Pith. sign in

REVIEW 3 major objections 6 minor 56 references

This paper introduces a construction-safety benchmark mined from over 100,000 real inspection records and argues that current multimodal language models are far from reliable, with the best overall score near 60.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review

2026-08-04 01:21 UTC pith:HHSO6E5V

load-bearing objection Solid, expert-verified safety benchmark with a useful mining pipeline; the 'temporal-robust' framing is softer than the paper's own caveats, but the core value survives. the 3 major comments →

arxiv 2608.00068 v1 pith:HHSO6E5V submitted 2026-07-29 cs.CV

SafeBuild-Bench: A Temporal-Robust Construction Safety Benchmark with Graph-Enhanced Data Mining

Yi Cui , Zilin Wang , Yijie Xu , Qianyi Cai , Huizai Yao , Shuai Jiang , Bingzhuo Zhong , Hui Xiong This is my paper
classification cs.CV
keywords BenchmarkData-Centric AITemporal RobustnessData EfficiencyConstruction SafetyLong-Tail HazardsMultimodal LLM EvaluationGraph-Based Data Selection
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

SafeBuild-Bench is an expert-verified benchmark of 3,314 construction-hazard task instances mined from over 100,000 real inspection image-text records, designed so that models can be scored per month and per site rather than on one fixed split. To build it affordably, the authors develop GEMS, a graph-based selection pipeline that ranks confusing, non-duplicate candidates for human review instead of labeling automatically. On the benchmark, today's multimodal models remain far from dependable: the best overall score is about 60 out of 100. In proxy experiments on public instruction data, training on a 1% GEMS-selected subset matches or exceeds full-data training on reliability-oriented evaluation suites. The paper's point is that large inspection archives are information-poor at the top and hazard-rich in the long tail, and that benchmarks must be built and scored accordingly.

Core claim

The paper's central claim is that a trustworthy construction-safety benchmark can be mined from redundant, long-tailed industrial archives, and that doing so exposes a large gap in current models. SafeBuild-Bench retains collection timestamps and site identifiers for every instance, so users can stratify scores by time and location; the authors report that fixed models move by roughly 8 to 13 accuracy points across July-to-November month slices. The same pipeline that selects benchmark candidates also keeps a small training subset informative: on a public instruction-tuning mix, a 1% GEMS-selected subset performs at about the full-data level on reliability-focused evaluations. The benchmark

What carries the argument

The engine is GEMS (Graph-Enhanced Multimodal Selection), a candidate-mining pipeline that turns a raw archive into an information core. It embeds image-text pairs jointly, scores each candidate by how uncertain a proxy model is about its own hazard analysis (length-normalized negative log-likelihood), builds a k-nearest-neighbour graph on the embeddings, and maximizes a submodular utility with a saturation threshold so that selected items cover many scene modes without near-duplicates. The greedy optimization carries a 1 - 1/e approximation guarantee. The other load-bearing piece is the expert-defined confusion groups used to build multiple-choice distractors from visually and semantically

Load-bearing premise

The benchmark's validity rests on the assumption that the roughly 3,000 images selected by embedding similarity and proxy-model confusion from over 100,000 raw records fairly represent the distribution of construction hazards that matters in deployment, rather than merely the scenes the selection machinery finds confusing or embed closely.

What would settle it

Take a random, expert-verified sample of 3,000 images from the same archives and run the same models: if scores jump well above 60 or the model ranking changes materially, then GEMS selection, not model weakness, would be driving the low benchmark numbers. Alternatively, a strict forward-chaining protocol (train on earlier months, test on later months) in which accuracy no longer drops would refute the temporal-robustness claim.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X LinkedIn Reddit HN

