Expert intervention frequency across different biological scales.

Evaluation Group	Agreement Score
Human-only (Fleiss’ Kappa)	0.733
Human + GPT-5 (Fleiss’ Kappa)	0.771
GPT-5 vs Gemini-2.5-pro (Spearman)	0.835
GPT-5 vs GPT-5 (Spearman)	0.912

Appendix

Table 14: Scoring agreement between human experts and LLMs.

Level	Agreement (Human-only)	Agreement (Human + GPT-5)
Organ-level	0.728	0.742
Cellular-level	0.740	0.739
Subcellular-level	0.691	0.727

Appendix

Table 15: Agreement across biological scales.

Parameter	Value
--train_batch_size	8
--max_train_steps	5000
--learning_rate	1e-5
--mixed_precision	bf16
--training_cfg_rate	0.1
--num_height	480

Appendix

Table 16: Detailed hyperparameters used during the training phase of MicroVerse.

Parameter	Value
--height	480
--width	832
--num_frames	81
--guidance_scale	5.0
--num_inference_steps	50

Appendix

Table 17: Standardized inference parameter settings for model evaluation.

Comparison Pair	Cohen’s Kappa
Expert1 vs. Expert2	0.87
Expert1 vs. Expert3	0.83
Expert2 vs. Expert3	0.81

Appendix

Table 18: Comparison of Cohen’s Kappa values between experts.

Figure 9: Example of a cell mitosis simulation video frame, taken from an excellent example video on YouTube.

Dimension	Criteria	Weight
Scientific Fidelity	The sequence of stages, segregation patterns, and ploidy changes in mitosis and meiosis are accurately represented; mitosis produces two genetically identical diploid daughter cells, whereas meiosis involves two successive divisions resulting in four genetically diverse haploid gametes.	+ 1.0
	The alignment of chromosomes at the metaphase plate and their segregation from metaphase to anaphase occur correctly; sister chromatids are distinctly differentiated from homologous chromosomes; the attachment of kinetochores to spindle microtubules and the direction of their tension conform to biological principles.	+ 1.0
	Structural details and dynamic coordination between the spindle apparatus and the centrosome (centriole) are accurate; spindle pole positioning, microtubule polarity, and force distribution are appropriate; the relationship between the microtubule-organizing center and cell polarity is correctly established.	+ 1.0
	The timing and mechanisms of DNA replication and genetic recombination are correctly presented; DNA replication occurs during the pre-mitotic S phase, homologous pairing and crossing over take place in prophase I of meiosis, and no DNA replication occurs between meiosis I and II.	+ 1.0
Visual Quality	The image demonstrates high clarity and fine presentation of microstructural details, with sharp edges of subcellular structures, well-defined layer separation, and absence of wax artifact noise.	+ 0.5
	The animation exhibits coherent motion with stable temporal rhythm, smooth phase transitions, and natural movement trajectories, without any stuttering or tearing.	+ 0.5

Appendix

Table 19: Rubric Example for Cell Mitosis Evaluation

Figure 10: Example of real biological video clips.

Category	Total	Train (80%)	Test (20%)
Microsimulation-related	1107	885	222
Non-microsimulation	1473	1178	295
Total	2580	2063	517

Appendix

Table 20: Dataset statistics for microsimulation classification.

Video Generation

Video generation studies models that create and edit moving images from text, images, audio, and structured controls.

See all 364 papers

Nano World Models: A Minimalist Implementation of Future Video Prediction
May 17, 2026Siqiao Huang, Partha Kaushik, Michael Chen +5Action-Conditioned Video Generation Long-Horizon Video Generation
Tsinghua University · Carnegie Mellon University · University of Bristol +2
MiniWorld: Democratizing the Training of Video World Models from Scratch
Aug 2, 2026Yian Zhao, Ruochong Zheng, Hongcan Guo +3Video Diffusion Models Streaming Video Generation
Peking University
Sci-VBench: Evaluating Knowledge- and Reasoning-Intensive Video Generation in Science Domains
Aug 10, 2026Diandian Zhang, Tingyu Song, Lin Fu +2VLM Evaluation Video Generation
Zhejiang University · UCAS · Tongji University +1

May 17, 2026J/K move · Enter open · S save

Nano World Models: A Minimalist Implementation of Future Video Prediction

Siqiao Huang, Partha Kaushik, Michael Chen, Hengkai Pan+4

Tsinghua University · Carnegie Mellon University · University of Bristol+2

Action-Conditioned Video Generation Long-Horizon Video Generation Video Prediction Video World Models

Open paper AskPDF arXiv

World models have become a central paradigm for learning predictive simulators that support generation, planning, and decision-making. Yet, despite rapid progress in industry-scale interactive video generation, the broader research community still lacks compact, reproducible, and easily extensible implementations for studying the design choices underlying modern world models. We introduce Nano World Models, a minimalist codebase for future video prediction centered around diffusion forcing. Nano World Models provides a unified interface for generative objectives, model scales, action-conditioning mechanisms, latent observation spaces, datasets, evaluation protocols, and long-horizon rollout procedures. This design enables controlled studies of world-modeling components that are often entangled across separate implementations. Through experiments across simple control environments, game simulation, and real-robot data, we examine how prediction parameterization, architecture scale, action injection, sampling budget, and domain complexity affect video prediction quality and autoregressive rollout behavior. By releasing code, configurations, evaluation scripts, and pretrained checkpoints, Nano World Models aims to provide a compact yet extensible experimental substrate for open, reproducible, and scientific world-model research.

Abstract

Video Generation

Explore similar work

Nano World Models: A Minimalist Implementation of Future Video Prediction

MiniWorld: Democratizing the Training of Video World Models from Scratch

Sci-VBench: Evaluating Knowledge- and Reasoning-Intensive Video Generation in Science Domains