Pith. sign in

REVIEW 3 major objections 3 minor 2 cited by

The paper claims that MLLM trustworthiness can be measured by a unified 32-task benchmark, that current multimodal models are less trustworthy than capable, and that multimodal training amplifies base-model risks; it also proposes a chain-o

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 ↗

MultiTrust-X is a new 32-task, 28-dataset benchmark over 30 multimodal LLMs claiming that trustworthiness lags capability, that multimodality amplifies base-model risks, and that its RESA alignment method reaches state-of-the-art safety.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection The submission is not a paper: the title and abstract describe an MLLM trustworthiness benchmark, but the body is an unrelated 3D reconstruction paper with different authors; there is nothing to referee. the 3 major comments →

arxiv 2508.15370 v1 pith:CF7KDLYR submitted 2025-08-21 cs.CL cs.AI

Unveiling Trust in Multimodal Large Language Models: Evaluation, Analysis, and Mitigation

Yichi Zhang , Yao Huang , Yifan Wang , Yitong Sun , Chang Liu , Zhe Zhao , Zhengwei Fang , Huanran Chen
show 5 more authors
Xiao Yang Xingxing Wei Hang Su Yinpeng Dong Jun Zhu
This is my paper
classification cs.CL cs.AI
keywords multimodal large language modelstrustworthiness benchmarksafety alignmentchain-of-thoughtmultimodal risksevaluation frameworkreasoning-enhanced safety alignmentMLLM
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.

The reading

This paper tries to establish that trustworthiness in multimodal large language models is not a side effect of capability, but a distinct property that can be measured, analyzed, and intentionally engineered. It introduces MultiTrust-X, a benchmark built on a three-dimensional taxonomy of five trustworthiness aspects, two novel risk types, and four mitigation perspectives, instantiated as 32 tasks over 28 curated datasets. Running 30 open-source and proprietary MLLMs through it, the authors claim three things: trustworthiness lags general capability; multimodal training and inference amplify risks already present in base LLMs; and current mitigation methods improve narrow aspects but rarely overall trustworthiness, often at the cost of utility. They then propose RESA, a reasoning-enhanced safety alignment method that uses chain-of-thought to make models discover latent risks, reporting state-of-the-art results.

Core claim

The paper's central claim is that a comprehensive, reusable measurement standard for MLLM trustworthiness is possible, and that measuring against it reveals systemic problems: current MLLMs are significantly less trustworthy than their general capabilities suggest, and the process of making them multimodal—both at training and inference—tends to amplify risks inherited from the underlying language model. The benchmark MultiTrust-X operationalizes this through a three-dimensional taxonomy covering five aspects (truthfulness, robustness, safety, fairness, privacy), two novel risk types (multimodal risks and cross-modal impacts), and four mitigation perspectives (data, architecture, training, i

What carries the argument

MultiTrust-X is the central instrument: a three-dimensional taxonomy (aspects × risk types × mitigation perspectives) that organizes 32 tasks and 28 curated datasets into a holistic evaluation of 30 MLLMs. The taxonomy defines the categories of trustworthiness, the ways multimodality introduces or transforms risk, and the kinds of interventions to compare. RESA is the paper's proposed mitigation: it augments alignment training so that the model generates chain-of-thought reasoning about potential risks before answering, turning risk discovery into an explicit reasoning step.

Load-bearing premise

The load-bearing premise is that the author-defined taxonomy and its 28 curated datasets actually measure trustworthiness as it matters in real multimodal deployments; if they do not, the capability gap, amplification, and mitigation findings are properties of the benchmark rather than of the models.

What would settle it

Construct an independent trustworthiness test set from real-world user reports, adversarial red-team logs, and documented deployment incidents, run the same 30 models on it, and check whether the rankings and the capability–trustworthiness gap match MultiTrust-X; a large divergence would show the benchmark captures a proxy, not trustworthiness.

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

If this is right

  • MultiTrust-X could serve as a reusable evaluation standard, letting future MLLM releases report trustworthiness alongside capability numbers.
  • The claimed capability–trustworthiness gap implies that high benchmark scores on general tasks should not be taken as evidence of safe deployment; trust needs separate measurement.
  • The finding that multimodal training and inference amplify base-LLM risks suggests safety should be re-audited after every modality-fusion stage, not only on the final model.
  • The observation that existing mitigations trade utility for safety supports a shift toward reasoning-based alignment objectives rather than simple data filtering or output censoring.
  • If RESA's reasoning-based risk discovery works as described, it offers an interpretable path to safety: models can articulate what risk they detected and how they avoided it.

