arxiv: 2601.01322 · v2 · submitted 2026-01-04 · 💻 cs.CV · cs.AI· cs.LG· cs.MM· eess.IV

Recognition: 2 theorem links

· Lean Theorem

LinMU: Multimodal Understanding Made Linear

Hongjie Wang , Niraj K. Jha

Authors on Pith no claims yet

Pith reviewed 2026-05-16 18:29 UTC · model grok-4.3

classification 💻 cs.CV cs.AIcs.LGcs.MMeess.IV

keywords linear complexityvision-language modelsstate-space modelsdistillationmultimodal understandingM-MATE blockself-attention replacementlong video processing

0 comments

The pith

LinMU replaces self-attention in vision-language models with a linear dual-branch module that matches teacher performance on multimodal benchmarks.

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

The paper establishes that vision-language models can reach state-of-the-art multimodal understanding using only linear-complexity operations instead of quadratic self-attention. It introduces LinMU, which substitutes every self-attention layer with an M-MATE block that pairs a bidirectional state-space model for global context with localized window attention for nearby tokens. A three-stage distillation process transfers knowledge from a pre-trained teacher model by initializing branches, jointly fine-tuning them, and then adapting the rest with LoRA while matching hidden states and logits. A sympathetic reader would care because this removes the quadratic cost barrier that currently blocks high-resolution images and minute-length videos from running efficiently on standard hardware.

Core claim

LinMU replaces every self-attention layer in the language model decoder with an M-MATE block: a dual-branch module that combines a bidirectional state-space model branch (Flex-MA) for global context with a localized Swin-style window attention branch (Local-Swin) for adjacent correlations. A three-stage distillation framework converts a pre-trained VLM into this architecture by first training the Flex-MA branch alone from self-attention weights, then jointly fine-tuning both branches, and finally adapting the remaining blocks with LoRA adapters while regressing on the teacher’s hidden states and token logits. On MMMU, TextVQA, LongVideoBench, Video-MME and similar benchmarks, the resulting模型

What carries the argument

M-MATE block: dual-branch replacement for self-attention consisting of a bidirectional state-space model for global context and localized Swin-style window attention for adjacent correlations.

Load-bearing premise

The combination of a bidirectional state-space model branch and localized window attention can fully substitute for global self-attention across multimodal reasoning tasks without measurable degradation after the three-stage distillation.

What would settle it

Accuracy measurements on a long-video or high-resolution image benchmark where LinMU falls measurably below the teacher model on tasks that require full global context.

Figures

Figures reproduced from arXiv: 2601.01322 by Hongjie Wang, Niraj K. Jha.

**Figure 2.** Figure 2: The overview of LinMU, in which the Vision Encoder is untouched but all the attention layers [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Figure 3: The receptive field comparison between native Mamba2 (suitable for text-only tasks), bidirectional [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: Window division, shifting, and re-combination in the local 3D Swin attention. After window [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

**Figure 5.** Figure 5: Our proposed three-stage distillation pipeline to replace attention layers in pre-trained VLMs with [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗

**Figure 6.** Figure 6: Results are measured on NVIDIA H100 GPUs. Note that we exclude the vision encoding time here [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗

**Figure 6.** Figure 6: Efficiency comparison between the quadratic-complexity teacher model NVILA and the linear [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗

**Figure 7.** Figure 7: Efficiency comparison between the quadratic-complexity teacher model Qwen2.5-VL-7B-Instruct [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗

