pith. sign in

arxiv: 2606.04349 · v2 · pith:SOMHVSRGnew · submitted 2026-06-03 · 💻 cs.CV · cs.AI

MorphoQuant: Modality-Aware Quantization for Omni-modal Large Language Models

Yue Wu , Changyuan Wang , Zixuan Wang , Shilin Ma , Yansong Tang This is my paper

Pith reviewed 2026-06-28 07:14 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords post-training quantizationomni-modal large language modelsmodality-aware quantizationoutlier compensation4-bit inferenceScienceQAmultimodal benchmarks
0
0 comments X

The pith

MorphoQuant lets 4-bit omni-modal LLMs surpass 16-bit accuracy on ScienceQA by absorbing modality-specific outliers into biases.

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

The paper claims that standard post-training quantization fails on omni-modal models because each modality produces its own extreme outlier patterns and distribution shapes. MorphoQuant counters this with two linked steps: Distribution-Aware Bias Compensation moves long-tailed outliers into per-channel biases so the remaining dense values can be discretized at high precision, and Morphology-Directed Quantization Function Optimization then tunes the quantization grid jointly with those biases to keep cross-modal alignment. On Qwen2.5-Omni the resulting W4A4 model reaches 76.63 percent on ScienceQA, beating both prior 4-bit methods and the 16-bit baseline. The same pattern holds on MMMU and Video-MME, showing that the accuracy-efficiency trade-off improves once modality morphology is explicitly protected during quantization.

Core claim

MorphoQuant is a modality-aware post-training quantization framework that preserves cross-modal morphology in omni-modal LLMs by using Distribution-Aware Bias Compensation to selectively absorb long-tailed outliers into channel-wise biases and Morphology-Directed Quantization Function Optimization to co-optimize the quantization grid with the bias mask, thereby enabling W4A4 models to outperform both state-of-the-art W4A4 methods and the W4A16 baseline on ScienceQA while maintaining performance across MMMU and Video-MME.

What carries the argument

Distribution-Aware Bias Compensation (DABC) that absorbs long-tailed outliers into channel-wise biases, paired with Morphology-Directed Quantization Function Optimization (MDQFO) that aligns the quantization grid to those biases across modalities.

If this is right

  • W4A4 quantization reaches 76.63 percent on ScienceQA and exceeds the W4A16 baseline.
  • The same W4A4 setting outperforms prior state-of-the-art 4-bit methods on MMMU and Video-MME.
  • Cross-modal distribution heterogeneity is mitigated without retraining the base model.
  • The accuracy-efficiency trade-off improves for any omni-modal LLM that exhibits similar per-modality outlier patterns.

Where Pith is reading between the lines

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

  • The same bias-absorption step could be tested on other multimodal architectures that mix vision, audio, and text to check whether the modality-specific handling generalizes.
  • If the method scales to even lower bit widths, it would directly reduce memory footprint for on-device omni-modal inference.
  • The co-optimization of grid and bias mask suggests that future quantizers may need joint search procedures rather than independent per-tensor scaling.

Load-bearing premise

Absorbing long-tailed outliers into channel-wise biases and co-optimizing the quantization grid will preserve cross-modal morphology and discretization accuracy without creating new errors or modality-specific degradation.

What would settle it

If the W4A4 MorphoQuant model scores below the W4A16 baseline on ScienceQA when re-run on the same Qwen2.5-Omni checkpoint and evaluation protocol, the central performance claim is falsified.

Figures

Figures reproduced from arXiv: 2606.04349 by Changyuan Wang, Shilin Ma, Yansong Tang, Yue Wu, Zixuan Wang.

