ShangHai JiAi Genetics & IVF Institute

Multimodal large language models (MLLMs), which integrate language and visual cues for problem-solving, are crucial for advancing artificial general intelligence (AGI). However, current benchmarks for measuring the intelligence of MLLMs suffer from limited scale, narrow coverage, and unstructured knowledge, offering only static and undifferentiated evaluations. To bridge this gap, we introduce MDK12-Bench, a large-scale multidisciplinary benchmark built from real-world K–12 exams spanning six disciplines with 141K instances and 6,225 knowledge points organized in a six-layer taxonomy. Covering five question formats with difficulty and year annotations, it enables comprehensive evaluation to capture the extent to which MLLMs perform over four dimensions: 1) difficulty levels, 2) temporal (cross-year) shifts, 3) contextual shifts, and 4) knowledge-driven reasoning. We propose a novel dynamic evaluation framework that introduces unfamiliar visual, textual, and question form shifts to challenge model generalization while improving benchmark objectivity and longevity by mitigating data contamination. We further evaluate knowledge-point reference-augmented generation (KP-RAG) to examine the role of knowledge in reasoning. Key findings reveal limitations in current MLLMs in multiple aspects and provide guidance for enhancing model reasoning, robustness, and AI-assisted education.

Women with advanced maternal age (AMA) face a significant decline in embryo euploidy rates, leading to reduced fertility and increased risks of adverse pregnancy outcomes. While preimplantation genetic testing for aneuploidy (PGT-A) aids in selecting euploid embryos, it does not address the underlying causes of aneuploidy. Growth hormone (GH) has shown promise in improving oocyte quality and embryo development, but its impact on the euploidy status of embryos remains underexplored. This randomized controlled trial aims to evaluate the effect of GH supplementation on the euploidy rate of blastocysts in AMA patients. A total of 400 women with AMA undergoing PGT-A will be randomized into two groups: the GH group (GnRH antagonist protocol with daily subcutaneous injection of 2 IU GH starting from the previous menstrual cycle) and the control group (GnRH antagonist protocol without GH). Embryos will be cultured in time-lapse monitoring (TLM), and morphokinetic parameters will be recorded. Trophectoderm biopsies will be analyzed using next-generation sequencing-based PGT-A to determine ploidy status. The primary outcome is the euploidy rate of blastocysts, while secondary outcomes include embryo developmental and morphokinetic parameters, as well as pregnancy outcomes after transfer. This prospective, single-center RCT will investigate the impact of GH supplementation on euploidy rates, morphokinetic parameters, and transfer outcomes in AMA patients. By combining PGT-A and TLM with randomization and blinding, the study will minimize bias and provide robust insights into how GH may improve reproductive outcomes in AMA women, offering high-quality clinical evidence for ART practice. NCT05447208, www.clinicaltrials.gov. Registered on July 3rd, 2022.

Functional encryption (FE) which covers the notion of attribute-based encryption (ABE), is the cryptographic tool to realize fine-grained control on the accessibility of encrypted data. The traditional FE requires a central trusted authority to issue secret keys. It depends on the full-trust model, and is vulnerable to the security issue caused by key-escrow. While the registered FE (Reg-FE) achieves the zero-trust model and addresses the security issue by removing the use of central authority. It allows users to generate secret keys themselves and join the system by registering corresponding public keys to a curator. This work introduces delegated Reg-FE, which is a primitive with a new registration paradigm. It allows the registration of certain authorities that can issue secret keys for their respective classical FE sub-systems, beyond the prior work of registering plain users. Delegated Reg-FE implements a hybrid trust model within a two-level hierarchy. By redefining key escrow as a functional mechanism rather than a security concern, this model employs a zero-trust upper level which removes key-escrow, while the subsystem of each authority is locally full-trust and retains key-escrow mechanism. We construct four delegated Reg-FE schemes for functionalities that can be described as the $$2\times 2$$ combinations of linear function and policy check. Namely, Delegated Reg-IPFE, Delegated Reg-ABE, Reg-IPFE with delegated ABE, and Reg-ABE with delegated IPFE. All concrete schemes support bounded registrations and delegations, and achieve standard adaptive security under $$\textsc {MDDH}$$ assumption on prime-order bilinear group. Furthermore, these schemes only rely on black-box techniques. Technically, these schemes relies on dual-system techniques as prior registration-based works. And we devise a new “hierarchically invoked dual-system” technique on schemes which have sub-ABE delegation systems. Furthermore, we present a generic construction of Delegated Reg-FE from the combination of Reg-FE and FE. The instantiations of this generic construction demonstrate the feasibility of delegated Reg-FE, supporting arbitrary functions as well as unbounded numbers of registrations and delegations. However, this approach requires non-black-box techniques and achieves weaker semi-adaptive security without malicious registration, where the semi-adaptive means the adversary claims the challenge after seeing common reference string but before making any query. Its security relies solely on the underlying assumptions of the Reg-FE and FE components.

Achieving generalizability in solving out-of-distribution tasks is one of the ultimate goals of learning robotic manipulation. Recent progress of Vision-Language Models (VLMs) has shown that VLM-based task planners can alleviate the difficulty of solving novel tasks, by decomposing the compounded tasks as a plan of sequentially executing primitive-level skills that have been already mastered. It is also promising for robotic manipulation to adapt such composable generalization ability, in the form of composable generalization agents (CGAs). However, the community lacks of reliable design of primitive skills and a sufficient amount of primitive-level data annotations. Therefore, we propose RH20T-P, a primitive-level robotic manipulation dataset, which contains about 38k video clips covering 67 diverse manipulation tasks in real-world scenarios. Each clip is manually annotated according to a set of meticulously designed primitive skills that are common in robotic manipulation. Furthermore, we standardize a plan-execute CGA paradigm and implement an exemplar baseline called RA-P on our RH20T-P, whose positive performance on solving unseen tasks validates that the proposed dataset can offer composable generalization ability to robotic manipulation agents. Project homepage: https://sites.google.com/view/rh20t-primitive/main.

