Search

Large Language Models (LLMs) face increasing loss spikes during scaling, undermining training stability and final performance. While gradient clipping mitigates this issue, traditional global approaches poorly handle parameter-specific gradient variations and decaying gradient norms. We propose **AdaGC**, an adaptive gradient clipping framework that automatically adjusts local thresholds per parameter through exponential moving average of gradient norms. Theoretical analysis proves AdaGC's convergence under non-convex conditions. Extensive experiments demonstrate significant improvements: On Llama-2 7B/13B, AdaGC completely eliminates loss spikes while reducing WikiText perplexity by 3.5% (+0.14pp LAMBADA accuracy) for 7B and achieving 0.65% lower training loss with 1.47% reduced validation perplexity for 13B compared to global clipping. For CLIP ViT-Base, AdaGC converges 25% faster than StableAdamW with full spike elimination. The method shows universal effectiveness across architectures (Llama-2 7B/13B) and modalities (CLIP), with successful integration into diverse optimizers like AdamW and Lion. Source code will be released on GitHub.

A variety of protective or reactive environmental gases have recently gained growing attention in laser-based metal additive manufacturing (AM) technologies due to their unique thermophysical properties and the potential improvements they can bring to the build processes. However, much remains unclear regarding the effects of different gas environments on critical phenomena in laser AM, such as rapid cooling, energy coupling, and defect generation. Through simultaneous high-speed synchrotron x-ray imaging and thermal imaging, we identify distinct effects of various environmental gases in laser AM and gained a deeper understanding of the underlying mechanisms. Compared to the commonly used protective gas, argon, it is found that helium has a negligible effect on cooling the part. However, helium can suppress unstable keyholes by decreasing effective energy absorption, thus mitigating keyhole porosity generation and reducing pore size under certain processing conditions. These observations provide guidelines for the strategic use of environmental gases in laser AM to produce parts with improved quality.

Code review remains a critical yet resource-intensive process in software development, particularly challenging in large-scale industrial environments. While Large Language Models (LLMs) show promise for automating code review, existing solutions face significant limitations in precision and practicality. This paper presents BitsAI-CR, an innovative framework that enhances code review through a two-stage approach combining RuleChecker for initial issue detection and ReviewFilter for precision verification. The system is built upon a comprehensive taxonomy of review rules and implements a data flywheel mechanism that enables continuous performance improvement through structured feedback and evaluation metrics. Our approach introduces an Outdated Rate metric that can reflect developers' actual adoption of review comments, enabling automated evaluation and systematic optimization at scale. Empirical evaluation demonstrates BitsAI-CR's effectiveness, achieving 75.0% precision in review comment generation. For the Go language which has predominant usage at ByteDance, we maintain an Outdated Rate of 26.7%. The system has been successfully deployed at ByteDance, serving over 12,000 Weekly Active Users (WAU). Our work provides valuable insights into the practical application of automated code review and offers a blueprint for organizations seeking to implement automated code reviews at scale.

Large language models (LLMs) have shown impressive performance across a range of natural language processing tasks. However, their vast number of parameters introduces significant memory challenges during training, particularly when using memory-intensive optimizers like Adam. Existing memory-efficient algorithms often rely on techniques such as singular value decomposition projection or weight freezing. While these approaches help alleviate memory constraints, they generally produce suboptimal results compared to full-rank updates. In this paper, we investigate the memory-efficient method beyond low-rank training, proposing a novel solution called Gradient Wavelet Transform (GWT), which applies wavelet transforms to gradients in order to significantly reduce the memory requirements for maintaining optimizer states. We demonstrate that GWT can be seamlessly integrated with memory-intensive optimizers, enabling efficient training without sacrificing performance. Through extensive experiments on both pre-training and fine-tuning tasks, we show that GWT achieves state-of-the-art performance compared with advanced memory-efficient optimizers and full-rank approaches in terms of both memory usage and training performance.

