# SteelBench: Evaluating Vision-Language Models in Real-World Industrial Environments

arXiv: https://arxiv.org/abs/2607.05264 · HTML: https://mlsanity.com/paper/2607.05264/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Placed in Chart understanding and anomaly detection, one of 16 directions in Multimodal LLMs / vision-language models.
- Builds on 21 of its 29 references that are in the map.
- First posted on arXiv 6 Jul 2026.

## Placed in

- Multimodal LLMs / vision-language models / [Chart understanding and anomaly detection](https://mlsanity.com/mllm/chart-understanding-and-anomaly-detection/)

## Builds on (most cited first, 21 of 21)

- [Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset](https://mlsanity.com/paper/1705.07750/) (2017)
- [Benchmarking Neural Network Robustness to Common Corruptions and Perturbations](https://mlsanity.com/paper/1903.12261/) (2019)
- [MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI](https://mlsanity.com/paper/2311.16502/) (2023)
- [The Kinetics Human Action Video Dataset](https://mlsanity.com/paper/1705.06950/) (2017)
- [Ego4D: Around the World in 3,000 Hours of Egocentric Video](https://mlsanity.com/paper/2110.07058/) (2021)
- [Video-MME: The First-Ever Comprehensive Evaluation Benchmark of Multi-modal LLMs in Video Analysis](https://mlsanity.com/paper/2405.21075/) (2024)
- [Datasheets for Datasets](https://mlsanity.com/paper/1803.09010/) (2018)
- [MVBench: A Comprehensive Multi-modal Video Understanding Benchmark](https://mlsanity.com/paper/2311.17005/) (2023)
- [Real-world Anomaly Detection in Surveillance Videos](https://mlsanity.com/paper/1801.04264/) (2018)
- [AVA: A Video Dataset of Spatio-temporally Localized Atomic Visual Actions](https://mlsanity.com/paper/1705.08421/) (2017)
- [Future Frame Prediction for Anomaly Detection -- A New Baseline](https://mlsanity.com/paper/1712.09867/) (2017)
- [Assembly101: A Large-Scale Multi-View Video Dataset for Understanding Procedural Activities](https://mlsanity.com/paper/2203.14712/) (2022)
- [Preference Leakage: A Contamination Problem in LLM-as-a-judge](https://mlsanity.com/paper/2502.01534/) (2025)
- [NaturalBench: Evaluating Vision-Language Models on Natural Adversarial Samples](https://mlsanity.com/paper/2410.14669/) (2024)
- [The MECCANO Dataset: Understanding Human-Object Interactions from Egocentric Videos in an Industrial-like Domain](https://mlsanity.com/paper/2010.05654/) (2020)
- [Just Put a Human in the Loop? Investigating LLM-Assisted Annotation for Subjective Tasks](https://mlsanity.com/paper/2507.15821/) (2025)
- [IndustryEQA: Pushing the Frontiers of Embodied Question Answering in Industrial Scenarios](https://mlsanity.com/paper/2505.20640/) (2025)
- [Vision Language Model for Interpretable and Fine-grained Detection of Safety Compliance in Diverse Workplaces](https://mlsanity.com/paper/2408.07146/) (2024)
- [The State of Data Curation at NeurIPS: An Assessment of Dataset Development Practices in the Datasets and Benchmarks Track](https://mlsanity.com/paper/2410.22473/) (2024)
- [iSafetyBench: A video-language benchmark for safety in industrial environment](https://mlsanity.com/paper/2508.00399/) (2025)
- [SH17: A Dataset for Human Safety and Personal Protective Equipment Detection in Manufacturing Industry](https://mlsanity.com/paper/2407.04590/) (2024)

## Built on by (most cited first, 1 of 1)

- [SafeSceneReason: A Multimodal Reasoning Benchmark Connecting Industrial Hazards with Accident Knowledge](https://mlsanity.com/paper/2608.09230/) (2026)

## Abstract (arXiv)

Existing video benchmarks evaluate action recognition on consumer videos, egocentric recordings, or simulated industrial environments. They do not test vision-language models under the visual and procedural conditions of real industrial CCTV, where workers appear as distant figures amid dust, steam, low light, glare, occlusion, and overlapping activities. We introduce STEELBENCH, a diagnostic benchmark for industrial surveillance that jointly evaluates per-worker activity recognition, safety-rule reasoning, and annotation provenance. SteelBench contains 1,345 densely annotated clips, curated from 149 hours of operational plant footage and 10,024 candidate clips using temporal deduplication, class balancing, and visibility-aware stratified sampling. Each clip includes dense per-worker action labels, PPE attributes, spatial context, and safety-rule annotations. Because model-assisted annotation can shape the labels later used for model evaluation, SteelBench includes a provenance-aware audit protocol. The protocol measures label influence, evaluates sensitivity to ground-truth provenance, and reports a human reference from expert-reviewed labels. Applying this audit, we find that unaudited VLM-sourced ground truth can inflate same-family model accuracy by up to 17 percentage points. Across nine VLMs from four architectural families, the best model reaches only 42.6% action accuracy, compared with an 84.6% human benchmark. Performance also fragments across recognition, robustness, calibration, and safety reasoning. Even when models predict the correct action, 37-58% of cases still yield incorrect safety judgments, and no model passes more than 2 of 5 diagnostic checks. The dataset is publicly available on Hugging Face.