If this is right

  • If the benchmark's claims hold, no current multimodal model is close to deployment-ready for construction-safety understanding; the best overall score near 60 sets a low ceiling.
  • Reporting only an aggregate score becomes misleading: fixed models vary by 8 to 13 accuracy points across month slices, so evaluations should report stratified time, site, category, and task views.
  • Small data budgets do not have to mean weaker models: a 1% GEMS-selected subset of a public instruction mix matches or exceeds full-data training on reliability-oriented evaluations.
  • Free-form hazard-description failures are mostly detection failures, not language failures, so improving hazard-presence detection should come first.
  • The release of metadata, evaluation scripts, and a fixed judge makes construction-safety evaluation reproducible and lets future models be compared on the same time and site slices.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because GEMS ranks candidates using one proxy model and one embedding space, the benchmark could be tilted toward scenes that happen to confuse that proxy; a multi-proxy or fully random expert-verified sample would reveal the size of that tilt.
  • Month-to-month score variation is evidence of temporal heterogeneity, but without strict forward-chaining it can also reflect differences in site mix and collection conditions across months; a true deployment test would train on early months and test on later ones.
  • The mine-then-verify recipe should transfer to other regulated visual domains such as mining, utilities, or warehousing, provided the hazard taxonomy and expert guidelines are rebuilt for local regulations.
  • If the judge audit generalizes, the failure pattern suggests that a reliable hazard-presence detector could be built first and plugged into description models—a testable architecture for safety systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. This paper introduces SafeBuild-Bench, an expert-verified construction-safety benchmark mined from over 100K industrial image-text inspection records. The release contains 3,314 task instances from over 3,000 images, covering multiple-choice hazard identification (2,200 instances) and free-form hazard description (1,114 instances), with each instance retaining temporal and site metadata. To make expert verification scalable, the authors propose GEMS, a graph-enhanced selection pipeline that scores candidates by proxy-model uncertainty (NLL) and selects a diverse, informative subset via a submodular objective on a k-NN graph. GEMS is validated on LLaVA-Instruct-150K and LLaVA-Mixed-665K, where 1% to 20% GEMS-selected subsets match or exceed full-data fine-tuning on robustness-oriented benchmarks such as VizWiz, MMMU, MME, and POPE. On SafeBuild-Bench, 15 evaluated MLLMs achieve overall scores between roughly 30 and 60, with Kimi-K2.5 best at 60.7. The paper also reports supplementary audits covering judge reliability, hard-case labeling ambiguity, subset composition, and month-stratified MCQ accuracy, and it releases the benchmark, evaluation scripts, and GEMS code.

Significance. If the release matches the description, SafeBuild-Bench fills a real gap: a public, expert-verified construction-safety benchmark with explicit temporal and site metadata and an emphasis on long-tail hazards. The GEMS approach is a plausible and useful contribution to data-centric MLLM benchmarking, and the proxy validation is nontrivial: small selected subsets matching or exceeding full-data training on multiple robustness-oriented benchmarks is concrete evidence that the curation signal carries information. The paper is also commendably careful in scope: it explicitly disclaims forward-chaining or held-out-site interpretations, reports judge and ambiguity audits, and separates the stable benchmark from supporting analyses. These strengths make the resource potentially valuable to the community, provided the methodological gaps below are addressed.

