PaperlayerView exact revision on arXiv

Source-linked AI summary

MicroVerse: A Preliminary Exploration Toward a Micro-World Simulation

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

arXiv:2603.00585v1cs.AIcs.CV

TL;DR

Video-generation models have advanced macroscopic simulation but remain unreliable for physically and biologically meaningful microscale dynamics. The paper introduces MicroWorldBench, MicroSim-10K, and MicroVerse to evaluate and improve microscale simulation, with MicroVerse showing improved scientific fidelity and consistency. The work presents a proof of concept for microscale simulation in biology, education, and scientific visualization.

  • Problem

    Video-generation research has focused mainly on macroscopic scenes, while existing models struggle to produce physically plausible and biologically meaningful microscale dynamics.

  • Method

    The paper constructs MicroWorldBench with expert-defined rubrics, builds the expert-verified MicroSim-10K dataset, and fine-tunes MicroVerse for microscale simulation.

  • Results

    Current models often maintain visual coherence but perform poorly on microscale biological plausibility, while MicroVerse surpasses the original model by more than +2.7 in scientific fidelity.

  • Takeaways & Limitations

    The proof of concept supports educational microscale simulations and potential applications in biology, biomedical research, and interactive scientific visualization.

  • Takeaways & Limitations

    The approach does not explicitly incorporate governing physical laws such as fluid mechanics, diffusion–reaction equations, or biomechanical constraints, limiting high-precision scientific simulation and prediction.

Abstract

from arXiv · show

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

1 INTRODUCTION

MicroWorldBench exposes that video-generation models, despite strong visual coherence, struggle to produce physically and biologically plausible microscale dynamics. MicroVerse addresses this gap with expert-verified microscale data and a tailored generation model.

  • MicroWorldBench: Across models, visual quality remains high while scientific fidelity is substantially weaker, revealing videos that look correct but violate physical and biological laws.Figure 1 highlights evident physical-law violations in outputs from Sora and Veo3 despite their visually correct appearance.
  • MicroWorldBench: MicroWorldBench evaluates 459 microscale simulation tasks spanning organ-level, cellular, and subcellular processes with task-specific expert criteria.The benchmark combines generated microscopic videos with evaluation criteria targeting scientific fidelity, visual quality, and instruction following.
  • MicroWorldBench: Current video-generation models often appear visually coherent and prompt-aligned but fail to generate biologically plausible microscale dynamics.These failures are attributed to insufficient grounding in microphysical principles and knowledge from predominantly human-scale training videos.
  • MicroVerse: MicroVerse is a microscale-focused video-generation model trained on MicroSim-10K, an expert-verified dataset containing 9,601 microscale scenarios.The dataset emphasizes physical plausibility and biological fidelity across diverse microscale mechanisms.
  • MicroVerse: MicroVerse surpasses the original model by more than +2.7 in scientific fidelity and improves scientific constraints alongside temporal and spatial consistency.The model is built upon Wan2.1 and fine-tuned using MicroSim-10K.

2 MICROWORLDBENCH: A RUBRIC-BASED BENCHMARK FOR MICROSCALE SIMULATION

MicroWorldBench is a rubric-based benchmark designed to evaluate microscale video simulation across organ, cellular, and subcellular processes. It combines expert-filtered tasks with criteria emphasizing scientific fidelity, then reveals that current models often produce visually convincing but scientifically weak simulations.

  • 2.1 TASK CHOICE: The benchmark samples three representative biological scales because practicality, impact, and data availability constrain broader hierarchical coverage.The selected hierarchy spans organ, cellular, and subcellular processes without discarding existing scientific frameworks.
  • 2.1 TASK CHOICE: 459 tasks cover organ-level, cellular-level, and subcellular-level simulations selected for diversity and practical relevance.The final suite contains 238 organ-level, 189 cellular-level, and 32 subcellular-level tasks.
  • 2.2 PROMPT SUITE: Over 8,000 YouTube videos yielded 8,162 generated tasks, which were filtered by scale diversity and agreement from at least two of three biology experts.Tasks were retained only when they satisfied both the diversity and practical-relevance criteria.
  • 2.3 RUBRIC CRITERIA: Each example pairs a task instruction with expert-refined rubric criteria evaluating scientific fidelity, visual quality, and instruction following.An LLM-based grader assigns standardized scores, with scientific fidelity emphasizing mechanistic accuracy over visual realism.
  • 2.3 RUBRIC CRITERIA: Criteria encode an evaluation dimension, description, polarity, and weight, with core scientific requirements receiving higher weights than auxiliary presentation criteria.The stated example weights are wi = 1.0 for core requirements, wi = 0.5 for secondary requirements, and wi = 0.2 for auxiliary or presentational criteria.
  • 2.3 RUBRIC CRITERIA: Normalized rubric scores cap performance at 100 and prevent minor positive criteria from offsetting severe scientific errors.The normalization is intended to make scores comparable across tasks.
  • 2.4 EVALUATION RESULTS AND ANALYSIS: Current models often score highly on visual quality but substantially lower on scientific fidelity, producing videos that look plausible without obeying biological and physical laws.Visual-quality scores are reported at 80–97, while most open-source models score 15–43 in scientific fidelity.
  • 2.4 EVALUATION RESULTS AND ANALYSIS: Cellular and subcellular tasks are harder than organ-level tasks, and scaling Wan models from 1.3B to 14B mainly improves visual quality rather than scientific fidelity.The paper attributes the hierarchical gap to stricter consistency requirements and scarce microscale training data, while concluding that parameter expansion alone is insufficient.

