AutoNeural: Co-Designing Vision-Language Models for NPU Inference

A recent study presents the Universal Weight Subspace Hypothesis, revealing that deep neural networks trained on various tasks converge to similar low-dimensional parametric subspaces. This research analyzed over 1,100 models, including Mistral-7B, Vision Transformers, and LLaMA-8B, demonstrating that these networks exploit shared spectral subspaces regardless of initialization or task.

A new framework called Fourier-Attentive Representation Learning (FARL) has been proposed to enhance few-shot generalization in Vision-Language Models (VLMs) by disentangling visual representations through Fourier analysis. This method utilizes a dual cross-attention mechanism to separately query structural and stylistic features of images, aiming to improve the adaptability of VLMs in various tasks.

A new method for movie trailer generation, named SSMP, has been proposed, which utilizes self-paced and self-corrective masked prediction to enhance the quality of trailers by employing bi-directional contextual modeling. This approach addresses the limitations of traditional selection-then-ranking methods that often lead to error propagation in trailer creation.

A new framework called COMET has been introduced for generating stable and controllable character motion in real-time, addressing challenges in computer animation related to fine-grained control and motion degradation over long sequences. This autoregressive model utilizes a Transformer-based conditional VAE to allow precise control over user-specified joints, enhancing tasks such as goal-reaching and in-betweening.

A new task called Cross-View Point Correspondence (CVPC) has been proposed to enhance spatial understanding in Vision-Language Models (VLMs). This initiative includes the introduction of CrossPoint-Bench, a benchmark designed to evaluate models based on human cognitive processes of perception, reasoning, and correspondence. Current state-of-the-art models, such as Gemini-2.5-Pro, show significant performance gaps compared to human accuracy, highlighting the need for improvement in point-level correspondence.

A new Transformer-based architecture called Small Building Model (SBM) has been introduced for layout synthesis in Building Information Modeling (BIM) scenes. This model addresses the challenge of tokenizing buildings by integrating diverse architectural features into sequences while maintaining their compositional structure, utilizing a sparse attribute-feature matrix to represent room properties.

A new method called Sliding-Window Merging (SWM) has been proposed to enhance the efficiency of large language models (LLMs) by compacting patch-redundant layers. This technique identifies and merges consecutive layers based on their functional similarity, thereby maintaining performance while simplifying model architecture. Extensive experiments indicate that SWM outperforms traditional pruning methods in zero-shot inference performance.

Autonomous Vehicles (AVs) are advancing rapidly, driven by improvements in intelligent perception and control systems, with a critical focus on reliable object detection in complex environments. Recent research highlights the integration of Vision-Language Models (VLMs) and Large Language Models (LLMs) as pivotal in overcoming existing challenges in multimodal perception and contextual reasoning.