In high-precision scenarios, vision language models suffer from Linguistic Priors Hallucination.When processing familiar text, models tend to over-rely on internal parametric knowledge, effectively "reciting" the content rather than "reading" the image.In this paper, we first systematically investigate this phenomenon by constructing the GlitchText Probing Dataset.We discover that the model's reliance on visual grounding diminishes significantly as the generation length increases.To mitigate this, we propose PAR (Positional Perturbation and Attention Recycling), a training-free, inferencetime intervention framework.PAR consists of two parts: (1) Positional Perturbation (PP) injects structured phase noise into the rotary positional embeddings; (2) Foveal Attention Recycling (FAR) detects over-confident linguistic priors and dynamically redistributes attention mass back to important visual regions.Extensive experiments across state-of-the-art models, demonstrate that PAR significantly reduces hallucination rates (reducing CER by 12%), particularly in long-context scenarios, while maintaining robust generalization on standard benchmarks.Our code is publicly available at https://github.com/Zoeyyao27/PAR-for- Faithful-OCR.

Singing, as a common facial movement second only to talking, can be regarded as a universal language across ethnicities and cultures, plays an important role in emotional communication, art, and entertainment. However, it is often overlooked in the field of audio-driven 3D facial animation due to the lack of singing head datasets and the domain gap between singing and talking in rhythm and amplitude. To this end, we collect a large-scale high-quality multi-modal singing head dataset,SingingHead, which consists of more than 27 hours of synchronized singing video, 3D facial motion, singing audio, and background music from 76 individuals and 8 types of music. Along with the SingingHead dataset, we benchmark existing audio-driven 3D facial animation methods and 2D talking head methods on the singing task. Existing 3D facial animation methods and 2D talking head methods fail to produce satisfactory singing results. Focusing on the 3D singing head animation, we first utilize the proposed singing-specific dataset to retrain the 3D facial animation methods, resulting in substantial performance improvements. Besides, considering the absence of background music and the slow generation speed of existing methods, we propose a simple but efficient non-autoregressive VAE-based framework with background music as an input signal to generate diverse and accurate 3D singing facial motions in real time. Extensive experiments demonstrate the significance of the SingingHead dataset in promoting the development of singing head animation. The dataset is released for research purposes at:https://wsj-sjtu.github.io/SingingHead/.

Clothed avatar generation has wide applications in virtual and augmented reality, filmmaking, and more. Previous methods have achieved success in generating diverse digital avatars, however, generating avatars with disentangled components (\eg, body, hair, and clothes) has long been a challenge. In this paper, we propose LayerAvatar, the first feed-forward diffusion-based method for generating component-disentangled clothed avatars. To achieve this, we first propose a layered UV feature plane representation, where components are distributed in different layers of the Gaussian-based UV feature plane with corresponding semantic labels. This representation supports high-resolution and real-time rendering, as well as expressive animation including controllable gestures and facial expressions. Based on the well-designed representation, we train a single-stage diffusion model and introduce constrain terms to address the severe occlusion problem of the innermost human body layer. Extensive experiments demonstrate the impressive performances of our method in generating disentangled clothed avatars, and we further explore its applications in component transfer. The project page is available at: https://olivia23333.github.io/LayerAvatar/

Characterizing quantum states is essential for advancing many quantum technologies. Recently, deep neural networks have been applied to learn quantum states by generating compressed implicit representations. Despite their success in predicting properties of the states, these representations remain a black box, lacking insights into strategies for experimental reconstruction. In this work, we aim to open this black box by developing explicit representations through generating surrogate state preparation circuits for property estimation. We design a reinforcement learning agent equipped with a Transformer-based architecture and a local fidelity reward function. Relying solely on measurement data from a few neighboring qubits, our agent accurately recovers properties of target states. We also theoretically analyze the global fidelity the agent can achieve when it learns a good local approximation. Extensive experiments demonstrate the effectiveness of our framework in learning various states of up to 100 qubits, including those generated by shallow Instantaneous Quantum Polynomial circuits, evolved by Ising Hamiltonians, and many-body ground states. Furthermore, the learned circuit representations can be applied to Hamiltonian learning as a downstream task utilizing a simple linear model.

