Ruixin Fan, Weining Wang, Jiahui Sun, Tailai Chen, Jing Liu
Unified multimodal models (UMMs) jointly support visual understanding and image generation, yet information flow between these two capabilities remains largely asymmetric: understanding representations routinely condition generation, whereas generative representations are rarely reused for multimodal reasoning. In this work, we investigate whether intermediate representations from a UMM's native diffusion branch can enhance spatial reasoning without compromising general multimodal understanding. We propose DiFuse, which selectively incorporates intermediate diffusion representations into the understanding pathway of UMMs. Through systematic layer-wise analysis, we identify which diffusion representations are most informative and where they can be integrated with minimal interference to semantic processing. Guided by these findings, DiFuse spatially aligns diffusion features with visual tokens and injects them into selected multimodal decoder layers via residual fusion. Extensive experiments on UniWorld-V1 show that, DiFuse consistently improves spatial reasoning over a matched SFT baseline while maintaining broadly comparable performance on general multimodal benchmarks. Further analyses show that these gains depend critically on the choice of diffusion representations, injection positions, and the preservation of spatial correspondence between diffusion features and visual tokens. Results on OmniGen2 further demonstrate the transferability of DiFuse across unified multimodal architectures. Overall, our findings show that native diffusion representations can provide complementary spatial cues for multimodal reasoning when selectively integrated in a structure-preserving manner.
Tailai Chen, Yuluxin Fu, Jie Deng, Shu Luo, Yanan Wu, Danfeng Hong
Reconstructing 1-m imagery from Sentinel-2 combines two difficulties. The input and the reference are observed by different satellites, which differ in spatial resolution, spectral response, radiometry, acquisition time, and local geometry, so no synthetic degradation operator links them. The required magnification of ten in each spatial dimension also far exceeds the factors of two to four at which remote sensing super-resolution is usually developed. We study this real-world $\times 10$ problem using YREB-SR, which pairs 2022 Sentinel-2 RGB composites with 2022 Jilin-1 RGB references resampled to a 1-m grid over the Yangtze River Economic Belt. The benchmark contains 11,559 geographically matched scene pairs, with 6,409, 2,135, and 3,015 assigned to training, validation, and testing, respectively. We develop CrossVAR, a visual autoregressive model that predicts the 1-m latent as a nine-scale hierarchy conditioned on the Sentinel-2 input and estimates the remaining quantization residual with a diffusion refiner run without injected noise. Complete 1600 $\times$ 1600 scenes are reconstructed in overlapping windows that start on the Sentinel-2 pixel grid, as every training patch does, and every window is decoded at the size on which the decoder was trained before the windows are blended with Gaussian weights. On all 3,015 test scenes, CrossVAR achieves 27.60 dB PSNR, 0.7725 SSIM, 0.1451 LPIPS, and 7.15 FID, the best value of each metric among the compared methods. These results establish a strong autoregressive baseline for large-scale cross-sensor reconstruction. To facilitate reproducibility, the source code and implementation are publicly available at https://github.com/SHANOAL/RSSR.
Prefilling computational costs pose a significant bottleneck for Large Language Models (LLMs) and Large Multimodal Models (LMMs) in long-context settings. While token pruning reduces sequence length, prior methods rely on heuristics that break compatibility with hardware-efficient kernels like FlashAttention. In this work, we observe that tokens evolve toward semantic fixing points, making further processing redundant. To this end, we introduce Delta Attention Selective Halting (DASH), a training-free policy that monitors the layer-wise update dynamics of the self-attention mechanism to selectively halt stabilized tokens. Extensive evaluation confirms that DASH generalizes across language and vision benchmarks, delivering significant prefill speedups while preserving model accuracy and hardware efficiency. Code will be released at https://github.com/verach3n/DASH.git .
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