Lei Tian, Chunzheng Zhu, Kai Wu, Yifan Zhang, Haihua Yang
Native multimodal large language models (native MLLMs) jointly encode visual and linguistic sequences within a unified autoregressive Transformer, emerging as a promising paradigm for multimodal understanding. Unlike conventional ViT-based MLLMs relying on external vision encoders, ViT-free native MLLMs tokenize pixel patches directly, potentially suffering severe visual token redundancy. A pilot study confirms this: unlike ViT-based models that degrade under compression, ViT-free MLLMs tolerate aggressive compression with near-lossless or even improved performance. Yet existing compression methods target ViT-based settings, leaving native MLLMs, which stand to gain the most, unexplored. To close this gap, we introduce an analysis pipeline centered on token discriminability and cross-dimensional heterogeneous redundancy, revealing that (i) redundancy is distributed independently and heterogeneously along height and width axes, and (ii) early-layer visual tokens are indiscriminable and uniformly redundant. Guided by these findings, we propose PACE (Pareto-Adaptive Compression for Efficient Native MLLMs), the first visual token compression framework tailored for native MLLMs with Multimodal-RoPE. Leveraging dimensional decoupling of positional encoding, PACE decomposes attention into two orthogonal spatial saliency estimates along height and width, gated by a positional attention signal, yielding direction-aware importance with negligible overhead. It further casts token budget allocation as Pareto optimization over performance preservation, inference efficiency, and spatial coverage, jointly identifying optimal compression depths and retention ratios. Experiments show PACE maintains near-lossless performance at only 25%-50% visual token retention, even improving several benchmarks, while delivering over 2x FLOPs reduction.
Yupeng Xiao, Yankai Wu, Jianjie Luo, Jiajie Lin, Yi Yu, Zhenguo Yang
Recent diffusion models have achieved impressive progress in single image to multi-view generation by synthesizing consistent novel views efficiently. Existing methods usually explicitly or implicitly couple geometry inference and texture synthesis, making generation vulnerable to inaccurate geometry prediction. Although several geometry-guided methods have explored decoupling, they usually treat predicted geometry as a fixed condition and do not explicitly model whether the geometry is reliable. We therefore propose Reliability-Guided Multi-View Generation (RGMV), which comprises a parameter-free Self-Consistency Geometry Gate (SCG) and a Reliability-aware Geometry Injection (RGI) module. Specifically, a geometry proposal generator produces multiple geometry predictions for the same reference image and target viewpoints under different stochastic sampling configurations. SCG measures their spatial agreement to derive a reliability map, which RGI uses to modulate the injection strength of geometry-conditioned residuals in the RGB synthesis model. Moreover, RGMV can adapt to different geometry proposal generators. As a result, RGMV preserves stronger geometric guidance in stable regions while attenuating unreliable geometry cues before they propagate into multi-view RGB generation. Extensive experiments show that reliability-aware geometry injection improves the geometric quality of multi-view generation. Code: https://github.com/WISLab-GDUT/RGMV.
Qian Zhang, Xinye Li, Xiaokai Wu, Junhao Xu, Zhanyue Qin Harbin Institute of Technology (Shenzhen) and Tencent, Qingbin Liu, Junxian Cai Tencent, Xi Chen 等 13 位作者
Model Editing, also known as knowledge editing, is receiving increasing attention in the field of Large Language Models (LLMs). However, existing model editing approaches predominantly focus on knowledge-level or static visual domains, overlooking dynamic semantics. This paper exploratively applies six representative model editing methods (FT, IKE, MEND, SERAC, MEMIT and AlphaEdit) to Video Large Language Models (Vid-LLMs) and introduces the first benchmark specifically designed for Vid-LLMs editing—VMEB (Vid-LLMs Model Editing Benchmark)—systematically extending model editing research from static modalities to dynamic video scenarios. We position this work as a forward-looking benchmark and a foundational diagnostic study: in the video paradigm, our evaluation dimensions encompass traditional metrics including Reliability, Locality, and Generality, while also introducing a video-specific metric: Robustness. Based on experimental results, we analyze the strengths and limitations of existing model editing approaches, and identify new challenges and research directions for the future development of the model editing field within the context of multimodal and video paradigms. Our benchmark is available at https://github.com/Sakabamrisa/VMEB.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表