4D printed constructs (4DPC) can transform into different forms and shapes upon exposure to environmental stimuli. However, most 4DPC can only achieve simple deformation under specific incentives currently. There is no effective strategy to enable the 4DPC to adapt to environmental changes autonomously. In this paper, an adaptive fuzzy controller based on expert knowledge is proposed, which will compensate for the disturbance and uncertainty caused by external stimuli to ensure the adaptability and robustness of the 4DPC in dynamic change and uncertainty. To facilitate the design of the controller, a long short-term memory (LSTM) neural network model is proposed to characterize the nonlinear deformation characteristics of the 4DPC. Finally, the environmental adaptability experiment of a self-made 4D printed color-changing structure(4DPCCS) shows that the 4DPCCS integrating perception and control functions can actively adapt to the dynamic changes of the environment. The deformation error of the regulation is$\le \pm 2.39\%$, and the 4DPCCS can adjust and maintain it adaptively according to changes in the external environment, achieving a camouflage function similar to that of a natural chameleon. Note to Practitioners—The aim of this study is to provide a new solution to meet the dynamic and real-time requirements of real environments for adaptive deformation of 4D printed constructions in dynamic environments. Most of the existing control methods act on the printing process of the constructions to guide the structural design of the 4DPC to achieve specific deformation to specific stimulus. This paper proposes a new method to realize the adaptive modulation of 4D printed constructions to external stimuli through the introduction of adaptive fuzzy control. First, we developed and designed a 4D printed color-changing structure with actuation, sensor and color-changing functions. Secondly, based on a large number of experimental data, we use the modeling method of the LSTM neural network to characterize the relationship between the complex deformation of 4D printed constructions and external stimuli. Finally, an adaptive fuzzy control system for 4D printed color-changing structure is developed and designed based on the neural network model. The experimental results show that this method is feasible. Under the action of the proposed control strategy, the constructions can actively identify the changes of the external environment and adjust and maintain adaptively, realizing the function of color-changing camouflage. In future studies, we will further research the autonomous adjustment method for variable performance or variable functionality of 4D printed constructions.

Despite the flourishing of 3D object generation, generating high-quality digital avatars with detailed geometry and texture that are free to animate remains a challenging task. Existing avatar generation techniques often suffer from limitations such as low-quality geometry and blurry texture. Thus, we propose HQ-Avatar, a novel method for generating animatable avatars with high-quality geometry and texture. We enhance the geometry quality by proposing an importance sampling strategy and capturing intricate details through learned normal maps. To achieve high-quality texture, we present a neural point-based avatar representation, which enables high-resolution rendering results at a resolution of 10242, allowing detailed supervision. Extensive experiments on THuman2.0 dataset demonstrate the superiority of our method over state-of-the-art techniques in generating high-quality avatars. Furthermore, we show the applicability of our method by employing it as 3D priors to simplify the human avatar reconstruction process from scans or even single images. Code is available at https://github.com/olivia23333/HQ-Avatar

Data-to-text (D2T) generation, the task of converting structured data into natural language, has extensive real-world applications. While supervised models have achieved promising results, they rely heavily on costly labeled training data. This paper investigates unsupervised D2T generation by leveraging the impressive general abilities of large language models (LLMs). We propose a framework for LLMs to collaboratively learn from unlabeled data through cascaded multi-model quality control. Specifically, one LLM, acting as a writer, generates candidate texts from input data. Additional LLMs, serving as checkers, validate output quality to filter high-quality samples for training the writer LLM. By cascading generation, checking, and meta-checking, the models extract linguistic knowledge and grounding ability from abundant unlabeled data. Experiments on established benchmarks demonstrate enhanced fluency, accuracy, and coherence compared to supervised baselines. This unsupervised approach circumvents labeled data dependence, unlocking readily available LLMs for on-demand D2T generation across diverse applications.