major comments (3)
  1. [§4.2, Eq. (3), Table 2(b)] The uncertainty weight λ is used in the implementation and ablation but never defined. The text says GEMS sets λ=0.6 and ablates λ∈{2,4} to 'increasingly up-weight model confusion in the selection objective,' yet Eq. (3) defines the utility F(S) without any λ. The exact form of the objective actually optimized—and how λ enters—must be specified, or the ablation and the main GEMS results are not reproducible from the paper text.
  2. [§4.3, Table 4; §1 contribution (1)] The month-stratified results are presented as evidence of 'temporal-robust' evaluation, but month is confounded with site identity, construction phase, weather, and the per-month GEMS candidate-pool composition. The Table 4 caption and §5 appropriately state that this is 'not a held-out-site or forward-chaining protocol,' yet the title and the contribution list still frame the benchmark as 'temporal-robust.' Because this temporal dimension is the benchmark's distinguishing feature, the authors should either add a site-matched or forward-chaining control analysis, or narrow the claims to 'metadata-stratified evaluation' throughout the title and contributions. Without that change, the observed July–November variation cannot support a conclusion about temporal distribution shift.
  3. [§3.3, §4.3] The benchmark claims site variation as a core contribution, but no site-stratified evaluation is reported anywhere. Since the temporal analysis is weakened by the confounds noted above, the site dimension is the remaining novel axis. The authors should report at least one site-stratified slice (e.g., accuracy per site for a few models), or explain why site-level results cannot be released without compromising privacy. Currently the site metadata is promised but its analytical utility is not demonstrated.
minor comments (6)
  1. [Table 1 / §4.3] Table 1 appears to show Gemini-3-Flash-Preview with Hazard Identification Accuracy 65.8, while the text says 'Gemini-3-Flash-Preview achieves the highest accuracy (55.8).' One of these is incorrect; please reconcile.
  2. [§2.2, §3.2] The text repeatedly refers to a 'dual-graph structure' in GEMS, but §3.2 and Algorithm 1 describe only a single k-NN graph. Either specify the second graph or remove the 'dual-graph' terminology to avoid confusing readers.
  3. [Table 4] The table would benefit from per-month sample counts, especially since the text notes October contains only 66 MCQ items. This would help readers assess the stability of the month-stratified estimates.
  4. [§4.2] Sensitivity analysis is provided for λ but not for the saturation threshold τ or the k in the k-NN graph. Since these are free parameters in Eq. (3) and Algorithm 1, a brief sensitivity note or a statement that they are fixed by validation would improve reproducibility.
  5. [§4.4] In the description-failure analysis, the criterion 'Final ≤ 0.5' is used without defining 'Final.' It appears to be the normalized Description Quality score; please state this explicitly.
  6. [Throughout] There are several formatting glitches: 'LLaV A' with a space in Table 2, 'GPT -4o' with a stray space, and 'Qwen2.5-VL-3Bas' in §4.2. These are minor but should be fixed.

Circularity Check

0 steps flagged

No significant circularity: GEMS validation is external and benchmark evaluation is zero-shot on expert-verified data.

full rationale

The derivation chain does not reduce to its inputs. GEMS is validated by training LLaVA-v1.5 on GEMS-selected public LLaVA instruction subsets and evaluating on external benchmarks (VizWiz, POPE, MMMU, MME) that are disjoint from the selection pool, so the validation is not circular. SafeBuild-Bench instances are selected by GEMS, then expert-verified, and all reported MLLM scores are zero-shot with no training on benchmark data; the headline result (best overall ~60) is therefore an independent measurement. The only self-citation is [15] (authors include H. Xiong), cited as the uncertainty-aware influence-maximization paradigm motivating GEMS's submodular objective; this is background and non-load-bearing because GEMS's effectiveness is demonstrated independently and the benchmark's claims do not depend on [15]. The temporal-slicing results in Table 4 are explicitly described as 'not a held-out-site or forward-chaining protocol,' so month/site/pool confounds are a validity limitation rather than a circular equation. No fitted parameter is renamed as a prediction, and no self-citation chain forces the central claim.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 1 invented entities

The central methodological claims rest on several domain assumptions about the embedding model, the uncertainty proxy, the submodular objective, and the raw archive's representativeness. The free parameters lambda, tau, and k are design choices that are not derived from first principles. The paper does not introduce a new physical or statistical entity, but the GEMS pipeline is the main invention.