Where Pith is reading between the lines

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

  • If the trustworthiness–capability gap is real across the evaluated 30 models, then capability leaderboards may mislead deployers; a testable consequence is that models ranked high on general benchmarks will not necessarily rank high on harm-avoidance in multimodal settings.
  • The amplification claim yields a direct experiment: take a base LLM with known biased or toxic text behavior, apply multimodal instruction tuning, and measure whether text-only toxicity probes become easier to trigger after the multimodal stage.
  • The taxonomy's 'cross-modal impacts' risk type generalizes naturally to audio, video, and other modalities; its structure could be reused to build trustworthiness suites for new MLLM generations without redefining categories.
  • The full text attached in this manuscript is a different paper ('DriveSplat', on 3D driving-scene reconstruction), so the summary above is grounded in the title and abstract of arXiv:2508.15370; readers should fetch the actual paper text before relying on implementation details.
Share X Bluesky LinkedIn Reddit HN

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 / 3 minor

Summary. The manuscript under review consists of an abstract claiming a comprehensive benchmark 'MultiTrust-X' for evaluating, analyzing, and mitigating trustworthiness in multimodal large language models (MLLMs), with 32 tasks, 28 datasets, 30 models, and a new method RESA achieving state-of-the-art results. The full text supplied, however, is a completely different paper: 'DriveSplat: Unified Neural Gaussian Reconstruction for Dynamic Driving Scenes', a 3D Gaussian splatting method for driving-scene reconstruction. The body contains no mention of MLLMs, trustworthiness, MultiTrust-X, the 32 tasks, the 28 datasets, the 30-model evaluation, or RESA. Consequently, every substantive claim in the abstract—the benchmark definition, the empirical findings (trustworthiness–capability gap, cross-modal risk amplification, limitations of existing mitigations), and the SOTA RESA result—is unsupported by the supplied text.

Significance. If the abstract's claims were accompanied by the described study, this could be a highly significant contribution: a comprehensive, structured benchmark for MLLM trustworthiness with a broad model suite, a proposed mitigation, and empirical findings of practical relevance. The claimed scale (32 tasks, 28 datasets, 30 models, 8 mitigation methods) would be valuable to the community, and the RESA result would be a potential advance in aligning multimodal models. However, the supplied manuscript does not contain this study. There are no protocols, no tables, no error analyses, no benchmark definitions, and no experiments. The significance cannot be assessed from the reviewable record; indeed, the record is internally inconsistent, as the title and abstract describe one paper and the body describes another. No credit can be given for verifiable artifacts because none of the claimed materials appear in the text.

major comments (3)
  1. [Abstract and Full Text] The abstract announces MultiTrust-X, a benchmark for MLLM trustworthiness with 32 tasks, 28 datasets, evaluations over 30 MLLMs, and a RESA method achieving state-of-the-art results. The full text provided is 'DriveSplat: Unified Neural Gaussian Reconstruction for Dynamic Driving Scenes' (arXiv:2508.15376v4), a 3D Gaussian reconstruction paper. None of the terms 'MultiTrust-X', 'RESA', 'MLLM', 'trustworthiness', or 'multimodal risk' appear in the body. This is not a case of incomplete support; the body is an unrelated manuscript. All central claims are therefore unsupported by the supplied text.
  2. [Abstract (RESA claim)] The abstract states that RESA 'achiev[es] state-of-the-art results' without reporting any metric, baseline, or comparison. The body contains no description of RESA, no chain-of-thought alignment protocol, and no experimental results. Because the method is not defined or evaluated anywhere in the manuscript, the SOTA claim is a bare assertion and cannot be verified or falsified from the reviewable record.
  3. [Abstract (benchmark validity)] Even taken at face value, the abstract's empirical conclusions ('gap between trustworthiness and general capabilities', 'amplification of potential risks in base LLMs by multimodal training and inference', 'few [mitigations] effectively address overall trustworthiness') are measured entirely against the author-defined three-dimensional taxonomy and the 28 curated datasets. The manuscript provides no definitions of the five trustworthiness aspects, the two risk types, or the dataset/task selection criteria, and no external validation (e.g., correlation with human judgments or deployment incidents). As supplied, the conclusions are properties of an unspecified instrument, not of MLLMs generally. A corrected manuscript would need to supply the full benchmark specification and ideally some external validation to make these claims load-bearing.
minor comments (3)
  1. [Section 3 header] The header reads 'Methology' instead of 'Methodology'.
  2. [Figures 14–16 captions] Several captions contain spacing artifacts, e.g., 'W aymo' and 'W aymo dataset'.
  3. [References] The reference list is entirely consistent with the DriveSplat paper (3D reconstruction and Gaussian splatting works). This reinforces the mismatch between abstract and body; the manuscript appears to be a different paper from the one described in the abstract.

