Search for: All records

As Large Language Models (LLMs) are increasingly deployed to handle various natural language processing (NLP) tasks, concerns regarding the potential negative societal impacts of LLM-generated content have also arisen. To evaluate the biases exhibited by LLMs, researchers have recently proposed a variety of datasets. However, existing bias evaluation efforts often focus on only a particular type of bias and employ inconsistent evaluation metrics, leading to difficulties in comparison across different datasets and LLMs. To address these limitations, we collect a variety of datasets designed for the bias evaluation of LLMs, and further propose CEB, a Compositional Evaluation Bechmark that covers different types of bias across different social groups and tasks. The curation of CEB is based on our newly proposed compositional taxonomy, which characterizes each dataset from three dimensions: bias types, social groups, and tasks. By combining the three dimensions, we develop a comprehensive evaluation strategy for the bias in LLMs. Our experiments demonstrate that the levels of bias vary across these dimensions, thereby providing guidance for the development of specific bias mitigation methods. 

Text watermarks for large language models (LLMs) have been commonly used to identify the origins of machine-generated content, which is promising for assessing liability when combating deepfake or harmful content. While existing watermarking techniques typically prioritize robustness against removal attacks, unfortunately, they are vulnerable to spoofing attacks: malicious actors can subtly alter the meanings of LLM-generated responses or even forge harmful content, potentially misattributing blame to the LLM developer. To overcome this, we introduce a bi-level signature scheme, Bileve, which embeds fine-grained signature bits for integrity checks (mitigating spoofing attacks) as well as a coarse-grained signal to trace text sources when the signature is invalid (enhancing detectability) via a novel rank-based sampling strategy. Compared to conventional watermark detectors that only output binary results, Bileve can differentiate 5 scenarios during detection, reliably tracing text provenance and regulating LLMs. The experiments conducted on OPT-1.3B and LLaMA-7B demonstrate the effectiveness of Bileve in defeating spoofing attacks with enhanced detectability. 

Deep neural network (DNN) models, despite their impressive performance, are vulnerable to exploitation by attackers who attempt to transfer them to other tasks for their own benefit. Current defense strategies mainly address this vulnerability at the model parameter level, leaving the potential of architectural-level defense largely unexplored. This paper, for the first time, addresses the issue of model protection by reducing transferability at the architecture level. Specifically, we present a novel neural architecture search (NAS)-enabled algorithm that employs zero-cost proxies and evolutionary search, to explore model architectures with low transferability. Our method, namely ArchLock, aims to achieve high performance on the source task, while degrading the performance on potential target tasks, i.e., locking the transferability of a DNN model. To achieve efficient cross-task search without accurately knowing the training data owned by the attackers, we utilize zero-cost proxies to speed up architecture evaluation and simulate potential target task embeddings to assist cross-task search with a binary performance predictor. Extensive experiments on NAS-Bench-201 and TransNAS-Bench-101 demonstrate that ArchLock reduces transferability by up to 30% and 50%, respectively, with negligible performance degradation on source tasks (<2%). The code is available at https://github.com/Tongzhou0101/ArchLock. 

Planar Josephson junctions (JJs), based on common superconductors and III–V semiconductors, are sought for Majorana states and fault-tolerant quantum computing. However, with gate-tunable spin–orbit coupling (SOC), we show that the range of potential applications of such JJs becomes much broader. The time-dependent SOC offers unexplored mechanisms for switching JJs, accompanied by the 2π-phase jumps and the voltage pulses corresponding to the single-flux-quantum transitions, key to high-speed and low-power superconducting electronics. In a constant applied magnetic field, with Rashba and Dresselhaus SOC, anharmonic current-phase relations, calculated microscopically in these JJs, yield a nonreciprocal transport and superconducting diode effect. Together with the time-dependent SOC, this allows us to identify a switching mechanism at no applied current bias, which supports fractional-flux-quantum superconducting circuits and neuromorphic computing. 

SUMMARY We present a new 3-D radially anisotropic seismic velocity model EARA2024 of the crust and mantle beneath East Asia and the northwestern Pacific using adjoint full-waveform inversion tomography. We construct the EARA2024 model by iteratively minimizing the waveform similarity misfit between the synthetic and observed waveforms from 142 earthquakes recorded by about 2000 broad-band stations in East Asia. Compared to previous studies, this new model renders significantly improved images of the subducted oceanic plate in the upper mantle, mantle transition zone, and uppermost lower mantle along the Kuril, Japan, Izu-Bonin and Ryukyu Trenches. Complex slab deformation and break-offs are observed at different depths. Moreover, our model provides new insights into the origins of intraplate volcanoes in East Asia, including the Changbaishan, Datong-Fengzhen, Tengchong and Hainan volcanic fields. 

A planar Josephson junction is a versatile platform to realize topological superconductivity over a large parameter space and host Majorana bound states. With a change in the Zeeman field, this system undergoes a transition from trivial to topological superconductivity accompanied by a jump in the superconducting phase difference between the two superconductors. A standard model of these Josephson junctions, which can be fabricated to have a nearly perfect interfacial transparency, predicts a simple universal behavior. In that model, at the same value of Zeeman field for the topological transition, there is a π phase jump and a minimum in the critical superconducting current, while applying a controllable phase difference yields a diamond-shaped topological region as a function of that phase difference and a Zeeman field. In contrast, even for a perfect interfacial transparency, we find a much richer and nonuniversal behavior as the width of the superconductor is varied or the Dresselhaus spin–orbit coupling is considered. The Zeeman field for the phase jump, not necessarily π, is different from the value for the minimum critical current, while there is a strong deviation from the diamond-like topological region. These Josephson junctions show a striking example of a nonreciprocal transport and superconducting diode effect, revealing the importance of our findings not only for topological superconductivity and fault-tolerant quantum computing but also for superconducting spintronics.