Although karyotyping technology has been implemented in genetic diagnosis for a long time, the comprehensive understanding of this technology is still limited. The aim of this study was to investigate the accuracy and resolution of karyotyping method in detecting chromosomal abnormalities by molecular genetic method. In this study, we conducted a retrospective analysis of embryonic molecular karyotypes and peripheral blood cytogenetic karyotypes from patients with balanced chromosomal rearrangements undergoing preimplantation genetic testing at our reproductive center. Blood karyotyping was performed using routine G-banding at the 400-band resolution by two well-trained technicians. The embryonic molecular karyotypes were detected using either high-throughput sequencing or single nucleotide polymorphism microarray method. We compared the breakpoint locations, determined by unbalanced rearrangements, in the embryonic molecular karyotypes with the corresponding rearranged chromosome bands in the peripheral blood karyotypes. A total of 508 cases were enrolled and 2078 embryos were detected, 404 cases were analyzed finally. We found that only 39.32% (289/735) of embryonic molecular breakpoints were located within the rearranged bands identified by peripheral blood karyotyping, while the remaining 60.68% (446/735) fell outside these regions. Our results showed only 73 cases exhibited accurate karyotyping results, indicating an accuracy rate of 18.07%. Furthermore, the average resolution of karyotyping technique was found to be approximately 9.01 megabases (Mb) pairs. These findings provide profound insight into the accuracy and resolution of karyotyping techniques, which can contribute to more precise genetic counseling.

Abstract Information scrambling refers to the phenomenon in which local quantum information in a many-body system becomes dispersed throughout the entire system under unitary evolution. It has been extensively studied in closed quantum systems, where it is quantified by operator size growth, revealing deep connections between condensed matter physics, high-energy physics, and quantum information. However, when extending the study of operator size dynamics to open quantum systems, two different definitions of operator size distributions emerge. These definitions are based on different treatments of the bath. In this work, we aim to establish a unified picture for operator size dynamics in open quantum systems, using the solvable Brownian SYK models at generic system size. In particular, we provide the conditions under which the signature of scrambling transition — discovered using one particular definition — appears in operator size dynamics under the other definition. Additionally, we extend previous studies by exploring finite-size effects that are not captured by the scramblon theory. Our results provide a refined understanding of operator size dynamics in open quantum systems.

Abstract Accurate cloud motion vector retrieval in multi‐layer clouds faces persistent challenges due to ambiguous height assignment. We developed a novel overlapped cloud motion vectors (OCMVs) retrieval method using Himawari‐8 observations. Multi‐layer cloud top heights (CTHs) were retrieved based on multi‐spectral radiance using neural networks and were assigned to upper ice and lower water cloud layers. Subsequently, CTHs from the two layers were used as respective tracers for deriving OCMVs based on the optical flow algorithm. Applied to Typhoon Mulan (2022), OCMVs showed strong vertical wind shear within the inner region, further depicting the kinematic structure of Typhoon Mulan. The vortex center of lower water OCMVs provided more valuable information on determining typhoon center than that of single‐layer CMVs. Additionally, the OCMVs demonstrated good consistency with dropsonde observations, exhibiting Root‐Mean‐Square‐Errors (RMSEs) of wind direction at ∼18.5°and wind speed at ∼5.2 m/s.

Modal tracking has evolved into a crucial research domain over the past decades. It synergizes with automatic modal parameter estimation (AMPE) to achieve automatic operational modal analysis (AOMA). Modal tracking aims to establish the dynamic correspondence between the currently identified mode and historical reference modes, thereby following the evolution of modal parameters over time. Established modal tracking methods involve two stages: reference mode determination and current mode linking. This paper reviews the developments of modal tracking from both methodological and application perspectives, emphasizing the pivotal challenges in existing results and suggestions for future research. Additionally, a detailed summary of the distinct strategies employed across two stages and the application of modal tracking in various types of structures is provided. Future research should focus on the adaptive updating of reference modes and linking thresholds while enhancing the capability to eliminate spurious modes and track rapidly changing physical modes. This review seeks to assist researchers in efficiently grasping the advancements in modal tracking and in developing more suitable and robust methods for civil engineering structures in the future.

Log-structured merge-trees (LSM-trees) are widely used to construct key-value stores. They periodically compact overlapping sorted runs to reduce the read amplification. Prior research on compaction policies has focused on the trade-off between write amplification (WA) and read amplification (RA). In this paper, we propose to treat the compaction operation in LSM-trees as a computational and I/O-bandwidth investment for improving the system's future query throughput, and thus rethink the compaction policy designs. A typical LSM-tree application handles a steady but moderate write stream and prioritizes resources for top-level flushes of small sorted runs to avoid data loss due to write stalls. The goal of the compaction policy, therefore, is to maintain an optimal number of sorted runs to maximize average query throughput. Because compaction and read operations compete for the CPU and I/O resources from the same pool, we must perform a joint optimization to determine the appropriate timing and aggressiveness of the compaction. We introduce a three-level model of an LSM-tree and propose EcoTune, an algorithm based on dynamic programming to find the optimal compaction policy according to workload characterizations. Our evaluation on RocksDB shows that EcoTune improves the average query throughput by 1.5x to 3x over the leveling policy and by up to 2.5x over the lazy-leveling policy on workloads with range/point query ratios.