what's new in generative ai

Marianne Arriola, Volodymyr Kuleshov

this paper presents set diffusion, a method that combines autoregressive and diffusion models for faster and more flexible decoding of discrete tokens. ml engineers working with sequence generation can use this to improve the efficiency and control over their generative models.

Zhenghao Zhou, Yiyan Li, Tao Xu +3

this research proposes a physics-guided llm framework to generate customized power distribution system feeders. this can help engineers design more efficient and reliable power grids by automating the generation of complex system configurations.

Wujian Peng, Lingchen Meng, Yuxuan

this paper presents a unified multimodal autoregressive model that uses a shared context and visual tokenizer for unification. ai developers can use this approach to build more cohesive and capable multimodal models that can process and generate across different data types like text and images.

Ritesh Thawkar, Shravan Venkatraman, Omkar Thawakar +5

this paper proposes a self-evolving framework for unified multimodal understanding and generation using self-consistency rewards. this approach allows models to improve their ability to both comprehend and create multimodal content, leading to more robust and versatile ai systems.

Jen-Hao Cheng, Yipeng Wang, Hao Zhang +2

this paper presents flex4dhuman, a multi-view video diffusion model for flexible 4d human reconstruction. for practitioners in computer graphics, animation, or virtual try-on, this provides a powerful tool for generating realistic and dynamic 3d human models from video input.

Vu Tuan Truong, Long Bao Le

this paper reveals a new vulnerability, "toobad," demonstrating how diffusion models can be backdoored with imperceptible triggers and low poison rates. for ai security researchers and developers, this highlights critical security risks in generative models, emphasizing the need for robust defenses against malicious attacks.

Jingxuan Li, Yuxuan Liang, Yifan Li +4

this paper shows how generative ai can learn human taste preferences from recipe data to design foods that are delicious, nutritious, and sustainable. for practitioners, this means ai can be used to innovate in food product development, potentially reducing development cycles and improving health and environmental outcomes.

Alexander Soen, Hisham Husain, Valentin De Bortoli +1

this paper presents a method to accelerate speculative decoding for large language models using block verification. for practitioners, this means faster inference times for llms, which can significantly reduce computational costs and improve user experience in production environments.

Sophia Tang, Yuchen Zhu, Molei Tao +1

this paper presents a2d2, a method for fine-tuning any-length discrete diffusion models for adaptive decoding in sequence generation. for practitioners in nlp or sequence modeling, this offers a stable and flexible framework for generating sequences of varying lengths, improving the adaptability of diffusion models.

evaluates and improves LLMs for complex inpatient decision-making by moving beyond question-answering to real clinical workflows. practitioners need this because it addresses the gap between medical benchmarks and the messy, multi-step reasoning required in actual hospital settings.

Parikshit Bansal, Sujay Sanghavi

this paper introduces token time continuous diffusion (ttcd), a novel diffusion language model operating in continuous space. researchers in nlp can explore this method as an alternative to traditional autoregressive models for generating text, potentially offering new ways to control and synthesize language.

Qwen Team

this technical report introduces qwen-image-2.0, an omni-capable image generation foundation model that unifies high-fidelity generation and precise editing. for practitioners in computer vision and creative fields, this model offers advanced capabilities for generating and manipulating images within a single framework, addressing challenges like ultra-long text prompts.