Sharpness-Aware Minimization (SAM) has proven highly effective in improving model generalization in machine learning tasks. However, SAM employs a fixed hyperparameter associated with the regularization to characterize the sharpness of the model. Despite its success, research on adaptive regularization methods based on SAM remains scarce. In this paper, we propose the SAM with Adaptive Regularization (SAMAR), which introduces a flexible sharpness ratio rule to update the regularization parameter dynamically. We provide theoretical proof of the convergence of SAMAR for functions satisfying the Lipschitz continuity. Additionally, experiments on image recognition tasks using CIFAR-10 and CIFAR-100 demonstrate that SAMAR enhances accuracy and model generalization.

With the growing demand for seamless connectivity and intelligent communication, the integration of artificial intelligence (AI) and sixth-generation (6G) communication networks has emerged as a transformative paradigm. By embedding AI capabilities across various network layers, this integration enables optimized resource allocation, improved efficiency, and enhanced system robust performance, particularly in intricate and dynamic environments. This paper presents a comprehensive overview of AI and communication for 6G networks, with a focus on emphasizing their foundational principles, inherent challenges, and future research opportunities. We first review the integration of AI and communications in the context of 6G, exploring the driving factors behind incorporating AI into wireless communications, as well as the vision for the convergence of AI and 6G. The discourse then transitions to a detailed exposition of the envisioned integration of AI within 6G networks, delineated across three progressive developmental stages. The first stage, AI for Network, focuses on employing AI to augment network performance, optimize efficiency, and enhance user service experiences. The second stage, Network for AI, highlights the role of the network in facilitating and buttressing AI operations and presents key enabling technologies, such as digital twins for AI and semantic communication. In the final stage, AI as a Service, it is anticipated that future 6G networks will innately provide AI functions as services, supporting application scenarios like immersive communication and intelligent industrial robots. In addition, we conduct an in-depth analysis of the critical challenges faced by the integration of AI and communications in 6G. Finally, we outline promising future research opportunities that are expected to drive the development and refinement of AI and 6G communications.

This paper introduces a comprehensive planning and navigation framework that address these limitations by integrating semantic mapping, adaptive coverage planning, dynamic obstacle avoidance and precise trajectory tracking. Our framework begins by generating panoptic occupancy local semantic maps and accurate localization information from data aligned between a monocular camera, IMU, and GPS. This information is combined with input terrain point clouds or preloaded terrain information to initialize the planning process. We propose the Radiant Field-Informed Coverage Planning algorithm, which utilizes a diffusion field model to dynamically adjust the robot's coverage trajectory and speed based on environmental attributes such as dirtiness and dryness. By modeling the spatial influence of the robot's actions using a Gaussian field, ensures a speed-optimized, uniform coverage trajectory while adapting to varying environmental conditions.

Due to the rapid development of quantum computing, many classical blockchain technologies are now considered insecure. The emergence of quantum blockchain holds promise for addressing this issue. Various quantum consensus algorithms have been proposed so far, but there has not yet been a quantum consensus algorithm tailored specifically for consortium blockchain scenarios. In this paper, we propose a novel quantum consensus mechanism, named Q-PnV. This consensus mechanism is based on the classical Proof of Vote (PoV), integrating quantum voting, quantum digital signature and quantum random number generators (QRNGs). By combining Q-PnV with a quantum blockchain using weighted hypergraph states, we propose a comprehensive quantum blockchain solution for consortium blockchain scenarios. Compared to the classical method, the quantum blockchain based on Q-PnV can resist quantum attacks and shows significant improvements in security and fairness, making it better suit-ed for the future quantum era.

This paper proposes a Bayesian factor-augmented bundle choice model to estimate joint consumption as well as the substitutability and complementarity of multiple goods in the presence of endogenous regressors. The model extends the two primary treatments of endogeneity in existing bundle choice models: (1) endogenous market-level prices and (2) time-invariant unobserved individual heterogeneity. A Bayesian sparse factor approach is employed to capture high-dimensional error correlations that induce taste correlation and endogeneity. Time-varying factor loadings allow for more general individual-level and time-varying heterogeneity and endogeneity, while the sparsity induced by the shrinkage prior on loadings balances flexibility with parsimony. Applied to a soda tax in the context of complementarities, the new approach captures broader effects of the tax that were previously overlooked. Results suggest that a soda tax could yield additional health benefits by marginally decreasing the consumption of salty snacks along with sugary drinks, extending the health benefits beyond the reduction in sugar consumption alone.

