Search

Large language models (LLMs) like GPTs, trained on vast datasets, have demonstrated impressive capabilities in language understanding, reasoning, and planning, achieving human-level performance in various tasks. Most studies focus on enhancing these models by training on ever-larger datasets to build more powerful foundation models. While training stronger models is important, enabling models to evolve during inference is equally crucial, a process we refer to as AI self-evolution. Unlike large-scale training, self-evolution may rely on limited data or interactions. Inspired by the columnar organization of the human cerebral cortex, we hypothesize that AI models could develop cognitive abilities and build internal representations through iterative interactions with their environment. To achieve this, models need long-term memory (LTM) to store and manage processed interaction data. LTM supports self-evolution by representing diverse experiences across environments and agents. In this report, we explore AI self-evolution and its potential to enhance models during inference. We examine LTM's role in lifelong learning, allowing models to evolve based on accumulated interactions. We outline the structure of LTM and the systems needed for effective data retention and representation. We also classify approaches for building personalized models with LTM data and show how these models achieve self-evolution through interaction. Using LTM, our multi-agent framework OMNE achieved first place on the GAIA benchmark, demonstrating LTM's potential for AI self-evolution. Finally, we present a roadmap for future research, emphasizing the importance of LTM for advancing AI technology and its practical applications., Comment: 56 pages, 13 figures

The clinical diagnosis of most mental disorders primarily relies on the conversations between psychiatrist and patient. The creation of such diagnostic conversation datasets is promising to boost the AI mental healthcare community. However, directly collecting the conversations in real diagnosis scenarios is near impossible due to stringent privacy and ethical considerations. To address this issue, we seek to synthesize diagnostic conversation by exploiting anonymous patient cases that are easier to access. Specifically, we design a neuro-symbolic multi-agent framework for synthesizing the diagnostic conversation of mental disorders with large language models. It takes patient case as input and is capable of generating multiple diverse conversations with one single patient case. The framework basically involves the interaction between a doctor agent and a patient agent, and achieves text generation under symbolic control via a dynamic diagnosis tree from a tool agent. By applying the proposed framework, we develop the largest Chinese mental disorders diagnosis dataset MDD-5k, which is built upon 1000 cleaned real patient cases by cooperating with a pioneering psychiatric hospital, and contains 5000 high-quality long conversations with diagnosis results as labels. To the best of our knowledge, it's also the first labelled Chinese mental disorders diagnosis dataset. Human evaluation demonstrates the proposed MDD-5k dataset successfully simulates human-like diagnostic process of mental disorders. The dataset and code will become publicly accessible in https://github.com/lemonsis/MDD-5k.

Interactions between particles normally induce the decay of the particles Bloch oscillations (BOs) in a periodic lattice. In the limit of strong on-site interactions, spin-$1/2$ fermions may form doublon bound states and undergo BOs in the presence of a tilted potential. Here we investigate the impact of nearest-neighbor interaction $V$ on the multi-doublon BOs in a mass-imbalanced extended Fermi-Hubbard model. We derive an effective Hamiltonian for doublons, and show that a slight change in $V$ can qualitatively alter their dynamic behaviors. Notably, at a resonance point, the doublons behave like free hard-core bosons. Under a tilted potential, the system may exhibit different types of multi-doublon BOs at or deviation from the resonance point. Numerical results are presented to demonstrate our conclusions in both one- and two-dimensional systems., Comment: 12 pages, 7 figures

IR drop on the power delivery network (PDN) is closely related to PDN's configuration and cell current consumption. As the integrated circuit (IC) design is growing larger, dynamic IR drop simulation becomes computationally unaffordable and machine learning based IR drop prediction has been explored as a promising solution. Although CNN-based methods have been adapted to IR drop prediction task in several works, the shortcomings of overlooking PDN configuration is non-negligible. In this paper, we consider not only how to properly represent cell-PDN relation, but also how to model IR drop following its physical nature in the feature aggregation procedure. Thus, we propose a novel graph structure, PDNGraph, to unify the representations of the PDN structure and the fine-grained cell-PDN relation. We further propose a dual-branch heterogeneous network, PDNNet, incorporating two parallel GNN-CNN branches to favorably capture the above features during the learning process. Several key designs are presented to make the dynamic IR drop prediction highly effective and interpretable. We are the first work to apply graph structure to deep-learning based dynamic IR drop prediction method. Experiments show that PDNNet outperforms the state-of-the-art CNN-based methods by up to 39.3% reduction in prediction error and achieves 545x speedup compared to the commercial tool, which demonstrates the superiority of our method.