3 MICROVERSE: TOWARD MICROSCALE SIMULATION VIA A EXPERT-VERIFIED DATASET

MicroVerse is developed from an expert-verified microscale video dataset to address limitations in modeling physically and biologically grounded microscopic mechanisms. The dataset combines filtered, captioned clips with realism checks, while training uses latent diffusion with text conditioning.

  • Motivation: MicroWorldBench identifies that existing models remain limited in modeling microscale mechanisms governed by physical and biological principles.This motivates domain-specific data and a model explicitly incorporating physical grounding and fine-grained biological dynamics.
  • Data Construction: The dataset is constructed by collecting YouTube videos, segmenting them into short clips, automatically filtering them, and applying expert review for physical consistency.The pipeline uses resolution and licensing filters, OpenCLIP-based segmentation, VideoMAE classification, OpenCV and EasyOCR filtering, and expert removal of meaningless or inconsistent clips.
  • MicroSim-10K Dataset: MicroSim-10K contains 9,601 expert-verified microscale video clips spanning organ, cellular, and subcellular mechanisms.Each clip is at least 720p and 5–60 seconds long, with detailed captions averaging around 150 words.
  • Dataset Metadata: MicroSim-10K retains views, likes, and comments to preserve the educational and communicative value of microscale simulations.The dataset’s videos are reported to have broad public accessibility and substantial engagement.
  • Realism and Distribution: MicroSim-10K lies close to real microscopy video distributions, with an FVD of 123.9 against real biological videos.The comparison uses 377 collected biological videos yielding 643 clips after preprocessing; lower FVD indicates closer alignment.
  • Training MicroVerse: MicroVerse is trained by fine-tuning Wan2.1 with a denoising diffusion objective over video latents conditioned on CLIP text embeddings.The model predicts noise at each diffusion timestep, and 10% text-conditioning masking enables classifier-free guidance training.

4 EXPERIMENTS

Experiments evaluate MicroVerse against video generation models using MicroWorldBench, human preferences, and consistency analyses. MicroVerse improves scientific fidelity, especially on subcellular tasks, while scaling model capacity and data diversity improves performance across dimensions.

  • Experiment Settings: MicroVerse is trained by fully fine-tuning Wan2.1-T2V-1.3B on text-conditioned microsimulation video generation.Training uses 8 NVIDIA H200 GPUs, a learning rate of 1e-5, and batch size 8.
  • Human Evaluation: Human evaluation compares MicroVerse with Sora and Veo3 across 60 samples spanning organ-level, cellular, and subcellular microsimulation.Three evaluators selected preferred outputs using instruction fidelity and visual clarity, or marked a tie.
  • Results: 43.0 Scientific Fidelity: MicroVerse outperforms all open-source models, while Visual Quality reaches 68.5 and Instruction Following reaches 49.3.The scientific-fidelity improvement is attributed to training on the physics-grounded MicroSim-10K dataset.
  • Results: 53.3 on subcellular-level tasks: MicroVerse surpasses all open-source models on this challenging task category.The authors associate this progress with the microscale simulation dataset.
  • Scaling Results: Scaling to 14B parameters and mixing MicroSim-10K with equal high-quality general-domain data significantly improves performance across all dimensions.The scaled configuration achieves state-of-the-art results among open-source models.
  • Evaluation Consistency: Cohen’s Kappa above 0.80 indicates strong agreement among the three independent experts, supporting the reliability of the scoring process.MicroWorldBench also evaluates consistency between models and humans using human preference labeling.

5 RELATED WORK

Related work covers world models, video generation, and rubric-based evaluation. These lines of research motivate simulation, scalable video modeling, and interpretable task-specific assessment.

  • World Model: World models simulate dynamic environments by predicting future states and estimating rewards from observations and actions.Their state-transition modeling has been extended to robotics and other real-world scenarios.
  • Video Generation: Video generation has rapidly improved in quality, coherence, and controllability, with systems such as Sora, Veo3, Kling, HunyuanVideo, and Hailuo.These systems are widely applied in video production, advertising, and education.
  • Rubric Evaluation: Rubric-based evaluation provides task-specific, interpretable criteria that improve grading consistency for open-ended tasks.Prior work extends rubric evaluation to large-scale clinical conversations and dynamically generated case-specific reward signals.