As the capabilities of code large language models (LLMs) continue to expand, their applications across diverse code intelligence domains are rapidly increasing. However, most existing datasets only evaluate limited application domains. To address this gap, we have developed a comprehensive code evaluation dataset FullStack Bench focusing on full-stack programming, which encompasses a wide range of application domains (e.g., basic programming, data analysis, software engineering, mathematics, and machine learning). Besides, to assess multilingual programming capabilities, in FullStack Bench, we design real-world instructions and corresponding unit test cases from 16 widely-used programming languages to reflect real-world usage scenarios rather than simple translations. Moreover, we also release an effective code sandbox execution tool (i.e., SandboxFusion) supporting various programming languages and packages to evaluate the performance of our FullStack Bench efficiently. Comprehensive experimental results on our FullStack Bench demonstrate the necessity and effectiveness of our FullStack Bench and SandboxFusion., Comment: 26 pages

Uncertainty estimation is a standard tool to quantify the reliability of modern deep learning models, and crucial for many real-world applications. However, efficient uncertainty estimation methods for spiking neural networks, particularly for regression models, have been lacking. Here, we introduce two methods that adapt the Average-Over-Time Spiking Neural Network (AOT-SNN) framework to regression tasks, enhancing uncertainty estimation in event-driven models. The first method uses the heteroscedastic Gaussian approach, where SNNs predict both the mean and variance at each time step, thereby generating a conditional probability distribution of the target variable. The second method leverages the Regression-as-Classification (RAC) approach, reformulating regression as a classification problem to facilitate uncertainty estimation. We evaluate our approaches on both a toy dataset and several benchmark datasets, demonstrating that the proposed AOT-SNN models achieve performance comparable to or better than state-of-the-art deep neural network methods, particularly in uncertainty estimation. Our findings highlight the potential of SNNs for uncertainty estimation in regression tasks, providing an efficient and biologically inspired alternative for applications requiring both accuracy and energy efficiency.

Full-field ultra-high-speed (UHS) x-ray imaging experiments have been well established to characterize various processes and phenomena. However, the potential of UHS experiments through the joint acquisition of x-ray videos with distinct configurations has not been fully exploited. In this paper, we investigate the use of a deep learning-based spatio-temporal fusion (STF) framework to fuse two complementary sequences of x-ray images and reconstruct the target image sequence with high spatial resolution, high frame rate, and high fidelity. We applied a transfer learning strategy to train the model and compared the peak signal-to-noise ratio (PSNR), average absolute difference (AAD), and structural similarity (SSIM) of the proposed framework on two independent x-ray datasets with those obtained from a baseline deep learning model, a Bayesian fusion framework, and the bicubic interpolation method. The proposed framework outperformed the other methods with various configurations of the input frame separations and image noise levels. With 3 subsequent images from the low resolution (LR) sequence of a 4-time lower spatial resolution and another 2 images from the high resolution (HR) sequence of a 20-time lower frame rate, the proposed approach achieved an average PSNR of 37.57 dB and 35.15 dB, respectively. When coupled with the appropriate combination of high-speed cameras, the proposed approach will enhance the performance and therefore scientific value of the UHS x-ray imaging experiments.

Predicting electronic band structures from crystal structures is crucial for understanding structure-property correlations in materials science. First-principles approaches are accurate but computationally intensive. Recent years, machine learning (ML) has been extensively applied to this field, while existing ML models predominantly focus on band gap predictions or indirect band structure estimation via solving predicted Hamiltonians. An end-to-end model to predict band structure accurately and efficiently is still lacking. Here, we introduce a graph Transformer-based end-to-end approach that directly predicts band structures from crystal structures with high accuracy. Our method leverages the continuity of the k-path and treat continuous bands as a sequence. We demonstrate that our model not only provides accurate band structure predictions but also can derive other properties (such as band gap, band center, and band dispersion) with high accuracy. We verify the model performance on large and diverse datasets., Comment: 8 pages, 3 figures