Figure 1
Figure 1. Figure 1: Comparison with tradi￾tional methods. (Top) Traditional mixed-precision methods necessitate memory-interleaved operations (blue￾INT4, green-FP16), suffering from fragmented memory access and se￾vere overhead. (Bottom) Our DABC strictly decouples the computation, skilfully dividing the activation into the vast majority in pure INT4 Ten￾sor Cores and the sparse outliers (red) undergoing a lightweight project… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our omni-modal quantization framework. (a) OLLM Architecture: Omni-modal inputs converge into Qwen2.5- Omni, leading to severe distribution disparity. (b) Modality-Aware Distribution Separation: We identify long-tailed channels via disper￾sion score Dc, separating activations into symmetric inliers and sparse outliers. (d) Distribution-Directed Bias Compensation: Outliers are projected and merg… view at source ↗
Figure 3
Figure 3. Figure 3: (Left) A layer with extreme long tail, while the inliers dominate a tiny range. (Middle) PCA of activation inputs (Video-MME), indicating cross￾modal disparity, long tail and sparse outliers. (Right) Quantization error of an attention map, in which the primary error lies along the modal boundary line. Based on these observation, we conclude that conventional methods ap￾pear substantial discretization error… view at source ↗
Figure 4
Figure 4. Figure 4: Hardware-friendly Execution Flow. (Top) Traditional mixed￾precision methods suffer from memory-interleaved operations. (Bottom) Our approach executes the vast majority of activations via standard dense 4-bit CUDA kernels, while integrating sparse outlier corrections through lightweight element-wise addition, ensuring hardware-friendly inference. 3.4 Morphology-Directed Quantization Function Optimization Bu… view at source ↗
Figure 5
Figure 5. Figure 5: PCA visualization of activation features (Thinker Layer 0). Com￾pared to (a) the FP16 oracle, (b) traditional W4A4 quantization like Q-VLM suffers from severe representation collapse, losing the significant outliers. Conversely, (c) our method faithfully restores the original feature manifold and preserved outliers, preserving cross-modal semantic alignment. Variant DABC Collab. Search Comp. Loss Accuracy … view at source ↗
Figure 6
Figure 6. Figure 6: Ablation study on hyper-parameters. (Left) The answering accu￾racy w.r.t. different λcos on ScienceQA dataset. (Right) The accuracy and com￾pensation percentage w.r.t. different expansion rate on ScienceQA dataset. 5 Conclusion In this paper, we proposed a novel PTQ framework specifically tailored for MLLMs, with successful deployment and rigorous evaluation on the Qwen2.5- Omni architecture. Departing fro… view at source ↗
read the original abstract

Conventional Post-Training Quantization (PTQ) methods struggle with 4-bit Omni-modal Large Language Models (OLLMs) due to the extreme distribution heterogeneity and disparate outlier patterns across modalities. To address this, we propose MorphoQuant, a modality-aware PTQ framework engineered to preserve cross-modal morphology and mitigate outlier loss. Specifically, we introduce Distribution-Aware Bias Compensation (DABC), which selectively absorbs long-tailed outliers into channel-wise biases. This mechanism safeguards outlier magnitudes while maintaining high-precision discretization for dense inliers, thereby preserving accurate discretization across diverse modal distribution. Complementing this, we propose Morphology-Directed Quantization Function Optimization (MDQFO) to co-optimize the quantization grid with the bias mask, ensuring fine-grained alignment across modalities. Extensive evaluations on Qwen2.5-Omni across benchmarks like MMMU and Video-MME demonstrate our approach's superiority. Notably, our W4A4 model achieves 76.63% on ScienceQA, significantly outperforming SOTA W4A4 methods and surprisingly surpassing the W4A16 baseline, which fully demonstrates the exceptional accuracy-efficiency trade-off of our framework.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes MorphoQuant, a modality-aware post-training quantization (PTQ) framework for omni-modal large language models. It introduces Distribution-Aware Bias Compensation (DABC) to selectively absorb long-tailed outliers into channel-wise biases and Morphology-Directed Quantization Function Optimization (MDQFO) to co-optimize the quantization grid with the bias mask. The central empirical claim is that the resulting W4A4 model on Qwen2.5-Omni reaches 76.63% on ScienceQA, outperforming both prior SOTA W4A4 methods and the W4A16 baseline while also showing gains on MMMU and Video-MME.