Circularity Check

0 steps flagged

No circularity found: the supplied full text is a different paper (DriveSplat), so the abstract's MultiTrust-X derivation chain is absent; absence of evidence is not circularity.

full rationale

The abstract of the claimed paper (arXiv:2508.15370) describes MultiTrust-X, a benchmark for MLLM trustworthiness with 32 tasks, 28 datasets, 30 models, and a RESA mitigation method. The supplied full text, however, is DriveSplat (arXiv:2508.15376v4), a 3D Gaussian reconstruction paper for driving scenes. None of the MultiTrust-X taxonomy, tasks, datasets, evaluations, or RESA experiments appear in the body. I therefore cannot walk the claimed derivation chain: there are no equations, tables, or ablation results connecting the abstract's definitions, empirical claims, or SOTA assertion to any concrete computation. This is a missing-evidence / internal-inconsistency problem, not a circular-reasoning problem. No step in the available text reduces a claimed prediction to its own inputs by construction. In particular, the abstract's statement that the authors define a three-dimensional framework and then 'based on the taxonomy' build tasks is the normal construction of a benchmark; reporting a trustworthiness-capability gap and risk amplification using that benchmark is an empirical claim about model scores, not a restatement of the definitions. Likewise, evaluating one's own method on one's own benchmark is standard practice and only becomes circular if the benchmark is rigged to favor the method, which is not shown here. The reviewable record is therefore internally inconsistent, but circularity is not established. The honest finding is no significant circularity, score 0; the central claims are unverifiable from the supplied text, but that is a correctness/evidence concern, not a circularity concern.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 2 invented entities

Abstract-only assessment. The central claims rest on the authors' own definitions and instrument: no external benchmark, human-correlation check, or independent dataset validates the taxonomy; the mitigation is judged on the same benchmark that defines the phenomenon. The full text supplied is an unrelated paper, so none of the ledger items can be confirmed against implementation details.

axioms (3)
  • domain assumption The five aspects (truthfulness, robustness, safety, fairness, privacy) and two risk types (multimodal risks, cross-modal impacts) form a complete and non-overlapping decomposition of MLLM trustworthiness.
    Stated in the abstract as the taxonomy of MultiTrust-X; if the taxonomy omits a dimension of trustworthiness, benchmark rankings and the conclusion that multimodal training amplifies base-LLM risks may not generalize.
  • domain assumption Performance on the 28 curated datasets proxies real-world trustworthiness of MLLMs.
    The abstract's overall findings (capability-trustworthiness gap, risk amplification) are inferred from benchmark scores; this assumes the benchmark's task distribution matches deployment risks.
  • domain assumption RESA's chain-of-thought risk discovery transfers to unseen prompts and models beyond the evaluated set.
    The abstract claims state-of-the-art results for RESA; this assumes the mitigation is not overfit to the benchmark's task distribution.
invented entities (2)
  • MultiTrust-X benchmark no independent evidence
    purpose: Standardized instrument to measure MLLM trustworthiness across five aspects and two risk types
    New benchmark built by the authors; the abstract provides no external validation (e.g., correlation with human judgments or deployment incidents) of its measurements.
  • Reasoning-Enhanced Safety Alignment (RESA) no independent evidence
    purpose: Training/alignment method that adds chain-of-thought reasoning to discover risks, claimed state-of-the-art
    Abstract reports SOTA on the authors' own benchmark without external or held-out validation; no details appear in the supplied text.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Unveiling Trust in Multimodal Large Language Models: Evaluation, Analysis, and Mitigation." pith.science (2026). https://pith.science/paper/CF7KDLYR