read the original abstract

Modern Vision-Language Models (VLMs) achieve impressive performance but are limited by the quadratic complexity of self-attention, which prevents their deployment on edge devices and makes their understanding of high-resolution images and long-context videos prohibitively expensive. To address this challenge, we introduce LinMU (Linear-complexity Multimodal Understanding), a VLM design that achieves linear complexity for the language model decoder without using any quadratic-complexity modules while maintaining the performance of global-attention-based VLMs. LinMU replaces every self-attention layer in the language model decoder with an M-MATE block: a dual-branch module that combines a bidirectional state-space model for global context (Flex-MA branch) with localized Swin-style window attention (Local-Swin branch) for adjacent correlations. To transform a pre-trained VLM into the LinMU architecture, we propose a three-stage distillation framework that (i) initializes both branches with self-attention weights and trains the Flex-MA branch alone, (ii) unfreezes the Local-Swin branch and fine-tunes it jointly with the Flex-MA branch, and (iii) unfreezes the remaining blocks and fine-tunes them using LoRA adapters, while regressing on hidden states and token-level logits of the frozen VLM teacher. On MMMU, TextVQA, LongVideoBench, Video-MME, and other benchmarks, LinMU matches the performance of teacher models, yet reduces Time-To-First-Token (TTFT) by up to 2.7$\times$ and improves token throughput by up to 9.0$\times$ on minute-length videos. Ablations confirm the importance of each distillation stage and the necessity of the two branches of the M-MATE block. The proposed framework demonstrates that state-of-the-art multimodal reasoning can be achieved without quadratic attention, thus opening up avenues for long-context VLMs that can deal with high-resolution images and long videos.

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.

Desk Editor's Note private letter to a colleague

LinMU swaps in a dual-branch M-MATE decoder and three-stage distillation to hit linear complexity with reported matching scores and clear speedups on long videos, but the long-range fidelity claim rests on unshown transfer details.

read the letter

LinMU replaces every self-attention layer in the VLM decoder with an M-MATE block that runs a bidirectional SSM branch for global context alongside a fixed-window local attention branch. A three-stage distillation process turns a frozen teacher into this linear form by first training the SSM alone, then jointly tuning the local branch, and finally applying LoRA while regressing on teacher hidden states and logits. The headline result is that the model matches teacher scores on MMMU, TextVQA, LongVideoBench, and Video-MME while cutting TTFT by up to 2.7x and raising throughput by up to 9x on minute-scale videos. Ablations show both branches and all three stages matter for the reported numbers. That architecture and the concrete distillation schedule are the actual new pieces; they assemble existing SSM and window ideas into a drop-in linear decoder rather than inventing a wholly new primitive. The efficiency numbers look usable for edge or long-context work if they hold. The soft spot is the untested assumption that the SSM-plus-window combination recovers the same long-range cross-modal mixing the teacher computes directly. No table or section quantifies how much of the teacher's distant token interactions survive the distillation on hard cases, and the abstract gives no error bars or full training curves. If the benchmarks lean on patterns that local-plus-recurrent approximations can solve, the equivalence claim could be narrower than stated. This is for groups already working on efficient VLMs or SSM replacements who need a tested recipe they can reproduce. A reader who wants concrete speed numbers on real video lengths will find value here. It deserves peer review because the architecture is explicit, the efficiency gains are measurable, and the gaps are fixable with additional ablations rather than fatal.

Referee Report

3 major / 2 minor

Summary. The paper introduces LinMU, a VLM architecture that replaces all self-attention layers in the language decoder with M-MATE blocks. Each block combines a bidirectional state-space model branch (Flex-MA) for global context with a localized Swin-style window attention branch (Local-Swin) for adjacent correlations, achieving linear complexity. A three-stage distillation process initializes the branches from a frozen teacher VLM, jointly tunes them, and applies LoRA while regressing on hidden states and logits. The authors report that LinMU matches teacher performance on MMMU, TextVQA, LongVideoBench, Video-MME and related benchmarks while delivering up to 2.7× lower TTFT and 9.0× higher token throughput on minute-length videos.

Significance. If the performance equivalence and robustness claims hold, the work would be significant for enabling efficient long-context and high-resolution multimodal models on edge devices. The dual-branch linear design and staged distillation offer a concrete route to remove quadratic attention without sacrificing multimodal reasoning, with potential impact on deployment of VLMs for video understanding.

major comments (3)

[Benchmark results] Benchmark results section: the claim of matching teacher performance on MMMU, TextVQA, LongVideoBench, and Video-MME is presented without error bars, standard deviations across runs, or explicit data-exclusion criteria; this weakens the equivalence assertion because the central efficiency claim rests on summarized point estimates whose statistical reliability cannot be assessed.
[§3] Three-stage distillation framework (described in §3): no quantitative metrics are supplied on how faithfully the Flex-MA bidirectional SSM recovers the teacher's long-range cross-modal token interactions after stages (i)–(iii); hidden-state cosine similarity or logit KL divergence broken down by token distance on minute-length videos would directly test whether the linear recurrence approximates the quadratic mixing that the headline claim requires.
[Ablation studies] Ablation studies: while the necessity of both branches and all three stages is asserted, the paper does not report the performance drop on global-reasoning subsets when the Local-Swin window size is reduced or when Flex-MA is replaced by a unidirectional SSM; such controls are needed to substantiate that the combination fully substitutes for global self-attention without measurable degradation.

