Multimodal learning faces challenges in effectively fusing information from diverse modalities, especially when modality quality varies across samples. Dynamic fusion strategies, such as attention mechanism in Transformers, aim to address such challenge by adaptively emphasizing modalities based on the characteristics of input data. However, through amounts of carefully designed experiments, we surprisingly observed that the dynamic adaptability of widely-used self-attention models diminishes. Model tends to prefer one modality regardless of data characteristics. This bias triggers a self-reinforcing cycle that progressively overemphasizes the favored modality, widening the distribution gap in attention keys across modalities and deactivating attention mechanism's dynamic properties. To revive adaptability, we propose a simple yet effective method Rolling Query (RollingQ), which balances attention allocation by rotating the query to break the self-reinforcing cycle and mitigate the key distribution gap. Extensive experiments on various multimodal scenarios validate the effectiveness of RollingQ and the restoration of cooperation dynamics is pivotal for enhancing the broader capabilities of widely deployed multimodal Transformers. The source code is available at https://github.com/GeWu-Lab/RollingQ_ICML2025.

Partitioning a polygonal mesh into meaningful parts can be challenging. Many applications require decomposing such structures for further processing in computer graphics. In the last decade, several methods were proposed to tackle this problem, at the cost of intensive computational times. Recently, machine learning has proven to be effective for the segmentation task on 3D structures. Nevertheless, these state-of-the-art methods are often hardly generalizable and require dividing the learned model into several specific classes of objects to avoid overfitting. We present a data-driven approach leveraging deep learning to encode a mapping function prior to mesh segmentation for multiple applications. Our network reproduces a neighborhood map using our knowledge of the \textsl{Shape Diameter Function} (SDF) method using similarities among vertex neighborhoods. Our approach is resolution-agnostic as we downsample the input meshes and query the full-resolution structure solely for neighborhood contributions. Using our predicted SDF values, we can inject the resulting structure into a graph-cut algorithm to generate an efficient and robust mesh segmentation while considerably reducing the required computation times.

This article was co-published with Understanding AI, a newsletter that explores how A.I. works and how it's changing our world. In 2011, venture capitalist Marc Andreessen published an essay that became a kind of manifesto for Silicon Valley during the 2010s. "Software is eating the world," Andreessen declared. Computers and the internet had already revolutionized a bunch of information-oriented businesses: books, movies, music, photography, telecommunications, and so forth. Software also played a major supporting role in more tangible industries. New cars had dozens of computer chips in them, for example, and the oil and gas industry made heavy use of software to discover new drilling sites. But Andreessen, co-founder of the venture capital firm Andreessen Horowitz, argued that the software revolution was only getting started.

If AI could let you speak with your deceased loved ones again, would you take the chance? A few companies and experts believe this will soon be possible, and some are even starting to market their own solutions. Nothing supernatural about their proposal; what they are offering is, rather, the ability to talk to a digital representation of the dead, fine-tuned by combining large language models like GPT-4, speech synthesis and AI generation tools. In China, as the Strait Times reports, some funeral companies are rapidly bringing the worship of the deceased into the digital age, allowing relatives to speak to a digital avatar of the dearly departed. Many have started offering this service in early April, around the time of the Qing Ming Festival (or Tomb Sweeping Day), a public holiday when people remember and honor the dead. One of them, funeral services provider Shanghai Fushouyun, started even earlier, conducting its first AI-assisted funeral in January 2022, when colleagues and students of a deceased Chinese surgeon had the opportunity to chat with his digital replica on a screen, for a final farewell.

Since the birth of commercial air travel, airlines have always been defined by the experiences they create for passengers. Those experiences haven't always looked and felt the same, of course. From the golden era of the 1950s–60s (where flying was for the well-off and the well-dressed) to today's dominance of low-cost carriers, the industry has been in a state of almost constant change. Amidst all this change, however, one thing has remained the same: the role of marketing. Whether luxury, comfort or compromise, marketing has traditionally been there to sell experiences.

I haven't been following suit the fate of DMOZ, the human-edited web catalogue. So I was a bit surprised that it closed down a few days ago. Not that I was much surprised. I always felt that there was something wrong with it. I recall the days when a DMOZ listing was still supposed to improve your Google Ranking. We had our student intern apply for a job as editor for the vacant and oddly maintained insurance pages in German language.