@misc{pith2026250815370,
  author       = {Pith},
  title        = {Pith review of: Unveiling Trust in Multimodal Large Language Models: Evaluation, Analysis, and Mitigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CF7KDLYR}},
  note         = {Machine review of arXiv:2508.15370}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

The trustworthiness of Multimodal Large Language Models (MLLMs) remains an intense concern despite the significant progress in their capabilities. Existing evaluation and mitigation approaches often focus on narrow aspects and overlook risks introduced by the multimodality. To tackle these challenges, we propose MultiTrust-X, a comprehensive benchmark for evaluating, analyzing, and mitigating the trustworthiness issues of MLLMs. We define a three-dimensional framework, encompassing five trustworthiness aspects which include truthfulness, robustness, safety, fairness, and privacy; two novel risk types covering multimodal risks and cross-modal impacts; and various mitigation strategies from the perspectives of data, model architecture, training, and inference algorithms. Based on the taxonomy, MultiTrust-X includes 32 tasks and 28 curated datasets, enabling holistic evaluations over 30 open-source and proprietary MLLMs and in-depth analysis with 8 representative mitigation methods. Our extensive experiments reveal significant vulnerabilities in current models, including a gap between trustworthiness and general capabilities, as well as the amplification of potential risks in base LLMs by both multimodal training and inference. Moreover, our controlled analysis uncovers key limitations in existing mitigation strategies that, while some methods yield improvements in specific aspects, few effectively address overall trustworthiness, and many introduce unexpected trade-offs that compromise model utility. These findings also provide practical insights for future improvements, such as the benefits of reasoning to better balance safety and performance. Based on these insights, we introduce a Reasoning-Enhanced Safety Alignment (RESA) approach that equips the model with chain-of-thought reasoning ability to discover the underlying risks, achieving state-of-the-art results.

discussion (0)

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

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Time-consistent catastrophe risk management under the path-dependent effects

    q-fin.RM 2025-08 unverdicted novelty 6.0

    Incorporating path-dependent effects via rough volatility and Hawkes processes increases demand for catastrophe insurance and causes underinsurance if ignored.

  2. Time-consistent catastrophe risk management under the path-dependent effects

    q-fin.RM 2025-08 unverdicted novelty 6.0

    A path-dependent catastrophe model combining rough volatility and a Hawkes process with a modified Omori kernel predicts that ignoring event history causes significant underinsurance, with effects that vary by investm...

Reference graph

Works this paper leans on

36 extracted references · 28 canonical work pages · cited by 1 Pith paper · 2 internal anchors

  1. [1]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Yang, Z., Chen, Y., Wang, J., Manivasagam, S., Ma, W.-C., Yang, A.J., Urtasun, R.: Unisim: A neural closed-loop sensor sim- ulator. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 1389–1399 (2023)

    work page 2023

  2. [2]

    Advances in Neural Information Processing Systems37, 122434–122457 (2024)

    Zhang, Z., Song, T., Lee, Y., Yang, L., Peng, C., Chellappa, R., Fan, D.: Lp-3dgs: Learning to prune 3d gaussian splatting. Advances in Neural Information Processing Systems37, 122434–122457 (2024)

    work page 2024

  3. [3]

    International Journal of Computer Vision126(5), 460–475 (2018) 18

    Han, K., Wong, K.-Y.K., Liu, M.: Dense reconstruction of transparent objects by altering incident light paths through refrac- tion. International Journal of Computer Vision126(5), 460–475 (2018) 18

    work page 2018

  4. [4]

    International Journal of Computer Vision133(9), 6332–6346 (2025)

    Cho, G., Kang, C., Soon, D., Joo, K.: Dogre- con: Canine prior-guided animatable 3d gaus- sian dog reconstruction from a single image: Dogrecon: Canine prior-guided animatable 3d gaussian... International Journal of Computer Vision133(9), 6332–6346 (2025)

    work page 2025

  5. [5]

    Communications of the ACM65(1), 99–106 (2021)

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM65(1), 99–106 (2021)

    2021

  6. [6]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp

    Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 5855– 5864 (2021)

    work page 2021

  7. [7]