minor comments (2)

[§3.1] Clarify in §3.1 the precise fusion mechanism inside the M-MATE block (e.g., how Flex-MA and Local-Swin outputs are combined before the feed-forward layer) and whether any additional parameters are introduced beyond the stated LoRA adapters.
[Supplementary material] Add full training curves and per-stage loss trajectories to the supplementary material to allow readers to verify convergence behavior of the distillation objectives.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point by point below and will incorporate revisions to strengthen the statistical rigor and design validation of the manuscript.

read point-by-point responses

Referee: Benchmark results section: the claim of matching teacher performance on MMMU, TextVQA, LongVideoBench, and Video-MME is presented without error bars, standard deviations across runs, or explicit data-exclusion criteria; this weakens the equivalence assertion because the central efficiency claim rests on summarized point estimates whose statistical reliability cannot be assessed.

Authors: We agree that error bars and standard deviations are important for assessing statistical reliability. In the revised manuscript we will report mean performance and standard deviations across multiple independent runs for all benchmarks. We will also add an explicit statement confirming that all results use standard benchmark splits with no data exclusion. revision: yes
Referee: Three-stage distillation framework (described in §3): no quantitative metrics are supplied on how faithfully the Flex-MA bidirectional SSM recovers the teacher's long-range cross-modal token interactions after stages (i)–(iii); hidden-state cosine similarity or logit KL divergence broken down by token distance on minute-length videos would directly test whether the linear recurrence approximates the quadratic mixing that the headline claim requires.

Authors: We thank the referee for this valuable suggestion. We will add quantitative analysis in the revised paper, including hidden-state cosine similarity and logit KL divergence metrics broken down by token distance on minute-length videos, to directly demonstrate that the Flex-MA branch recovers the teacher's long-range interactions after each distillation stage. revision: yes
Referee: Ablation studies: while the necessity of both branches and all three stages is asserted, the paper does not report the performance drop on global-reasoning subsets when the Local-Swin window size is reduced or when Flex-MA is replaced by a unidirectional SSM; such controls are needed to substantiate that the combination fully substitutes for global self-attention without measurable degradation.

Authors: We agree these targeted controls would better substantiate the design. In the revised manuscript we will include additional ablation results on global-reasoning subsets, reporting performance when the Local-Swin window size is reduced and when Flex-MA is replaced by a unidirectional SSM variant, to confirm the necessity of the bidirectional global branch and chosen window size. revision: yes

Circularity Check

0 steps flagged

No significant circularity; architecture and distillation are independent of target metrics

full rationale

The paper defines the M-MATE block (Flex-MA bidirectional SSM + Local-Swin window attention) as a direct architectural replacement for self-attention, with linear complexity arising from the SSM recurrence and fixed-window design rather than any fitted quantity. The three-stage distillation explicitly regresses on hidden states and logits from a frozen external teacher VLM, so benchmark matching is an empirical outcome of that objective rather than a prediction forced by construction. No equations reduce a claimed result to its inputs, no uniqueness theorem is imported from self-citation, and no ansatz is smuggled via prior work by the same authors. The efficiency gains (TTFT and throughput) follow directly from the linear modules once the weights are obtained. This is a standard distillation setup with independent architectural choices; the derivation chain does not collapse.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 3 invented entities

The design rests on the assumption that state-space models plus local windows can capture the same information as global attention after distillation; no new physical entities are postulated, but the M-MATE block itself is an invented architectural component whose sufficiency is justified only by the reported benchmarks.

free parameters (1)

LoRA rank and scaling factors
Introduced in stage three of distillation; their values are chosen to match teacher hidden states and logits.

axioms (2)

domain assumption Bidirectional state-space models can represent global context at linear cost
Invoked when replacing self-attention with the Flex-MA branch.
domain assumption Localized window attention suffices for adjacent token correlations
Basis for the Local-Swin branch.

invented entities (3)

