Large Language Models (LLMs) are now integral across various domains and have demonstrated impressive performance. Progress, however, rests on the premise that benchmark scores are both accurate and reproducible. We demonstrate that the reproducibility of LLM performance is fragile: changing system configuration, such as evaluation batch size, GPU count, and GPU version, can introduce significant differences in the generated responses. This issue is especially pronounced in reasoning models, where minor rounding differences in early tokens can cascade into divergent chains of thought, ultimately affecting accuracy. For instance, under bfloat16 precision with greedy decoding, a reasoning model like DeepSeek-R1-Distill-Qwen-7B can exhibit up to 9\% variation in accuracy and 9,000 tokens difference in response length due to differences in GPU count, type, and evaluation batch size. We trace the root cause of this variability to the non-associative nature of floating-point arithmetic under limited numerical precision. This work presents the first systematic investigation into how numerical precision affects reproducibility in LLM inference. Through carefully controlled experiments across various hardware, software, and precision settings, we quantify when and how model outputs diverge. Our analysis reveals that floating-point precision—while critical for reproducibility—is often neglected in evaluation practices. Inspired by this, we develop a lightweight inference pipeline, dubbed LayerCast, that stores weights in 16-bit precision but performs all computations in FP32, balancing memory efficiency with numerical stability. Code is available at https://github.com/nanomaoli/llm_reproducibility.

Large text-to-image models have revolutionized the ability to generate imagery using natural language. However, particularly unique or personal visual concepts, such as pets and furniture, will not be captured by the original model. This has led to interest in how to personalize a text-to-image model. Despite significant progress, this task remains a formidable challenge, particularly in preserving the subject's identity. Most researchers attempt to address this issue by modifying model architectures. These methods are capable of keeping the subject structure and color but fail to preserve identity details. Towards this issue, our approach takes a data-centric perspective. We introduce a novel regularization dataset generation strategy on both the text and image level. This strategy enables the model to preserve fine details of the desired subjects, such as text and logos. Our method is architecture-agnostic and can be flexibly applied on various text-to-image models. We show on established benchmarks that our data-centric approach forms the new state of the art in terms of identity preservation and text alignment.

We have witnessed significant progress in deep learning-based 3D vision, ranging from neural radiance field (NeRF) based 3D representation learning to applications in novel view synthesis (NVS). However, existing scene-level datasets for deep learning-based 3D vision, limited to ei-ther synthetic environments or a narrow selection of real-world scenes, are quite insufficient. This insufficiency not only hinders a comprehensive benchmark of existing methods but also caps what could be explored in deep learning-based 3D analysis. To address this critical gap, we present DL3DV-10K, a large-scale scene dataset, featuring 51.2 million frames from 10,510 videos captured from 65 types of point- of-interest (POI) locations, covering both bounded and unbounded scenes, with different levels of reflection, transparency, and lighting. We conducted a comprehensive benchmark of recent NVS methods on DL3DV-10K, which revealed valuable insights for future research in NVS. In addition, we have obtained encouraging results in a pilot study to learn generalizable NeRF from DL3DV-10K, which manifests the necessity of a large-scale scene-level dataset to forge a path toward a foundation model for learning 3D representation. Our DL3DV-10K dataset, benchmark results, and models will be publicly accessible.

With the increasing demand to deploy convolutional neural networks (CNNs) on 5G mobile platforms, architecture designs with efficient sparse kernels (SKs) were proposed, which can save more parameters than the standard convolution while maintaining the high accuracy. Despite the great potential, neural network designs with SKs still require a lot of expert knowledge and take ample time. In this paper, we first propose a search scheme that effectively reduces the SK design space based on three aspects: composition, performance, and efficiency. Meanwhile, we completely eliminate the model training from our search scheme. Instead, an easily measurable quantity, the information field, is identified and used to predict the model accuracy in the searching process. Additionally, we provide a detailed efficiency analysis on the final designs found by our scheme. Second, based on the analysis we propose a model transformation scheme to better utilize the SK designs on existing models to either reduce the number of parameters or increase the accuracy. Last, considering the extra programming overhead and the expert knowledge required by the model transformation scheme, we develop a compiler prototype to automate the entire process, given the source code of an existing model. Experimental results show that models composed of the sparse kernel designs searched by our search scheme can beat state-of-the-art networks such as ResNets in terms of the accuracy and the efficiency. Also by using our model transformation scheme we can easily improve the accuracy (the same number of parameters) or the efficiency (the same accuracy) upon existing state-of-the-art models.

Recent studies have highlighted audio adversarial examples as a ubiquitous threat to state-of-the-art automatic speech recognition systems. Thorough studies on how to effectively generate adversarial examples are essential to prevent potential attacks. Despite many research on this, the efficiency and the robustness of existing works are not yet satisfactory. In this paper, we propose weighted-sampling audio adversarial examples, focusing on the numbers and the weights of distortion to reinforce the attack. Further, we apply a denoising method in the loss function to make the adversarial attack more imperceptible. Experiments show that our method is the first in the field to generate audio adversarial examples with low noise and high audio robustness at the minute time-consuming level 1.

MOTIVATION: With the rapid development of high-throughput technologies, parallel acquisition of large-scale drug-informatics data provides significant opportunities to improve pharmaceutical research and development. One important application is the purpose prediction of small-molecule compounds with the objective of specifying the therapeutic properties of extensive purpose-unknown compounds and repurposing the novel therapeutic properties of FDA-approved drugs. Such a problem is extremely challenging because compound attributes include heterogeneous data with various feature patterns, such as drug fingerprints, drug physicochemical properties and drug perturbation gene expressions. Moreover, there is a complex non-linear dependency among heterogeneous data. In this study, we propose a novel domain-adversarial multi-task framework for integrating shared knowledge from multiple domains. The framework first uses an adversarial strategy to learn target representations and then models non-linear dependency among several domains. RESULTS: Experiments on two real-world datasets illustrate that our approach achieves an obvious improvement over competitive baselines. The novel therapeutic properties of purpose-unknown compounds that we predicted have been widely reported or brought to clinics. Furthermore, our framework can integrate various attributes beyond the three domains examined herein and can be applied in industry for screening significant numbers of small-molecule drug candidates. AVAILABILITY AND IMPLEMENTATION: The source code and datasets are available at https://github.com/JohnnyY8/DAMT-Model. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Recently convolutional neural networks (CNNs) achieve great accuracy in visual recognition tasks. DenseNets become one of the most popular CNN models due to its effectiveness in the feature-reuse. However, like other CNN models, DenseNets also face the overfitting problem if not more severe. Existing dropout methods can be applied but not effective. In particular, the property of the feature-reuse in DenseNets will be impeded, and the dropout effect will be weakened by the spatial correlation inside feature maps. To address these problems, we craft the design of a specialized dropout method from three aspects, the dropout location, the dropout granularity, and the dropout probability. The insights attained here could potentially be applied as a general approach for boosting the accuracy of other CNN models with similar shortcut connections. Experimental results show that DenseNets with our specialized dropout method yield better accuracies compared to vanilla DenseNets and state-of-the-art CNN models, and such accuracy boost increases with the model depth.