free parameters (3)
  • Uncertainty weight lambda = 0.6 (ablated at 0.2 and 0.4 in experiments)
    This controls the trade-off between uncertainty and graph diversity in GEMS and is chosen by the authors; no theory fixes it.
  • Saturation threshold tau = Not stated
    The saturation threshold in the submodular utility function is a design choice; its value is not reported.
  • k in k-NN graph = Not stated
    The number of neighbors for graph construction is not reported in the paper.
axioms (5)
  • domain assumption The Qwen3-VL-Embedding encoder provides a joint image-text embedding that captures semantic proximity for construction-safety scenes.
    Used in GEMS graph construction; if this embedding does not reflect safety-relevant semantics, the graph edges and diversity enforcement are built on the wrong geometry.
  • domain assumption The proxy model's length-normalized negative log-likelihood is a useful informativeness/confusion signal for mining hard safety cases.
    The paper explicitly notes that NLL can be raised by blur, rain, or domain mismatch; it relies on expert verification to filter these, but the selection ranking itself depends on this signal.
  • domain assumption The submodular utility function F(S) with saturation threshold tau is an appropriate objective for the informational core of an industrial stream.
    The approximation guarantee relies on submodularity, but the choice of F(S) as a coverage objective is a modeling decision, not a proven characterization of 'informativeness'.
  • domain assumption The 100K+ archived inspection records from 50+ sites are representative of construction-safety deployment conditions.
    The benchmark inherits any biases in the raw archive (site selection, safety culture, collection protocol). The paper acknowledges the data is grounded in Chinese construction regulations, so generalization to other regulatory contexts requires new taxonomies.
  • domain assumption Expert-verified labels on the selected images are ground truth for construction-safety hazards.
    The paper relies on experts to assign hazard labels; while it audits a small number of cases for ambiguity, the 3,000+ images are not independently verified by multiple experts with measured inter-rater agreement.
invented entities (1)
  • GEMS (Graph-Enhanced Multimodal Selection) pipeline no independent evidence
    purpose: A new selection algorithm combining uncertainty scoring, NLL-based confusion signal, and dual graph diversity to mine informative candidates from redundant data streams.
    GEMS is validated solely through the paper's own experiments; it is not an independently verified or formally defined method with external benchmarks. The components (CELF, submodular selection, uncertainty sampling) are known, but the specific combination is new.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of SafeBuild-Bench: A Temporal-Robust Construction Safety Benchmark with Graph-Enhanced Data Mining." pith.science (2026). https://pith.science/paper/HHSO6E5V

@misc{pith2026260800068,
  author       = {Pith},
  title        = {Pith review of: SafeBuild-Bench: A Temporal-Robust Construction Safety Benchmark with Graph-Enhanced Data Mining},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HHSO6E5V}},
  note         = {Machine review of arXiv:2608.00068}
}
Share X LinkedIn Reddit HN
read the original abstract

Construction-safety models must handle concrete deployment risks, such as a worker standing near a scaffold edge without guardrails, rather than only recognize common objects in curated images. Yet real inspection archives are redundant, long-tailed, and collected across changing sites and months. We introduce SafeBuild-Bench, a metadata-driven benchmark for evaluating multimodal large language models on construction safety under realistic temporal and site variation. It is mined from 100K+ industrial image-text records and contains 3,314 task instances from over 3,000 expert-verified images, covering multiple-choice hazard identification and free-form hazard description. To make expert verification scalable, we develop GEMS, a graph-enhanced multimodal selection pipeline that combines a proxy-model confusion signal with graph-based diversity to identify informative candidates from redundant streams. On public instruction-tuning data, GEMS-selected subsets preserve robustness-oriented performance under small data budgets. On SafeBuild-Bench, current MLLMs remain far from reliable construction-safety understanding, with the best overall score near 60. We release the benchmark, evaluation scripts, and GEMS codebase at https://github.com/safebuild/gems.

Figures

Figures reproduced from arXiv: 2608.00068 by Bingzhuo Zhong, Hui Xiong, Huizai Yao, Qianyi Cai, Shuai Jiang, Yi Cui, Yijie Xu, Zilin Wang.