6 CONCLUSIONS

The paper frames microscale video generation as a gap left by progress on natural and human-centered macroscopic scenes. It combines expert evaluation, physics-oriented data, and a tailored model to pursue biologically meaningful simulation.

  • Conclusion: MicroWorldBench provides 459 expert-curated tasks with defined rubric criteria for microscale video-generation evaluation.The benchmark targets faithful microscale dynamics rather than only superficial visual quality.
  • Conclusion: MicroSim-10K and MicroVerse demonstrate performance on microscale simulation tasks by integrating physical constraints and expert supervision.The stated application scope includes biomedical research, education, and interactive scientific visualization.

LIMITATION

The work targets educational microscale simulations of biological mechanisms rather than reproducing wet-lab results. Its approach does not explicitly encode several governing physical laws, limiting high-precision scientific simulation and prediction.

  • Scope: MicroVerse is intended for educational microscale simulations, not reproduction of results observed in wet-lab experiments.The stated scope concerns biological mechanisms rather than experimental-result replication.
  • Limitation: The approach omits explicit physical laws for blood-flow fluid mechanics, molecular diffusion–reaction, and cellular biomechanical constraints.This limits applicability to scenarios requiring high-precision scientific simulation and prediction.

ETHICS STATEMENT

The dataset is released for research use under copyright, privacy, and content-handling constraints.

  • The data are publicly available under YouTube-compliant terms, with personal and sensitive content excluded.Captions were auto-generated and manually verified to remove inappropriate or identifiable material.
  • The dataset is restricted to research purposes and does not include raw copyrighted data.Only URLs linking to the data will be publicly released because the authors do not own the data copyright.

A DATA FILTERING PIPELINE

The authors construct and validate a filtering and deduplication pipeline for microsimulation videos and benchmark tasks, finding stable results on a clean test set while acknowledging potential source-data overlap for private models.

  • 128K microsimulation videos undergo cleaning to remove irrelevant samples, temporal inconsistencies, black borders, and hard-coded subtitles.The process uses OpenCV for border cropping and EasyOCR for subtitle detection and removal.
  • A three-level deduplication pipeline reports zero overlap at high similarity thresholds across source, text, and vision levels.Text- and vision-level thresholds were later lowered to 0.60 to remove potentially overlapping tasks.
  • 445 clean-test tasks produce highly stable performance scores, with only negligible decimal-level drops compared with the original evaluation.The authors therefore retain the 459-task benchmark.
  • MicroWorldBench may include data that private models such as Veo3 have previously seen on YouTube.The authors argue that expert-weighted scientific-fidelity evaluation remains relevant and plan to add specialized microscopy databases.

B ABLATION STUDY

Ablation studies show that data quality and scale improve scientific fidelity, while excessive guidance harms performance; training stabilizes around 5,000 steps and uses 81 frames as a practical compromise.

  • Dataset filtering: Filtered MicroSim-10K data yields more balanced and reliable gains than raw data, improving scientific fidelity without major visual-quality or instruction-following losses.
  • Dataset size: Larger high-quality training datasets steadily improve scientific fidelity and instruction following, with only minor visual-quality tradeoffs.
  • Training recipes (CFG rate): Excessive classifier-free guidance sharply degrades both scientific fidelity and visual quality.
  • Training recipes (training steps): 5,000 training steps are adopted because the model converges rapidly and then stabilizes around that point.
  • Training recipes (number of frames): 81 frames balance sufficient duration for complete biological events, temporal coherence, computational cost, and Wan2.1’s native temporal window.

C ANALYSIS OF THE PROCESS OF EXPERT REVISION AND VALIDATION

The evaluation framework combines expert revision with LLM-assisted rubric generation and validation, emphasizing scientific fidelity across biological scales. Expert and human–LLM agreement analyses support the reliability of the scoring process under standardized model evaluation settings.

  • Expert background: All evaluation experts hold Ph.D. degrees and have extensive cellular and molecular biology research experience.
  • Expert revision process: Experts most often adjusted rubric weights, while supplementing missing criteria was relatively infrequent because it required additional scientific-criterion identification.
  • Expert revision process: Experts filtered trivial criteria, adjusted weights to prioritize mechanisms, and supplemented missing physical constraints; only about 7.6% of 459 tasks needed added scientific criteria.
  • Expert revision process: Subcellular tasks received more frequent expert intervention to enforce stricter physical constraints where GPT-5 was less reliable.
  • Validation: Including GPT-5 maintains or slightly improves inter-rater agreement, and its agreement with experts is comparable to human evaluators across biological scales.
  • Rubric design: The rubric evaluates Scientific Fidelity, Visual Quality, and Instruction Following using criterion descriptions, polarity, and importance weights.
  • Validation: Three experts show strong agreement under Cohen’s Kappa, supporting the reliability of the scoring process.
Loading 2603.00585v1…