M-MATE block no independent evidence
purpose: Dual-branch replacement for self-attention that achieves linear complexity
New architectural unit combining Flex-MA and Local-Swin branches.
Flex-MA branch no independent evidence
purpose: Bidirectional state-space model for global context
Invented component within M-MATE.
Local-Swin branch no independent evidence
purpose: Localized window attention for adjacent correlations
Invented component within M-MATE.

pith-pipeline@v0.9.0 · 5659 in / 1633 out tokens · 40908 ms · 2026-05-16T18:29:39.852441+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

LinMU replaces every self-attention layer ... with an M-MATE block: a dual-branch module that combines a bidirectional state-space model ... with localized Swin-style window attention
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

three-stage distillation framework that (i) initializes both branches with self-attention weights and trains the Flex-MA branch alone

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages · 13 internal anchors

[1]

Qwen2.5-VL Technical Report

Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2.5-VL technical report.arXiv preprint arXiv:2502.13923,

work page internal anchor Pith review Pith/arXiv arXiv
[2]

Efficient hybrid mamba- transformer reasoning model.arXiv preprint arXiv:2508.14444, 2025

Aarti Basant, Abhijit Khairnar, Abhijit Paithankar, Abhinav Khattar, Adithya Renduchintala, Aditya Malte, Akhiad Bercovich, Akshay Hazare, Alejandra Rico, Aleksander Ficek, et al. NVIDIA Nemotron Nano 2: An accurate and efficient hybrid Mamba-Transformer reasoning model.arXiv preprint arXiv:2508.14444,

work page arXiv
[3]

Longformer: The Long-Document Transformer

Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer.arXiv preprint arXiv:2004.05150,

work page internal anchor Pith review Pith/arXiv arXiv 2004
[4]

Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling

Jieneng Chen, Luoxin Ye, Ju He, Zhao-Yang Wang, Daniel Khashabi, and Alan Yuille. Efficient large multi- modal models via visual context compression.Advances in Neural Information Processing Systems, 37: 73986–74007, 2024a. Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2...

work page internal anchor Pith review Pith/arXiv arXiv
[5]

NVIDIA Nemotron Nano V2 VL.arXiv preprint arXiv:2511.03929,

Amala Sanjay Deshmukh, Kateryna Chumachenko, Tuomas Rintamaki, Matthieu Le, Tyler Poon, Da- nial Mohseni Taheri, Ilia Karmanov, Guilin Liu, Jarno Seppanen, Guo Chen, et al. NVIDIA Nemotron Nano V2 VL.arXiv preprint arXiv:2511.03929,

work page arXiv
[6]

Streaming video question-answering with in-context video KV-cache re- trieval.arXiv preprint arXiv:2503.00540,

Shangzhe Di, Zhelun Yu, Guanghao Zhang, Haoyuan Li, Tao Zhong, Hao Cheng, Bolin Li, Wanggui He, Fangxun Shu, and Hao Jiang. Streaming video question-answering with in-context video KV-cache re- trieval.arXiv preprint arXiv:2503.00540,

work page arXiv
[7]

Efficiently Modeling Long Sequences with Structured State Spaces

Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces.arXiv preprint arXiv:2111.00396,

work page internal anchor Pith review Pith/arXiv arXiv
[8]

PruneVid: Visual token pruning for efficient video large language models

Xiaohu Huang, Hao Zhou, and Kai Han. PruneVid: Visual token pruning for efficient video large language models. InFindings of the Association for Computational Linguistics: ACL 2025, pp. 19959–19973,

work page 2025
[9]

GPT-4o System Card

Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. GPT-4o system card.arXiv preprint arXiv:2410.21276,

work page internal anchor Pith review Pith/arXiv arXiv
[10]

A survey on vision-language-action models for autonomous driving

Sicong Jiang, Zilin Huang, Kangan Qian, Ziang Luo, Tianze Zhu, Yang Zhong, Yihong Tang, Menglin Kong, Yunlong Wang, Siwen Jiao, et al. A survey on vision-language-action models for autonomous driving. arXiv preprint arXiv:2506.24044,

work page arXiv
[11]

OpenVLA: An Open-Source Vision-Language-Action Model

Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. OpenVLA: An open-source vision-language- action model.arXiv preprint arXiv:2406.09246,

work page internal anchor Pith review Pith/arXiv arXiv
[12]