Video Moment Retrieval (VMR) aims at retrieving the most relevant events from an untrimmed video with natural language queries. Existing VMR methods suffer from two defects: (1) massive expensive temporal annotations are required to obtain satisfying performance; (2) complicated cross-modal interaction modules are deployed, which lead to high computational cost and low efficiency for the retrieval process. To address these issues, we propose a novel method termed Cheaper and Faster Moment Retrieval (CFMR), which well balances the retrieval accuracy, efficiency, and annotation cost for VMR. Specifically, our proposed CFMR method learns from point-level supervision where each annotation is a single frame randomly located within the target moment. It is 6 times cheaper than the conventional annotations of event boundaries. Furthermore, we also design a concept-based multimodal alignment mechanism to bypass the usage of cross-modal interaction modules during the inference process, remarkably improving retrieval efficiency. The experimental results on three widely used VMR benchmarks demonstrate the proposed CFMR method establishes new state-of-the-art with point-level supervision. Moreover, it significantly accelerates the retrieval speed with more than 100 times FLOPs compared to existing approaches with point-level supervision., Comment: 10 pages, 7 figures

It is becoming increasingly popular to elicit informative priors on the basis of historical data. Popular existing priors, including the power prior, commensurate prior, and robust meta-analytic prior provide blanket discounting. Thus, if only a subset of participants in the historical data are exchangeable with the current data, these priors may not be appropriate. In order to combat this issue, propensity score (PS) approaches have been proposed. However, PS approaches are only concerned with the covariate distribution, whereas exchangeability is typically assessed with parameters pertaining to the outcome. In this paper, we introduce the latent exchangeability prior (LEAP), where observations in the historical data are classified into exchangeable and non-exchangeable groups. The LEAP discounts the historical data by identifying the most relevant subjects from the historical data. We compare our proposed approach against alternative approaches in simulations and present a case study using our proposed prior to augment a control arm in a phase 3 clinical trial in plaque psoriasis with an unbalanced randomization scheme.

We establish stability and characteristics of two-dimensional (2D) vortex ring-shaped quantum droplets (QDs) formed by binary Bose-Einstein condensates (BECs). The system is modeled by the Gross-Pitaevskii (GP) equation with the cubic term multiplied by a logarithmic factor (as produced by the Lee-Huang-Yang correction to the mean-field theory) and a potential which is a periodic function of the radial coordinate. Narrow vortex rings with high values of the topological charge, trapped in particular circular troughs of the radial potential, are produced. These results suggest an experimentally relevant method for the creation of vortical QDs (thus far, only zero-vorticity ones have been reported). The 2D GP equation for the narrow rings is approximately reduced to the 1D form, which makes it possible to study the modulational stability of the rings against azimuthal perturbations. Full stability areas are delineated for these modes. The trapping capacity of the circular troughs is identified for the vortex rings with different winding numbers (WNs). Stable compound states in the form of mutually nested concentric multiple rings are constructed too, including ones with opposite signs of the WNs. Other robust compound states combine a modulationally stable narrow ring in one circular potential trough and an azimuthal soliton performing orbital motion in an adjacent one. The results may be used to design a device employing coexisting ring-shaped modes with different WNs for data storage., Comment: 15 pages, 11 figures,to be published in New Journal of Physics

Dielectric loss is one of the major decoherence sources of superconducting qubits. Contemporary high-coherence superconducting qubits are formed by material systems mostly consisting of superconducting films on substrate with low dielectric loss, where the loss mainly originates from the surfaces and interfaces. Among the multiple candidates for material systems, a combination of titanium nitride (TiN) film and sapphire substrate has good potential because of its chemical stability against oxidization, and high quality at interfaces. In this work, we report a TiN film deposited onto sapphire substrate achieving low dielectric loss at the material interface. Through the systematic characterizations of a series of transmon qubits fabricated with identical batches of TiN base layers, but different geometries of qubit shunting capacitors with various participation ratios of the material interface, we quantitatively extract the loss tangent value at the substrate-metal interface smaller than $8.9 \times 10^{-4}$ in 1-nm disordered layer. By optimizing the interface participation ratio of the transmon qubit, we reproducibly achieve qubit lifetimes of up to 300 $\mu$s and quality factors approaching 8 million. We demonstrate that TiN film on sapphire substrate is an ideal material system for high-coherence superconducting qubits. Our analyses further suggest that the interface dielectric loss around the Josephson junction part of the circuit could be the dominant limitation of lifetimes for state-of-the-art transmon qubits.