Significance. If the DABC/MDQFO mechanisms can be shown to preserve inlier discretization fidelity and cross-modal interactions without introducing new modality-specific errors, the work would be significant for practical deployment of omni-modal LLMs. The reported outperformance of a 4-bit model over a 16-bit baseline is a strong result if the supporting formulations, ablations, and error analysis are provided.

major comments (2)
  1. [Abstract] Abstract: the headline claim that W4A4 reaches 76.63% on ScienceQA and surpasses the W4A16 baseline is load-bearing for the accuracy-efficiency argument, yet the abstract supplies no explicit update rule for the bias mask in DABC, no objective function for MDQFO, and no derivation showing that outlier absorption is norm-preserving; without these the observed gain cannot be attributed to improved quantization rather than an artifact of bias injection.
  2. [Method] Method section (DABC/MDQFO): the assumption that selectively absorbing long-tailed outliers into channel-wise biases and co-optimizing the grid preserves cross-modal morphology and accurate inlier discretization is central but unsupported by any stated proof, norm-preservation argument, or ablation isolating the contribution of each component from potential side-effects.
minor comments (2)
  1. [Abstract] Abstract: the terms 'cross-modal morphology' and 'modality-aware' are used without operational definition or reference to prior usage.
  2. [Abstract] Abstract: 'W4A4' and 'W4A16' should be defined on first use rather than assumed known.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback highlighting the need for greater clarity on the theoretical foundations and empirical isolation of our proposed components. We address each major comment below and outline planned revisions to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the headline claim that W4A4 reaches 76.63% on ScienceQA and surpasses the W4A16 baseline is load-bearing for the accuracy-efficiency argument, yet the abstract supplies no explicit update rule for the bias mask in DABC, no objective function for MDQFO, and no derivation showing that outlier absorption is norm-preserving; without these the observed gain cannot be attributed to improved quantization rather than an artifact of bias injection.

    Authors: We agree the abstract's brevity omits explicit equations. The bias-mask update rule appears in Eq. (2), the MDQFO objective in Eq. (4), and the norm-preservation argument via selective absorption is derived in Sec. 3.1. Ablations in Sec. 4.3 and error analysis in Sec. 4.4 attribute gains to the mechanisms rather than bias artifacts. We will revise the abstract to reference these elements concisely while preserving length constraints. revision: partial

  2. Referee: [Method] Method section (DABC/MDQFO): the assumption that selectively absorbing long-tailed outliers into channel-wise biases and co-optimizing the grid preserves cross-modal morphology and accurate inlier discretization is central but unsupported by any stated proof, norm-preservation argument, or ablation isolating the contribution of each component from potential side-effects.

    Authors: The manuscript already provides component-wise ablations (Table 2, Fig. 5) isolating DABC and MDQFO contributions, plus modality-specific error breakdowns in Sec. 4.4 showing preserved inlier fidelity and cross-modal consistency. While a formal proof is absent, we will add a short norm-preservation derivation based on the bias-compensation formulation and expand the ablation discussion to explicitly rule out side-effects. revision: yes

Circularity Check

0 steps flagged

No circularity: claims rest on empirical benchmarks, not self-referential definitions or fitted inputs

full rationale

The paper introduces DABC and MDQFO as new PTQ components for handling modality-specific distributions and outliers in OLLMs, then reports concrete benchmark numbers (e.g., W4A4 at 76.63% on ScienceQA). No equation or step is shown to define a quantity in terms of itself, rename a fitted parameter as a prediction, or rely on a self-citation chain whose prior result is unverified. The derivation chain consists of proposed algorithmic mechanisms whose validity is asserted via external evaluation rather than by construction from the target metrics.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities; the approach implicitly assumes that modality-specific distribution handling can be achieved through bias compensation and grid optimization without further justification.

pith-pipeline@v0.9.1-grok · 5743 in / 1090 out tokens · 29407 ms · 2026-06-28T07:14:33.323485+00:00 · methodology

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

38 extracted references · 15 canonical work pages · 4 internal anchors

  1. [1]

    Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution

    Wang, P., Bai, S., Tan, S., Wang, S., Fan, Z., Bai, J., Chen, K., Liu, X., Wang, J., Ge, W., et al.: Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution. arXiv preprint arXiv:2409.12191 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  2. [2]

    Qwen3-Omni Technical Report

    Xu, J., Guo, Z., Hu, H., Chu, Y., Wang, X., He, J., Wang, Y., Shi, X., He, T., Zhu, X., et al.: Qwen3-omni technical report. arXiv preprint arXiv:2509.17765 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  3. [3]

    Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context

    Team, G., Georgiev, P., Lei, V.I., Burnell, R., Bai, L., Gulati, A., Tanzer, G., Vincent, D., Pan, Z., Wang, S., et al.: Gemini 1.5: Unlocking multimodal un- derstanding across millions of tokens of context. arXiv preprint arXiv:2403.05530 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  4. [4]

    Visual Intelligence 3(1) (2025) 27

    Jin, Y., Li, J., Gu, T., Liu, Y., Zhao, B., Lai, J., Gan, Z., Wang, Y., Wang, C., Tan, X., et al.: Efficient multimodal large language models: A survey. Visual Intelligence 3(1) (2025) 27

    2025

  5. [5]

    arXiv preprint arXiv:2404.07214 (2024)

    Ghosh, A., Acharya, A., Saha, S., Jain, V., Chadha, A.: Exploring the frontier of vision-language models: A survey of current methodologies and future directions. arXiv preprint arXiv:2404.07214 (2024)

    work page arXiv 2024

  6. [6]

    Findings of the association for computational linguistics: ACL 2024 (2024) 13590– 13618

    Caffagni, D., Cocchi, F., Barsellotti, L., Moratelli, N., Sarto, S., Baraldi, L., Cornia, M., Cucchiara, R.: The revolution of multimodal large language models: A survey. Findings of the association for computational linguistics: ACL 2024 (2024) 13590– 13618

    2024

  7. [7]

    IEEE Transactions on Pattern Analysis and Machine Intelligence45(10) (2023) 12113–12132

    Xu, P., Zhu, X., Clifton, D.A.: Multimodal learning with transformers: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence45(10) (2023) 12113–12132

    2023

  8. [8]

    IEEE Access (2025)

    Bhuyan, M.S.M., Hossain, E., Sathi, K.A., Hossain, M.A., Dewan, M.A.A.: Bvqa: connecting language and vision through multimodal attention for open-ended ques- tion answering. IEEE Access (2025)

    2025

  9. [9]

    arXiv preprint arXiv:2505.05108 (2025)

    Feng, Z., Xue, R., Yuan, L., Yu, Y., Ding, N., Liu, M., Gao, B., Sun, J., Zheng, X., Wang, G.: Multi-agent embodied ai: Advances and future directions. arXiv preprint arXiv:2505.05108 (2025)

    work page arXiv 2025

  10. [10]

    arXiv preprint arXiv:2508.08706 (2025)

    Cheng, Z., Zhang, Y., Zhang, W., Li, H., Wang, K., Song, L., Zhang, H.: Omnivtla: Vision-tactile-language-action model with semantic-aligned tactile sensing. arXiv preprint arXiv:2508.08706 (2025)

    work page arXiv 2025

  11. [11]

    Artificial Intelligence Review58(12) (2025) 403

    Liu, Y., Cao, J., Liu, C., Ding, K., Jin, L.: Datasets for large language models: A comprehensive survey. Artificial Intelligence Review58(12) (2025) 403

    2025

  12. [12]

    arXiv preprint arXiv:2408.08632 (2024)

    Li, J., Lu, W., Fei, H., Luo, M., Dai, M., Xia, M., Jin, Y., Gan, Z., Qi, D., Fu, C., et al.: A survey on benchmarks of multimodal large language models. arXiv preprint arXiv:2408.08632 (2024)

    work page arXiv 2024

  13. [13]

    Advances in Neural Information Processing Systems 36(2023) 26650–26685

    Yin, Z., Wang, J., Cao, J., Shi, Z., Liu, D., Li, M., Huang, X., Wang, Z., Sheng, L., Bai, L., et al.: Lamm: Language-assisted multi-modal instruction-tuning dataset, framework, and benchmark. Advances in Neural Information Processing Systems 36(2023) 26650–26685

    2023

  14. [14]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Ye, W., Wu, Q., Lin, W., Zhou, Y.: Fit and prune: Fast and training-free visual token pruning for multi-modal large language models. In: Proceedings of the AAAI Conference on Artificial Intelligence. Volume 39. (2025) 22128–22136