LLaVA-OneVision: Easy Visual Task Transfer

Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. LlaVA-OneVision: Easy visual task transfer.arXiv preprint arXiv:2408.03326,

work page internal anchor Pith review Pith/arXiv arXiv
[13]

Improved baselines with visual instruction tuning

Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 26296–26306, 2024a. HaotianLiu, ChunyuanLi, YuhengLi, BoLi, YuanhanZhang, ShengShen, andYongJaeLee. LLaVA-NeXT: Improved reasoning, OCR, and world knowledge, January 2024b. URLhttps...

work page 2024
[14]

MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts

Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. MathVista: Evaluating mathematical reasoning of foundation models in visual contexts.arXiv preprint arXiv:2310.02255,

work page internal anchor Pith review Pith/arXiv arXiv
[15]

RWKV: Reinventing RNNs for the Transformer Era

Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Stella Biderman, Huanqi Cao, Xin Cheng, Michael Chung, Matteo Grella, et al. RWKV: Reinventing rnns for the transformer era.arXiv preprint arXiv:2305.13048,

work page internal anchor Pith review Pith/arXiv arXiv
[16]

VL-Mamba: Exploring state space models for multimodal learning.arXiv preprint arXiv:2403.13600,

21 Yanyuan Qiao, Zheng Yu, Longteng Guo, Sihan Chen, Zijia Zhao, Mingzhen Sun, Qi Wu, and Jing Liu. VL-Mamba: Exploring state space models for multimodal learning.arXiv preprint arXiv:2403.13600,

work page arXiv
[17]

LLaVA-Scissor: Token compression with semantic connected components for video LLMs.arXiv preprint arXiv:2506.21862,

Boyuan Sun, Jiaxing Zhao, Xihan Wei, and Qibin Hou. LLaVA-Scissor: Token compression with semantic connected components for video LLMs.arXiv preprint arXiv:2506.21862,

work page arXiv
[18]

Retentive Network: A Successor to Transformer for Large Language Models

Yutao Sun, Li Dong, Shaohan Huang, Shuming Ma, Yuqing Xia, Jilong Xue, Jianyong Wang, and Furu Wei. Retentive Network: A successor to transformer for large language models.arXiv preprint arXiv:2307.08621,

work page internal anchor Pith review Pith/arXiv arXiv
[19]

ECoFLaP: Efficient coarse-to-fine layer-wise pruning for vision-language models.arXiv preprint arXiv:2310.02998,

Yi-Lin Sung, Jaehong Yoon, and Mohit Bansal. ECoFLaP: Efficient coarse-to-fine layer-wise pruning for vision-language models.arXiv preprint arXiv:2310.02998,

work page arXiv
[20]

CoViPAL: Layer-wise contextualized visual token pruning for large vision-language models

Zicong Tang, Ziyang Ma, Suqing Wang, Zuchao Li, Lefei Zhang, Hai Zhao, Yun Li, and Qianren Wang. CoViPAL: Layer-wise contextualized visual token pruning for large vision-language models. InFindings of the Association for Computational Linguistics: EMNLP 2025, pp. 20701–20714,

work page 2025
[21]

Hongjie Wang, Chih-Yao Ma, Yen-Cheng Liu, Ji Hou, Tao Xu, Jialiang Wang, Felix Juefei-Xu, Yaqiao Luo, Peizhao Zhang, Tingbo Hou, Peter Vajda, Xiaoliang Dai, and Niraj K. Jha. LinGen-Uni: A universal linear-complexity framework for high-resolution minute-length text-to-video generation.Research Square Preprint, 2025a. Hongjie Wang, Chih-Yao Ma, Yen-Cheng L...

work page internal anchor Pith review Pith/arXiv arXiv
[22]

SparseVLM: Visual Token Sparsification for Efficient Vision-Language Model Inference

Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, et al. SparseVLM: Visual token sparsification for efficient vision-language model inference.arXiv preprint arXiv:2410.04417, 2024a. Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video in...

work page internal anchor Pith review Pith/arXiv arXiv
[23]

MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models

Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. MiniGPT-4: Enhancing vision- language understanding with advanced large language models.arXiv preprint arXiv:2304.10592,

work page internal anchor Pith review Pith/arXiv arXiv