Figure 1
Figure 1. Figure 1: The dichotomy of construction safety data. (Top) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the GEMS framework. The pipeline involves: (1) Graph Construction: Modeling the data manifold [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Category distribution of samples in SafeBuild￾Bench. The benchmark covers 19 fine-grained hazard cate￾gories, which are further grouped into five higher-level safety domains: Site Management, Elevation Work, Facilities, No Hazard, and Others. safety experts. For each selected sample, experts assign the pri￾mary hazard category and provide a concise reference description following established construction s… view at source ↗
Figure 4
Figure 4. Figure 4: Category-wise identification accuracy on the [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative Examples from SafeBuild-Bench. Two MCQ samples testing precise hazard identification. The correct option is highlighted. Task: Open-Ended Description & Evaluation Ground Truth: The safety netting is improperly installed, with visible gaps that could lead to falls from height. Model Prediction: There is a hazard due to exposed and tangled electrical cables, which pose a risk of tripping, electri… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative Examples from SafeBuild-Bench. Two reasoning samples requiring complex analysis of Temporary Electrical Supply and Working At Heights hazards. Key risk factors are bolded [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

56 extracted references · 1 canonical work pages · 1 internal anchor

  1. [1]

    Josh Achiam et al . 2023. GPT-4 Technical Report. arXiv:2303.08774 [cs.CL] doi:10.48550/ARXIV.2303.08774

    Pith/arXiv arXiv doi:10.48550/arxiv.2303.08774 2023

  2. [2]

    Gonzalez, and Qipei Mei

    Muhammad Adil, Gaang Lee, Vicente A. Gonzalez, and Qipei Mei. 2025. Us- ing Vision Language Models for Safety Hazard Identification in Construction. arXiv:2504.09083 [cs.CV] https://arxiv.org/abs/2504.09083

    Pith/arXiv arXiv 2025

  3. [3]

    2025.Claude 4.5 Sonnet System Card

    Anthropic. 2025.Claude 4.5 Sonnet System Card. Technical Report. Anthropic

    2025

  4. [4]

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, et al . 2025. Qwen3-VL Technical Report. arXiv:2511.21631 [cs.CV] https://arxiv.org/abs/ 2511.21631

    Pith/arXiv arXiv 2025

  5. [5]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, et al

  6. [6]

    Powei Chang, Jinpeng Zhang, Bowen Chen, Chenyu Wang, Chenlu Guo, Yixing Zhang, Yukang Gao, JianXiang Xiang, Yue Gao, Chaoqun Sun, Yiyi Chen, and Dongying Kong. 2026. SPICE: Submodular Penalized Information-Conflict Se- lection for Efficient Large Language Model Training. arXiv:2601.23155 [cs.LG] https://arxiv.org/abs/2601.23155

    Pith/arXiv arXiv 2026

  7. [7]

    Jierun Chen, Fangyun Wei, Jinjing Zhao, Sizhe Song, Bohuai Wu, Zhuoxuan Peng, S-H Gary Chan, and Hongyang Zhang. 2025. Revisiting referring expression comprehension evaluation in the era of large multimodal models. InProceedings of the Computer Vision and Pattern Recognition Conference. 513–524

    2025

  8. [8]

    Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al

  9. [9]

    Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre-Emmanuel Mazaré, Maria Lomeli, Lucas Hosseini, and Hervé Jégou. 2025. The Faiss library. arXiv:2401.08281 [cs.LG] https://arxiv.org/abs/2401.08281

    Pith/arXiv arXiv 2025

  10. [10]

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities

  11. [11]

    Chaoyou Fu, Jun Chen, et al. 2023. MME: A Comprehensive Evaluation Bench- mark for Multimodal Large Language Models.arXiv preprint arXiv:2306.13394 (2023). doi:10.48550/ARXIV.2306.13394

    Pith/arXiv arXiv doi:10.48550/arxiv.2306.13394 2023

  12. [12]

    Dan Feldman. 2020. Core-sets: An Updated Survey.arXiv preprint arXiv:2011.09384(2020). doi:10.48550/ARXIV.2011.09384