    Interna- tional Journal of Computer Vision133(6), 3203–3221 (2025)

    Li, J., Yu, J., Wang, R., Gao, S.: Pseudo- plane regularized signed distance field for neural indoor scene reconstruction. Interna- tional Journal of Computer Vision133(6), 3203–3221 (2025)

    work page 2025

  8. [8]

    Inter- national Journal of Computer Vision133(1), 106–128 (2025)

    Xu, R., Yao, M., Chen, C., Wang, L., Xiong, Z.: Continuous spatial-spectral reconstruc- tion via implicit neural representation. Inter- national Journal of Computer Vision133(1), 106–128 (2025)

    work page 2025

  9. [9]

    ACM Transactions on Graphics (TOG)42(4), 1–14 (2023)

    Kerbl, B., Kopanas, G., Leimk¨ uhler, T., Dret- takis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG)42(4), 1–14 (2023)

    work page 2023

  10. [10]

    International Journal of Computer Vision (2026)

    Zhan, C., Zhang, Y., Lin, Y., Wang, G., Wang, H.: Rdg-gs: Relative depth guidance with gaussian splatting for real-time sparse- view 3d rendering. International Journal of Computer Vision (2026)

    work page 2026

  11. [11]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Yu, Z., Chen, A., Huang, B., Sattler, T., Geiger, A.: Mip-splatting: Alias-free 3d gaussian splatting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 19447–19456 (2024)

    work page 2024

  12. [12]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Lu, T., Yu, M., Xu, L., Xiangli, Y., Wang, L., Lin, D., Dai, B.: Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 20654–20664 (2024)

    work page 2024

  13. [13]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2025)

    Ren, K., Jiang, L., Lu, T., Yu, M., Xu, L., Ni, Z., Dai, B.: Octree-gs: Towards consis- tent real-time rendering with lod-structured 3d gaussians. IEEE Transactions on Pattern Analysis and Machine Intelligence (2025)

    work page 2025

  14. [14]

    In: European Conference on Computer Vision, pp

    Zhang, D., Wang, C., Wang, W., Li, P., Qin, M., Wang, H.: Gaussian in the wild: 3d gaus- sian splatting for unconstrained image collec- tions. In: European Conference on Computer Vision, pp. 341–359 (2024). Springer

    work page 2024

  15. [15]

    ACM Transactions on Graphics (TOG)43(4), 1–15 (2024)

    Kerbl, B., Meuleman, A., Kopanas, G., Wim- mer, M., Lanvin, A., Drettakis, G.: A hier- archical 3d gaussian representation for real- time rendering of very large datasets. ACM Transactions on Graphics (TOG)43(4), 1–15 (2024)

    work page 2024

  16. [16]

    arXiv preprint arXiv:2406.18198 (2024)

    Li, H., Li, J., Zhang, D., Wu, C., Shi, J., Zhao, C., Feng, H., Ding, E., Wang, J., Han, J.: Vdg: vision-only dynamic gaus- sian for driving simulation. arXiv preprint arXiv:2406.18198 (2024)

    Pith/arXiv arXiv 2024

  17. [17]

    arXiv preprint arXiv:2311.18561 (2023)

    Chen, Y., Gu, C., Jiang, J., Zhu, X., Zhang, L.: Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561 (2023)

    Pith/arXiv arXiv 2023

  18. [18]

    arXiv preprint arXiv:2405.20323 (2024)

    Huang, N., Wei, X., Zheng, W., An, P., Lu, M., Zhan, W., Tomizuka, M., Keutzer, K., Zhang, S.: S3gaussian: Self-supervised street gaussians for autonomous driving. arXiv preprint arXiv:2405.20323 (2024)

    Pith/arXiv arXiv 2024

  19. [19]

    arXiv preprint arXiv:2401.01339 (2024)

    Yan, Y., Lin, H., Zhou, C., Wang, W., Sun, H., Zhan, K., Lang, X., Zhou, X., Peng, S.: Street gaussians for model- ing dynamic urban scenes. arXiv preprint arXiv:2401.01339 (2024)

    Pith/arXiv arXiv 2024

  20. [20]

    10318– 10327 (2021)