Superconducting qubits provide a promising path toward building large-scale quantum computers. The simple and robust transmon qubit has been the leading platform, achieving multiple milestones. However, fault-tolerant quantum computing calls for qubit operations at error rates significantly lower than those exhibited in the state of the art. Consequently, alternative superconducting qubits with better error protection have attracted increasing interest. Among them, fluxonium is a particularly promising candidate, featuring large anharmonicity and long coherence times. Here, we engineer a fluxonium-based quantum processor that integrates high qubit-coherence, fast frequency-tunability, and individual-qubit addressability for reset, readout, and gates. With simple and fast gate schemes, we achieve an average single-qubit gate fidelity of 99.97% and a two-qubit gate fidelity of up to 99.72%. This performance is comparable to the highest values reported in the literature of superconducting circuits. Thus our work, for the first time within the realm of superconducting qubits, reveals an approach toward fault-tolerant quantum computing that is alternative and competitive to the transmon system.

Electronic systems for qubit control and measurement serve as a bridge between quantum programming language and quantum information processors. With the rapid development of superconducting quantum circuit (SQC) technology, synchronization in a large-scale system, low-latency execution, and low noise are required for electronic systems. Here, we present a field-programmable gate array (FPGA)-based electronic system with a distributed synchronous clock and trigger architecture. The system supports synchronous control of qubits with jitters of approximately 5 ps. We implement a real-time digital signal processing system in the FPGA, enabling precise timing control, arbitrary waveform generation, IQ demodulation for qubit state discrimination, and the generation of real-time qubit-state-dependent trigger signals for feedback/feedforward control. The hardware and firmware low-latency design reduces the feedback/feedforward latency of the electronic system to 125 ns, significantly less than the decoherence times of the qubit. Finally, we demonstrate the functionalities and low-noise performance of this system using a fluxonium quantum processor.

Determining the extent to which a patient is benefiting from cancer therapy is challenging. Criteria for quantifying the extent of "tumor response" observed within a few cycles of treatment have been established for various types of solid as well as hematologic malignancies. These measures comprise the primary endpoints of phase II trials. Regulatory approvals of new cancer therapies, however, are usually contingent upon the demonstration of superior overall survival with randomized evidence acquired with a phase III trial comparing the novel therapy to an appropriate standard of care treatment. With nearly two thirds of phase III oncology trials failing to achieve statistically significant results, researchers continue to refine and propose new surrogate endpoints. This article presents a Bayesian framework for studying relationships among treatment, patient subgroups, tumor response and survival. Combining classical components of mediation analysis with Bayesian model averaging (BMA), the methodology is robust to model mis-specification among various possible relationships among the observable entities. Posterior inference is demonstrated via application to a randomized controlled phase III trial in metastatic colorectal cancer. Moreover, the article details posterior predictive distributions of survival and statistical metrics for quantifying the extent of direct and indirect, or tumor response mediated, treatment effects., Comment: 25 pages, 3 figures and 3 tables in the main manuscript. Supplementary materials included

Basket designs are prospective clinical trials that are devised with the hypothesis that the presence of selected molecular features determine a patient's subsequent response to a particular "targeted" treatment strategy. Basket trials are designed to enroll multiple clinical subpopulations to which it is assumed that the therapy in question offers beneficial efficacy in the presence of the targeted molecular profile. The treatment, however, may not offer acceptable efficacy to all subpopulations enrolled. Moreover, for rare disease settings, such as oncology wherein these trials have become popular, marginal measures of statistical evidence are difficult to interpret for sparsely enrolled subpopulations. Consequently, basket trials pose challenges to the traditional paradigm for trial design, which assumes inter-patient exchangeability. The R-package \pkg{basket} facilitates the analysis of basket trials by implementing multi-source exchangeability models. By evaluating all possible pairwise exchangeability relationships, this hierarchical modeling framework facilitates Bayesian posterior shrinkage among a collection of discrete and pre-specified subpopulations. Analysis functions are provided to implement posterior inference of the response rates and all possible exchangeability relationships between subpopulations. In addition, the package can identify "poolable" subsets of and report their response characteristics. The functionality of the package is demonstrated using data from an oncology study with subpopulations defined by tumor histology., Comment: 18 pages, 4 figures, 3 tables, submitted to the Journal of Open Source Software