    Pith/arXiv arXiv doi:10.48550/arxiv.2011.09384 2020

  13. [13]

    Yash Goyal, Tanishq Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh

  14. [14]

    2025.Gemini 3 Flash Model Card

    Google. 2025.Gemini 3 Flash Model Card. Technical Report. Google

    2025

  15. [15]

    Jindong Han, Hao Liu, Jun Fang, Naiqiang Tan, and Hui Xiong. 2025. Automatic Instruction Data Selection for Large Language Models via Uncertainty-Aware Influence Maximization. InProceedings of the ACM on Web Conference 2025 (WWW ’25). ACM, 4969–4979. doi:10.1145/3696410.3714817

    arXiv 2025

  16. [16]

    Hudson and Christopher D

    Drew A. Hudson and Christopher D. Manning. 2019. GQA: A New Dataset for Real-World Visual Reasoning and Compositional Question Answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2019

  17. [17]

    Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. 2018. Vizwiz grand challenge: Answering visual questions from blind people. InProceedings of the IEEE conference on computer vision and pattern recognition. 3608–3617

    2018

  18. [18]

    Jaewoo Lee, Boyang Li, and Sung Ju Hwang. 2024. Concept-skill transferability- based data selection for large vision-language models.arXiv preprint arXiv:2406.10995(2024)

    Pith/arXiv arXiv 2024

  19. [19]

    Yeo-Reum Lee, Seung-Hwan Jung, Kyung-Su Kang, Han-Cheol Ryu, and Han- Guk Ryu. 2023. Deep learning-based framework for monitoring wearing personal protective equipment on construction sites.Journal of Computational Design and Engineering10, 2 (2023), 905–917

    2023

  20. [21]

    Mingxin Li, Yanzhao Zhang, Dingkun Long, Keqin Chen, Sibo Song, Shuai Bai, Zhibo Yang, Pengjun Xie, An Yang, Dayiheng Liu, Jingren Zhou, and Junyang Lin

  21. [22]

    Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. 2023. Evaluating Object Hallucination in Large Vision-Language Models. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP). 292–305. doi:10.18653/v1/2023.emnlp-main.20

    doi:10.18653/v1/2023.emnlp-main.20 2023

  22. [23]

    Jure Leskovec, Andreas Krause, Carlos Guestrin, Christos Faloutsos, Jeanne Van- briesen, and Natalie Glance. 2007. Cost-effective outbreak detection in networks. Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining420-429, 420–429. doi:10.1145/1281192.1281239

    arXiv 2007

  23. [24]

    Lawrence Zitnick, and Piotr Dollár

    Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dollár

  24. [25]

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2023. Improved Baselines with Visual Instruction Tuning. arXiv:2310.03744 [cs.CV] https://arxiv.org/abs/ 2310.03744

    Pith/arXiv arXiv 2023

  25. [26]

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2024. Improved baselines with visual instruction tuning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 26296–26306

    2024

  26. [27]

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. 2017. Focal Loss for Dense Object Detection. InProceedings of the IEEE International Conference on Computer Vision. 2980–2988

    2017

  27. [28]

    Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin

  28. [29]

    Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. 2022. Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answering. arXiv preprint arXiv:2209.09513(2022). doi:10.48550/ARXIV.2209.09513

    Pith/arXiv arXiv doi:10.48550/arxiv.2209.09513 2022

  29. [30]

    Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty

  30. [31]

    OpenAI. 2024. GPT-4o System Card. arXiv:2410.21276 [cs.CL] https://arxiv.org/ abs/2410.21276

    Pith/arXiv arXiv 2024

  31. [32]

    Liangxin Liu, Xuebo Liu, Derek F Wong, Dongfang Li, Ziyi Wang, Baotian Hu, and Min Zhang. 2024. Selectit: Selective instruction tuning for llms via uncertainty- aware self-reflection.Advances in Neural Information Processing Systems37 (2024), 97800–97825