Chiral phonons are vibrational modes in a crystal that possess a well-defined handedness or chirality, typically found in materials that lack inversion symmetry. Here we report the discovery of chiral phonon modes in the kagome ferromagnetic Weyl semimetal Co3Sn2S2, a material that preserves inversion symmetry but breaks time-reversal symmetry. Using helicity-resolved magneto-Raman spectroscopy, we observe the spontaneous splitting of the doubly degenerate in-plane Eg modes into two distinct chiral phonon modes of opposite helicity when the sample is zero-field cooled below the Curie temperature, in the absence of an external magnetic field. As we sweep the out-of-plane magnetic field, this Eg phonon splitting exhibits a well-defined hysteresis loop directly correlated with the material's magnetization. The observed spontaneous splitting reaches up to 1.27 cm-1 at low temperatures, progressively diminishes with increasing temperature, and completely vanishes near the Curie temperature. Our findings highlight the role of the magnetic order in inducing chiral phonons, paving the way for novel methods to manipulate chiral phonons through magnetization and vice versa. Additionally, our work introduces new possibilities for controlling chiral Weyl fermions using chiral phonons.

Code large language models (LLMs) have made significant progress in code debugging by directly generating the correct code based on the buggy code snippet. Programming benchmarks, typically consisting of buggy code snippet and their associated test cases, are used to assess the debugging capabilities of LLMs. However, many existing benchmarks primarily focus on Python and are often limited in terms of language diversity (e.g., DebugBench and DebugEval). To advance the field of multilingual debugging with LLMs, we propose the first massively multilingual debugging benchmark, which includes 3.6K test samples of 18 programming languages and covers the automated program repair (APR) task, the code review (CR) task, and the bug identification (BI) task. Further, we introduce the debugging instruction corpora MDEVAL-INSTRUCT by injecting bugs into the correct multilingual queries and solutions (xDebugGen). Further, a multilingual debugger xDebugCoder trained on MDEVAL-INSTRUCT as a strong baseline specifically to handle the bugs of a wide range of programming languages (e.g. "Missing Mut" in language Rust and "Misused Macro Definition" in language C). Our extensive experiments on MDEVAL reveal a notable performance gap between open-source models and closed-source LLMs (e.g., GPT and Claude series), highlighting huge room for improvement in multilingual code debugging scenarios., Comment: 15 pages

This paper investigates the roles of gradient normalization and clipping in ensuring the convergence of Stochastic Gradient Descent (SGD) under heavy-tailed noise. While existing approaches consider gradient clipping indispensable for SGD convergence, we theoretically demonstrate that gradient normalization alone without clipping is sufficient to ensure convergence. Furthermore, we establish that combining gradient normalization with clipping offers significantly improved convergence rates compared to using either technique in isolation, notably as gradient noise diminishes. With these results, our work provides the first theoretical evidence demonstrating the benefits of gradient normalization in SGD under heavy-tailed noise. Finally, we introduce an accelerated SGD variant incorporating gradient normalization and clipping, further enhancing convergence rates under heavy-tailed noise.

Modeling trajectory data with generic-purpose dense representations has become a prevalent paradigm for various downstream applications, such as trajectory classification, travel time estimation and similarity computation. However, existing methods typically rely on trajectories from a single spatial view, limiting their ability to capture the rich contextual information that is crucial for gaining deeper insights into movement patterns across different geospatial contexts. To this end, we propose MVTraj, a novel multi-view modeling method for trajectory representation learning. MVTraj integrates diverse contextual knowledge, from GPS to road network and points-of-interest to provide a more comprehensive understanding of trajectory data. To align the learning process across multiple views, we utilize GPS trajectories as a bridge and employ self-supervised pretext tasks to capture and distinguish movement patterns across different spatial views. Following this, we treat trajectories from different views as distinct modalities and apply a hierarchical cross-modal interaction module to fuse the representations, thereby enriching the knowledge derived from multiple sources. Extensive experiments on real-world datasets demonstrate that MVTraj significantly outperforms existing baselines in tasks associated with various spatial views, validating its effectiveness and practical utility in spatio-temporal modeling.