    2025

  15. [15]

    arXiv preprint arXiv:2401.16160 (2024)

    Chen, S., Jie, Z., Ma, L.: Llava-mole: Sparse mixture of lora experts for mitigating data conflicts in instruction finetuning mllms. arXiv preprint arXiv:2401.16160 (2024)

    work page arXiv 2024

  16. [16]

    arXiv preprint arXiv:2401.00625 (2024) 16

    Bai, G., Chai, Z., Ling, C., Wang, S., Lu, J., Zhang, N., Shi, T., Yu, Z., Zhu, M., Zhang, Y., et al.: Beyond efficiency: A systematic survey of resource-efficient large language models. arXiv preprint arXiv:2401.00625 (2024) 16

    work page arXiv 2024

  17. [17]

    In: Inter- national conference on machine learning, PMLR (2023) 38087–38099

    Xiao, G., Lin, J., Seznec, M., Wu, H., Demouth, J., Han, S.: Smoothquant: Accu- rate and efficient post-training quantization for large language models. In: Inter- national conference on machine learning, PMLR (2023) 38087–38099

    2023

  18. [18]

    Proceedings of machine learning and systems6 (2024) 87–100

    Lin, J., Tang, J., Tang, H., Yang, S., Chen, W.M., Wang, W.C., Xiao, G., Dang, X., Gan, C., Han, S.: Awq: Activation-aware weight quantization for on-device llm compression and acceleration. Proceedings of machine learning and systems6 (2024) 87–100

    2024

  19. [19]

    In: Proceedings of the 32nd ACM International Conference on Multimedia

    Xie, J., Zhang, Y., Lin, M., Cao, L., Ji, R.: Advancing multimodal large lan- guage models with quantization-aware scale learning for efficient adaptation. In: Proceedings of the 32nd ACM International Conference on Multimedia. (2024) 10582–10591

    2024

  20. [21]

    In: Proceedings of the 33rd ACM International Conference on Multimedia

    Yu, J., Zhou, S., Yang, D., Li, S., Wang, S., Hu, X., Xu, C., Xu, Z., Shu, C., Yuan, Z.: Mquant: Unleashing the inference potential of multimodal large language models via static quantization. In: Proceedings of the 33rd ACM International Conference on Multimedia. (2025) 1783–1792

    2025

  21. [22]

    In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition

    Wang, C., Wang, Z., Xu, X., Tang, Y., Zhou, J., Lu, J.: Towards accurate post- training quantization for diffusion models. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition. (2024) 16026–16035

    2024

  22. [23]

    Advances in Neural Information Processing Systems37(2024) 114553–114573

    Wang, C., Wang, Z., Xu, X., Tang, Y., Zhou, J., Lu, J.: Q-vlm: Post-training quantization for large vision-language models. Advances in Neural Information Processing Systems37(2024) 114553–114573

    2024

  23. [24]

    Xu, J., Guo, Z., He, J., Hu, H., He, T., Bai, S., Chen, K., Wang, J., Fan, Y., Dang, K., Zhang, B., Wang, X., Chu, Y., Lin, J.: Qwen2.5-omni technical report (2025)

    2025

  24. [25]

    Visual Intelligence2(1) (2024) 36

    Huang, W., Zheng, X., Ma, X., Qin, H., Lv, C., Chen, H., Luo, J., Qi, X., Liu, X., Magno, M.: An empirical study of llama3 quantization: From llms to mllms. Visual Intelligence2(1) (2024) 36

    2024

  25. [26]

    int8 (): 8-bit matrix multiplication for transformers at scale

    Dettmers, T., Lewis, M., Belkada, Y., Zettlemoyer, L.: Gpt3. int8 (): 8-bit matrix multiplication for transformers at scale. Advances in neural information processing systems35(2022) 30318–30332

    2022

  26. [27]

    arXiv preprint arXiv:2306.03078 (2023)

    Dettmers, T., Svirschevski, R., Egiazarian, V., Kuznedelev, D., Frantar, E., Ashk- boos, S., Borzunov, A., Hoefler, T., Alistarh, D.: Spqr: A sparse-quantized repre- sentation for near-lossless llm weight compression. arXiv preprint arXiv:2306.03078 (2023)

    work page arXiv 2023

  27. [28]