    2024

  32. [33]

    Ahmed Bin Kabir Rabbi and Idris Jeelani. 2024. AI integration in construction safety: Current state, challenges, and future opportunities in text, vision, and audio based applications.Automation in Construction164 (2024), 105443

    2024

  33. [34]

    Florian Schroff, Dmitry Kalenichenko, and James Philbin. 2015. FaceNet: A Unified Embedding for Face Recognition and Clustering. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 815–823

    2015

  34. [35]

    Abhinav Shrivastava, Abhinav Gupta, and Ross Girshick. 2016. Training Region- Based Object Detectors with Online Hard Example Mining. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 761–769

    2016

  35. [36]

    Amanpreet Singh, Vivek Natarajan, Yu Jiang, Xinlei Chen, Manohar Shah, Marcus Rohrbach, Dhruv Batra, and Devi Parikh. 2019. Towards VQA Models That Can Read. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2019

  36. [37]

    Zhiqing Sun, Sheng Shen, et al. 2023. Aligning Large Multimodal Models with Factually Augmented RLHF.arXiv preprint arXiv:2309.14525(2023). doi:10.48550/ ARXIV.2309.14525

    Pith/arXiv arXiv doi:10.48550/arxiv.2309.14525 2023

  37. [38]

    Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, et al. 2025. Gemma 3 technical report.arXiv preprint arXiv:2503.19786 (2025)

    Pith/arXiv arXiv 2025

  38. [39]

    Zhenhui Ou, Dawei Li, Zhen Tan, Wenlin Li, Huan Liu, and Siyuan Song

  39. [40]

    InProceedings of the 34th ACM International Conference on Information and Knowledge Management(Seoul, Republic of Korea)(CIKM ’25)

    Building Safer Sites: A Large-Scale Multi-Level Dataset for Construc- tion Safety Benchmark. InProceedings of the 34th ACM International Conference on Information and Knowledge Management(Seoul, Republic of Korea)(CIKM ’25). Association for Computing Machinery, New York, NY, USA, 6508–6512. doi:10.1145/3746252.3761652

    arXiv

  40. [41]

    Antoine J. P. Tixier and Matthew R. Hallowell. 2023. Safer Together: Machine Learning Models Trained on Shared Accident Datasets Predict Construction Injuries Better than Company-Specific Models. arXiv:2301.03567 [cs.LG] https: //arxiv.org/abs/2301.03567

    Pith/arXiv arXiv 2023

  41. [42]

    Mengzhou Xia, Sadhika Malladi, Suchin Gururangan, Sanjeev Arora, and Danqi Chen. 2024. Selecting Influential Data for Targeted Instruction Tuning.arXiv preprint arXiv:2402.04333(2024). doi:10.48550/ARXIV.2402.04333

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.04333 2024

  42. [43]

    Lu Yang, He Jiang, Qing Song, and Jun Guo. 2022. A survey on long-tailed visual recognition.International Journal of Computer Vision130, 7 (2022), 1837–1872

    2022

  43. [44]

    Huaxiu Yao, Caroline Choi, Bochuan Cao, Yoonho Lee, Pang Wei W Koh, and Chelsea Finn. 2022. Wild-time: A benchmark of in-the-wild distribution shift over time.Advances in Neural Information Processing Systems35 (2022), 10309–10324

    2022

  44. [45]

    Qifan Yu, Zhebei Shen, Zhongqi Yue, Yang Wu, Wenqiao Zhang, Yunfei Li, Juncheng Li, Siliang Tang, and Yueting Zhuang. 2024. Mastering Collabora- tive Multi-modal Data Selection: A Focus on Informativeness, Uniqueness, and Representativeness.arXiv preprint arXiv:2412.06293(2024). doi:10.48550/ARXIV. 2412.06293

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2412.06293 2024

  45. [46]

    Tianyu Yu, Zefan Wang, Chongyi Wang, Fuwei Huang, Wenshuo Ma, Zhihui He, Tianchi Cai, Weize Chen, Yuxiang Huang, Yuanqian Zhao, Bokai Xu, Junbo Cui, Yingjing Xu, Liqing Ruan, Luoyuan Zhang, Hanyu Liu, Jingkun Tang, Hongyuan Liu, Qining Guo, Wenhao Hu, Bingxiang He, Jie Zhou, Jie Cai, Ji Qi, Zonghao Guo, Chi Chen, Guoyang Zeng, Yuxuan Li, Ganqu Cui, Ning D...