Recently, backdoor attack has become an increasing security threat to deep neural networks and drawn the attention of researchers. Backdoor attacks exploit vulnerabilities in third-party pretrained models during the training phase, enabling them to behave normally for clean samples and mispredict for samples with specific triggers. Existing backdoor attacks mainly focus on balanced datasets. However, real-world datasets often follow long-tailed distributions. In this paper, for the first time, we explore backdoor attack on such datasets. Specifically, we first analyze the influence of data imbalance on backdoor attack. Based on our analysis, we propose an effective backdoor attack named Dynamic Data Augmentation Operation (D$^2$AO). We design D$^2$AO selectors to select operations depending jointly on the class, sample type (clean vs. backdoored) and sample features. Meanwhile, we develop a trigger generator to generate sample-specific triggers. Through simultaneous optimization of the backdoored model and trigger generator, guided by dynamic data augmentation operation selectors, we achieve significant advancements. Extensive experiments demonstrate that our method can achieve the state-of-the-art attack performance while preserving the clean accuracy.

Preference alignment in Large Language Models (LLMs) has significantly improved their ability to adhere to human instructions and intentions. However, existing direct alignment algorithms primarily focus on relative preferences and often overlook the qualitative aspects of responses. Striving to maximize the implicit reward gap between the chosen and the slightly inferior rejected responses can cause overfitting and unnecessary unlearning of the high-quality rejected responses. The unawareness of the reward scores also drives the LLM to indiscriminately favor the low-quality chosen responses and fail to generalize to responses with the highest rewards, which are sparse in data. To overcome these shortcomings, our study introduces reward-conditioned LLM policies that discern and learn from the entire spectrum of response quality within the dataset, helping extrapolate to more optimal regions. We propose an effective yet simple data relabeling method that conditions the preference pairs on quality scores to construct a reward-augmented dataset. This dataset is easily integrated with existing direct alignment algorithms and is applicable to any preference dataset. The experimental results across instruction-following benchmarks including AlpacaEval, MT-Bench, and Arena-Hard-Auto demonstrate that our approach consistently boosts the performance of DPO by a considerable margin across diverse models. Additionally, our method improves the average accuracy on various academic benchmarks. When applying our method to on-policy data, the resulting DPO model achieves SOTA results on AlpacaEval. Through ablation studies, we demonstrate that our method not only maximizes the utility of preference data but also mitigates the issue of unlearning, demonstrating its broad effectiveness beyond mere dataset expansion. Our code is available at https://github.com/shenao-zhang/reward-augmented-preference.

The emergence of Vision-Language Models (VLMs) represents a significant advancement in integrating computer vision with Large Language Models (LLMs) to generate detailed text descriptions from visual inputs. Despite their growing importance, the security of VLMs, particularly against backdoor attacks, is under explored. Moreover, prior works often assume attackers have access to the original training data, which is often unrealistic. In this paper, we address a more practical and challenging scenario where attackers must rely solely on Out-Of-Distribution (OOD) data. We introduce VLOOD (Backdooring Vision-Language Models with Out-of-Distribution Data), a novel approach with two key contributions: (1) demonstrating backdoor attacks on VLMs in complex image-to-text tasks while minimizing degradation of the original semantics under poisoned inputs, and (2) proposing innovative techniques for backdoor injection without requiring any access to the original training data. Our evaluation on image captioning and visual question answering (VQA) tasks confirms the effectiveness of VLOOD, revealing a critical security vulnerability in VLMs and laying the foundation for future research on securing multimodal models against sophisticated threats., Comment: ICLR 2025

Large language models (LLMs) have become increasingly pivotal across various domains, especially in handling complex data types. This includes structured data processing, as exemplified by ChartQA and ChatGPT-Ada, and multimodal unstructured data processing as seen in Visual Question Answering (VQA). These areas have attracted significant attention from both industry and academia. Despite this, there remains a lack of unified evaluation methodologies for these diverse data handling scenarios. In response, we introduce BabelBench, an innovative benchmark framework that evaluates the proficiency of LLMs in managing multimodal multistructured data with code execution. BabelBench incorporates a dataset comprising 247 meticulously curated problems that challenge the models with tasks in perception, commonsense reasoning, logical reasoning, and so on. Besides the basic capabilities of multimodal understanding, structured data processing as well as code generation, these tasks demand advanced capabilities in exploration, planning, reasoning and debugging. Our experimental findings on BabelBench indicate that even cutting-edge models like ChatGPT 4 exhibit substantial room for improvement. The insights derived from our comprehensive analysis offer valuable guidance for future research within the community. The benchmark data can be found at https://github.com/FFD8FFE/babelbench.