    Proceedings of Machine Learning and Systems6(2024) 196–209

    Zhao, Y., Lin, C.Y., Zhu, K., Ye, Z., Chen, L., Zheng, S., Ceze, L., Krishnamurthy, A., Chen, T., Kasikci, B.: Atom: Low-bit quantization for efficient and accurate llm serving. Proceedings of Machine Learning and Systems6(2024) 196–209

    2024

  28. [29]

    Ad- vances in neural information processing systems35(2022) 27168–27183

    Yao, Z., Yazdani Aminabadi, R., Zhang, M., Wu, X., Li, C., He, Y.: Zeroquant: Efficient and affordable post-training quantization for large-scale transformers. Ad- vances in neural information processing systems35(2022) 27168–27183

    2022

  29. [30]

    arXiv preprint arXiv:2006.16669 (2020)

    Wu, D., Tang, Q., Zhao, Y., Zhang, M., Fu, Y., Zhang, D.: Easyquant: Post-training quantization via scale optimization. arXiv preprint arXiv:2006.16669 (2020)

    work page arXiv 2006

  30. [31]

    Billm: Pushing the limit of post-training quantization for llms.arXiv preprint arXiv:2402.04291, 2024

    Huang, W., Liu, Y., Qin, H., Li, Y., Zhang, S., Liu, X., Magno, M., Qi, X.: Billm: Pushing the limit of post-training quantization for llms. arXiv preprint arXiv:2402.04291 (2024)

    work page arXiv 2024

  31. [32]

    Advances in neural information processing systems36 (2023) 10088–10115 17

    Dettmers, T., Pagnoni, A., Holtzman, A., Zettlemoyer, L.: Qlora: Efficient fine- tuning of quantized llms. Advances in neural information processing systems36 (2023) 10088–10115 17

    2023

  32. [33]

    Omniquant: Omnidirectionally calibrated quantization for large language models.arXiv preprint arXiv:2308.13137, 2023

    Shao, W., Chen, M., Zhang, Z., Xu, P., Zhao, L., Li, Z., Zhang, K., Gao, P., Qiao, Y., Luo, P.: Omniquant: Omnidirectionally calibrated quantization for large language models. arXiv preprint arXiv:2308.13137 (2023)

    work page arXiv 2023

  33. [34]

    SpinQuant: LLM quantization with learned rotations

    Liu, Z., Zhao, C., Fedorov, I., Soran, B., Choudhary, D., Krishnamoorthi, R., Chan- dra, V., Tian, Y., Blankevoort, T.: Spinquant: Llm quantization with learned rotations. arXiv preprint arXiv:2405.16406 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  34. [35]

    Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery15(3) (2025) e70036

    Shinde, G., Ravi, A., Dey, E., Sakib, S., Rampure, M., Roy, N.: A survey on efficient vision-language models. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery15(3) (2025) e70036

    2025

  35. [36]

    Bi-vlm: Pushing ultra-low precision post-training quantization boundaries in vision-language models

    Wang, X., Huang, J., Abdalla, R., Zhang, C., Xian, R., Manocha, D.: Bi-vlm: Push- ing ultra-low precision post-training quantization boundaries in vision-language models. arXiv preprint arXiv:2509.18763 (2025)

    work page arXiv 2025

  36. [37]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Li, S., Hu, Y., Ning, X., Liu, X., Hong, K., Jia, X., Li, X., Yan, Y., Ran, P., Dai, G., et al.: Mbq: Modality-balanced quantization for large vision-language models. In: Proceedings of the Computer Vision and Pattern Recognition Conference. (2025) 4167–4177

    2025

  37. [38]

    In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing

    Wei, X., Zhang, Y., Li, Y., Zhang, X., Gong, R., Guo, J., Liu, X.: Outlier suppres- sion+: Accurate quantization of large language models by equivalent and effective shifting and scaling. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. (2023) 1648–1665

    2023

  38. [39]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.: Lavt: Language- aware vision transformer for referring image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. (2022) 18155– 18165

    2022