    Pith/arXiv arXiv 2025

  46. [47]

    Kimi Team, Tongtong Bai, Yifan Bai, Yiping Bao, S. H. Cai, Yuan Cao, Y. Charles, H. S. Che, Cheng Chen, Guanduo Chen, Huarong Chen, Jia Chen, Jiahao Chen, Jianlong Chen, Jun Chen, Kefan Chen, Liang Chen, Ruijue Chen, Xinhao Chen, Yanru Chen, Yanxu Chen, Yicun Chen, Yimin Chen, Yingjiang Chen, Yuankun Chen, Yujie Chen, Yutian Chen, Zhirong Chen, Ziwei Chen...

    Pith/arXiv arXiv 2026

  47. [48]

    V Team, Wenyi Hong, Wenmeng Yu, Xiaotao Gu, Guo Wang, Guobing Gan, Haomiao Tang, Jiale Cheng, Ji Qi, Junhui Ji, Lihang Pan, Shuaiqi Duan, Weihan Wang, Yan Wang, Yean Cheng, Zehai He, Zhe Su, Zhen Yang, Ziyang Pan, Aohan Zeng, Baoxu Wang, Bin Chen, Boyan Shi, Changyu Pang, Chenhui Zhang, Da Yin, Fan Yang, Guoqing Chen, Jiazheng Xu, Jiale Zhu, Jiali Chen, J...

    Pith/arXiv arXiv 2025

  48. [49]

    Xing Zi, Jinghao Xiao, Yunxiao Shi, Xian Tao, Jun Li, Ali Braytee, and Mukesh Prasad. 2025. RSVLM-QA: A Benchmark Dataset for Remote Sensing Vision Language Model-based Question Answering. InProceedings of the 33rd ACM International Conference on Multimedia. 12905–12911. A Release, Audit, and Ethics Notes Dataset composition and release.SafeBuild-Benchcon...

    2025

  49. [55]

    Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. 2024. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 9556–9567

    2024

  50. [56]

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, et al. 2025. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models. arXiv preprint arXiv:2504.10479(2025)

    Pith/arXiv arXiv 2025

  51. [2015]

    arXiv:1405.0312 [cs.CV] https://arxiv.org/abs/1405.0312

    Microsoft COCO: Common Objects in Context. arXiv:1405.0312 [cs.CV] https://arxiv.org/abs/1405.0312

    Pith/arXiv arXiv

  52. [2017]

    InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    Making the V in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

  53. [2019]

    In ICDAR

    OCR-VQA: Visual Question Answering by Reading Text in Images. In ICDAR

  54. [2023]

    doi:10.48550/ARXIV.2307.06281

    MMBench: Is Your Multi-modal Model an All-around Player?arXiv preprint arXiv:2307.06281(2023). doi:10.48550/ARXIV.2307.06281

    Pith/arXiv arXiv doi:10.48550/arxiv.2307.06281 2023

  55. [2025]

    arXiv:2502.13923 [cs.CV] https://arxiv.org/ abs/2502.13923

    Qwen2.5-VL Technical Report. arXiv:2502.13923 [cs.CV] https://arxiv.org/ abs/2502.13923

    Pith/arXiv arXiv

  56. [2026]

    Qwen3-VL-Embedding and Qwen3-VL-Reranker: A Unified Framework for State-of-the-Art Multimodal Retrieval and Ranking.arXiv(2026)

    2026

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.