In various domains, the increasing application of machine learning allows researchers to access inexpensive predictive data, which can be utilized as auxiliary data for statistical inference. Although such data are often unreliable compared to gold-standard datasets, Prediction-Powered Inference (PPI) has been proposed to ensure statistical validity despite the unreliability. However, the challenge of `data silos' arises when the private gold-standard datasets are non-shareable for model training, leading to less accurate predictive models and invalid inferences. In this paper, we introduces the Federated Prediction-Powered Inference (Fed-PPI) framework, which addresses this challenge by enabling decentralized experimental data to contribute to statistically valid conclusions without sharing private information. The Fed-PPI framework involves training local models on private data, aggregating them through Federated Learning (FL), and deriving confidence intervals using PPI computation. The proposed framework is evaluated through experiments, demonstrating its effectiveness in producing valid confidence intervals.

Speech enhancement (SE) improves communication in noisy environments, affecting areas such as automatic speech recognition, hearing aids, and telecommunications. With these domains typically being power-constrained and event-based while requiring low latency, neuromorphic algorithms in the form of spiking neural networks (SNNs) have great potential. Yet, current effective SNN solutions require a contextual sampling window imposing substantial latency, typically around 32ms, too long for many applications. Inspired by Dual-Path Spiking Neural Networks (DPSNNs) in classical neural networks, we develop a two-phase time-domain streaming SNN framework -- the Dual-Path Spiking Neural Network (DPSNN). In the DPSNN, the first phase uses Spiking Convolutional Neural Networks (SCNNs) to capture global contextual information, while the second phase uses Spiking Recurrent Neural Networks (SRNNs) to focus on frequency-related features. In addition, the regularizer suppresses activation to further enhance energy efficiency of our DPSNNs. Evaluating on the VCTK and Intel DNS Datasets, we demonstrate that our approach achieves the very low latency (approximately 5ms) required for applications like hearing aids, while demonstrating excellent signal-to-noise ratio (SNR), perceptual quality, and energy efficiency.

Ultrasonic processing of solidifying metals in additive manufacturing can provide grain refinement and advantageous mechanical properties. However, the specific physical mechanisms of microstructural refinement relevant to laser-based additive manufacturing have not been directly observed because of sub-millimeter length scales and rapid solidification rates associated with melt pools. Here, high-speed synchrotron X-ray imaging is used to observe the effect of ultrasonic vibration directly on melt pool dynamics and solidification of Al6061 alloy. The high temporal and spatial resolution enabled direct observation of cavitation effects driven by a 20.2 kHz ultrasonic source. We utilized multiphysics simulations to validate the postulated connection between ultrasonic treatment and solidification. The X-ray results show a decrease in melt pool and keyhole depth fluctuations during melting and promotion of pore migration toward the melt pool surface with applied sonication. Additionally, the simulation results reveal increased localized melt pool flow velocity, cooling rates, and thermal gradients with applied sonication. This work shows how ultrasonic treatment can impact melt pools and its potential for improving part quality.

STOchastic Recursive Momentum (STORM)-based algorithms have been widely developed to solve one to $K$-level ($K \geq 3$) stochastic optimization problems. Specifically, they use estimators to mitigate the biased gradient issue and achieve near-optimal convergence results. However, there is relatively little work on understanding their generalization performance, particularly evident during the transition from one to $K$-level optimization contexts. This paper provides a comprehensive generalization analysis of three representative STORM-based algorithms: STORM, COVER, and SVMR, for one, two, and $K$-level stochastic optimizations under both convex and strongly convex settings based on algorithmic stability. Firstly, we define stability for $K$-level optimizations and link it to generalization. Then, we detail the stability results for three prominent STORM-based algorithms. Finally, we derive their excess risk bounds by balancing stability results with optimization errors. Our theoretical results provide strong evidence to complete STORM-based algorithms: (1) Each estimator may decrease their stability due to variance with its estimation target. (2) Every additional level might escalate the generalization error, influenced by the stability and the variance between its cumulative stochastic gradient and the true gradient. (3) Increasing the batch size for the initial computation of estimators presents a favorable trade-off, enhancing the generalization performance.