    Zhou, X., Lin, Z., Shan, X., Wang, Y., Sun, D., Yang, M.-H.: Drivinggaussian: Composite gaussian splatting for surrounding dynamic 19 Vision and Pattern Recognition, pp. 10318– 10327 (2021)

    work page 2021

  21. [36]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, pp

    Park, S., Son, M., Jang, S., Ahn, Y.C., Kim, J.-Y., Kang, N.: Temporal interpolation is all you need for dynamic neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, pp. 4212–4221 (2023)

    work page 2023

  22. [37]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Li, Z., Wang, Q., Cole, F., Tucker, R., Snavely, N.: Dynibar: Neural dynamic image- based rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 4273–4284 (2023)

    work page 2023

  23. [38]

    In: SIGGRAPH Asia 2022 Conference Papers, pp

    Lin, H., Peng, S., Xu, Z., Yan, Y., Shuai, Q., Bao, H., Zhou, X.: Efficient neural radiance fields for interactive free-viewpoint video. In: SIGGRAPH Asia 2022 Conference Papers, pp. 1–9 (2022)

    work page 2022

  24. [39]

    ACM Transactions on Graphics (ToG)40(4), 1–13 (2021)

    Lombardi, S., Simon, T., Schwartz, G., Zoll- hoefer, M., Sheikh, Y., Saragih, J.: Mixture of volumetric primitives for efficient neural rendering. ACM Transactions on Graphics (ToG)40(4), 1–13 (2021)

    work page 2021

  25. [40]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Peng, S., Yan, Y., Shuai, Q., Bao, H., Zhou, X.: Representing volumetric videos as dynamic mlp maps. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 4252– 4262 (2023)

    work page 2023

  26. [41]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Cao, A., Johnson, J.: Hexplane: A fast repre- sentation for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 130–141 (2023)

    work page 2023

  27. [42]

    arXiv preprint arXiv:2308.09713 (2023)

    Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713 (2023)

    Pith/arXiv arXiv 2023

  28. [43]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene recon- struction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 20331–20341 (2024)

    work page 2024

  29. [44]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 20310–20320 (2024)

    work page 2024

  30. [45]

    FSGS: Real-Time Few-shot View Synthesis using Gaussian Splatting

    Zhu, Z., Fan, Z., Jiang, Y., Wang, Z.: Fsgs: Real-time few-shot view synthesis using gaussian splatting. arXiv preprint arXiv:2312.00451 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  31. [46]

    arXiv preprint arXiv:2311.17977 (2023)

    Jiang, Y., Tu, J., Liu, Y., Gao, X., Long, X., Wang, W., Ma, Y.: Gaussianshader: 3d gaussian splatting with shading func- tions for reflective surfaces. arXiv preprint arXiv:2311.17977 (2023)

    Pith/arXiv arXiv 2023

  32. [47]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp

    Wei, Y., Liu, S., Rao, Y., Zhao, W., Lu, J., Zhou, J.: Nerfingmvs: Guided optimiza- tion of neural radiance fields for indoor multi-view stereo. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 5610–5619 (2021)

    work page 2021

  33. [48]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Roessle, B., Barron, J.T., Mildenhall, B., Srinivasan, P.P., Nießner, M.: Dense depth priors for neural radiance fields from sparse input views. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 12892–12901 (2022)

    work page 2022

  34. [49]

    SparseNeRF: Distilling Depth Ranking for Few-shot Novel View Synthesis

    Wang, G., Chen, Z., Loy, C.C., Liu, Z.: Sparsenerf: Distilling depth ranking for few- shot novel view synthesis. arXiv preprint arXiv:2303.16196 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  35. [50]

    arXiv preprint arXiv:2403.17822 (2024)

    Turkulainen, M., Ren, X., Melekhov, I., Seiskari, O., Rahtu, E., Kannala, J.: Dn- splatter: Depth and normal priors for gaus- sian splatting and meshing. arXiv preprint arXiv:2403.17822 (2024)

    Pith/arXiv arXiv 2024

  36. [51]

    In: ACM SIG- GRAPH 2024 Conference Papers, pp

    Huang, B., Yu, Z., Chen, A., Geiger, A., Gao, S.: 2d gaussian splatting for geometri- cally accurate radiance fields. In: ACM SIG- GRAPH 2024 Conference Papers, pp. 1–11 (2024) 21

    work page 2024

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.