Speculative decoding has emerged as a potential solution for speeding up inferences using large language models (LLMs). However, training-based speculative methods require the creation of draft models that typically lack generalizability, whereas training-free speculative frameworks offer only limited speed-up gains. SPECTRA, a novel speculative framework developed by researchers from JAIST, accelerates LLM inferences without additional training and surpasses existing speculative decoding methods across diverse functionalities, making it a practical tool for real-world LLM applications.

Kyoto, Japan -- Northern fur seals are long-distance migrants, moving seasonally between breeding colonies and wintering grounds. After leaving breeding colonies in autumn, they travel south for winter foraging, returning in early spring. This allows them to avoid sea ice and harsh conditions in northern waters while relocating to ice-free areas.

The southward migration of these seals has been relatively well-documented. One of their primary wintering areas is the Sea of Japan, where they can find abundant prey. They have also been reported to damage fishing gear and remove catch directly from nets, potentially competing with local fisheries.

The seals' northbound migration from wintering grounds back to breeding colonies in spring remains poorly understood, however. Factors such as tag loss, limited battery life, and the difficulty of capturing animals at sea pose particular challenges.

Many animals, including humans, can flexibly modulate their responsiveness to sounds according to different situations. This ability allows them to optimize the use of their limited brain resources by prioritizing the processing of critical information at any given moment. In line with this idea, a research team at Nagoya University in Japan has demonstrated that when female fruit flies are eager to mate, dopaminergic signals can influence their sensitivity to sounds, including courtship sounds from males, which are an essential source of information for species reproduction.

We have proposed a new therapeutic effect of CAR-T-cell therapy through a lymphoma case treated with bispecific antibody (BsAb). Memory-formed T cells had been developed in the body after CAR-T-cell therapy. Interestingly, bystander CAR-negative CD8⁺ T cells in the CAR-T-cell product successfully expanded in periphery and lymph node, and mostly propagated after treatment with BsAb, which resulted in complete remission. Therefore, such sequential approach using CAR-T/BsAb might have a potential to further enhance antitumor effects.

A group at Niigata University (Japan) has investigated the pathogenic mechanisms of nephrotic syndrome, one of the most prevalent kidney diseases, which presents severe proteinuria, and revealed that elevated activity of cdc42, a small GTPase, caused by anti-nephrin antibody is a critical initiation event leading to proteinuria. The major parts of nephrotic syndrome are reported to be induced by autoantibody against nephrin. Suppression of cdc42 activity could be a promising therapy for nephrotic syndrome.

What happens when things combine? This question lies at the heart of the Borell-Brascamp-Lieb inequality (BBL), a mathematical relation widely applied across many fields of mathematics, science and beyond. Now, researchers from the Okinawa Institute of Science and Technology (OIST), University of Tokyo and University of Florence have provided a new proof for this powerful inequality, taking an unconventional approach based on heat and diffusion equations.

Inspired by biology, researchers have achieved the highest performing underwater adhesive hydrogel technology to date through a data mining and machine learning approach.

Researchers from the National Institute of Physiological Sciences found that ‘denoising’, a method used to clean diffusion-weighted MRI (dMRI) data, alters the appearance of dMRI images and the estimated signal-to-noise ratio of the data, but had limited impacts on the ability to detect optic tract abnormalities in glaucoma.

Origami, or the art of paper folding, finds its roots in Japanese tradition, where it is widely practiced even today. Drawing inspiration from this traditional art form, an emerging field of theoretical computer science, known as Computational Origami, has taken shape. Building on this idea, Professor Ryuhei Uehara from Japan Advanced Institute of Science and Technology, Japan, is working towards building algorithms for finding effective solutions to complex puzzles and patterns.

At Japan Advanced Institute of Science and Technology (JAIST), Professor Toshiaki Taniike’s lab – Laboratory of Materials Informatics, is pioneering a new era of materials discovery by combining automation, custom-built instruments, and data-driven methods. Their cutting-edge approach is accelerating the search for entirely new reactions, paving the way for carbon-neutral technologies, energy innovation, and a more sustainable future. His efforts are focused on developing a smarter, faster, and more sustainable approach to materials design.