Muon is Scalable for LLM Training

Recently, the Muon optimizer based on matrix orthogonalization has demonstrated strong results in training small-scale language models, but the scalability to larger models has not been proven. We identify two crucial techniques for scaling up Muon: (1) adding weight decay and (2) carefully adjusting the per-parameter update scale. These techniques allow Muon to work out-of-the-box on large-scale training without the need of hyper-parameter tuning. Scaling law experiments indicate that Muon achieves $\sim\!2\times$ computational efficiency compared to AdamW with compute optimal training. Based on these improvements, we introduce Moonlight, a 3B/16B-parameter Mixture-of-Expert (MoE) model trained with 5.7T tokens using Muon. Our model improves the current Pareto frontier, achieving better performance with much fewer training FLOPs compared to prior models. We open-source our distributed Muon implementation that is memory optimal and communication efficient. We also release the pretrained, instruction-tuned, and intermediate checkpoints to support future research.

MoBA: Mixture of Block Attention for Long-Context LLMs

Scaling the effective context length is essential for advancing large language models (LLMs) toward artificial general intelligence (AGI). However, the quadratic increase in computational complexity inherent in traditional attention mechanisms presents a prohibitive overhead. Existing approaches either impose strongly biased structures, such as sink or window attention which are task-specific, or radically modify the attention mechanism into linear approximations, whose performance in complex reasoning tasks remains inadequately explored. In this work, we propose a solution that adheres to the ``less structure'' principle, allowing the model to determine where to attend autonomously, rather than introducing predefined biases. We introduce Mixture of Block Attention (MoBA), an innovative approach that applies the principles of Mixture of Experts (MoE) to the attention mechanism. This novel architecture demonstrates superior performance on long-context tasks while offering a key advantage: the ability to seamlessly transition between full and sparse attention, enhancing efficiency without the risk of compromising performance. MoBA has already been deployed to support Kimi's long-context requests and demonstrates significant advancements in efficient attention computation for LLMs. Our code is available at https://github.com/MoonshotAI/MoBA.

A Plug-and-Play Bregman ADMM Module for Inferring Event Branches in Temporal Point Processes

An event sequence generated by a temporal point process is often associated with a hidden and structured event branching process that captures the triggering relations between its historical and current events. In this study, we design a new plug-and-play module based on the Bregman ADMM (BADMM) algorithm, which infers event branches associated with event sequences in the maximum likelihood estimation framework of temporal point processes (TPPs). Specifically, we formulate the inference of event branches as an optimization problem for the event transition matrix under sparse and low-rank constraints, which is embedded in existing TPP models or their learning paradigms. We can implement this optimization problem based on subspace clustering and sparse group-lasso, respectively, and solve it using the Bregman ADMM algorithm, whose unrolling leads to the proposed BADMM module. When learning a classic TPP (e.g., Hawkes process) by the expectation-maximization algorithm, the BADMM module helps derive structured responsibility matrices in the E-step. Similarly, the BADMM module helps derive low-rank and sparse attention maps for the neural TPPs with self-attention layers. The structured responsibility matrices and attention maps, which work as learned event transition matrices, indicate event branches, e.g., inferring isolated events and those key events triggering many subsequent events. Experiments on both synthetic and real-world data show that plugging our BADMM module into existing TPP models and learning paradigms can improve model performance and provide us with interpretable structured event branches. The code is available at \url{https://github.com/qingmeiwangdaily/BADMM_TPP}.

Layer Importance and Hallucination Analysis in Large Language Models via Enhanced Activation Variance-Sparsity

Evaluating the importance of different layers in large language models (LLMs) is crucial for optimizing model performance and interpretability. This paper first explores layer importance using the Activation Variance-Sparsity Score (AVSS), which combines normalized activation variance and sparsity to quantify each layer's contribution to overall model performance. By ranking layers based on AVSS and pruning the least impactful 25\%, our experiments on tasks such as question answering, language modeling, and sentiment classification show that over 90\% of the original performance is retained, highlighting potential redundancies in LLM architectures. Building on AVSS, we propose an enhanced version tailored to assess hallucination propensity across layers (EAVSS). This improved approach introduces Hallucination-Specific Activation Variance (HSAV) and Hallucination-Specific Sparsity (HSS) metrics, allowing precise identification of hallucination-prone layers. By incorporating contrastive learning on these layers, we effectively mitigate hallucination generation, contributing to more robust and efficient LLMs(The maximum performance improvement is 12\%). Our results on the NQ, SciQ, TriviaQA, TruthfulQA, and WikiQA datasets demonstrate the efficacy of this method, offering a comprehensive framework for both layer importance evaluation and hallucination mitigation in LLMs.

AVSS: Layer Importance Evaluation in Large Language Models via Activation Variance-Sparsity Analysis

The evaluation of layer importance in deep learning has been an active area of research, with significant implications for model optimization and interpretability. Recently, large language models (LLMs) have gained prominence across various domains, yet limited studies have explored the functional importance and performance contributions of individual layers within LLMs, especially from the perspective of activation distribution. In this work, we propose the Activation Variance-Sparsity Score (AVSS), a novel metric combining normalized activation variance and sparsity to assess each layer's contribution to model performance. By identifying and removing approximately the lowest 25% of layers based on AVSS, we achieve over 90% of original model performance across tasks such as question answering, language modeling, and sentiment classification, indicating that these layers may be non-essential. Our approach provides a systematic method for identifying less critical layers, contributing to efficient large language model architectures.

FlamePINN-1D: Physics-informed neural networks to solve forward and inverse problems of 1D laminar flames

Given the existence of various forward and inverse problems in combustion studies and applications that necessitate distinct methods for resolution, a framework to solve them in a unified way is critically needed. A promising approach is the integration of machine learning methods with governing equations of combustion systems, which exhibits superior generality and few-shot learning ability compared to purely data-driven methods. In this work, the FlamePINN-1D framework is proposed to solve the forward and inverse problems of 1D laminar flames based on physics-informed neural networks. Three cases with increasing complexity have been tested: Case 1 are freely-propagating premixed (FPP) flames with simplified physical models, while Case 2 and Case 3 are FPP and counterflow premixed (CFP) flames with detailed models, respectively. For forward problems, FlamePINN-1D aims to solve the flame fields and infer the unknown eigenvalues (such as laminar flame speeds) under the constraints of governing equations and boundary conditions. For inverse problems, FlamePINN-1D aims to reconstruct the continuous fields and infer the unknown parameters (such as transport and chemical kinetics parameters) from noisy sparse observations of the flame. Our results strongly validate these capabilities of FlamePINN-1D across various flames and working conditions. Compared to traditional methods, FlamePINN-1D is differentiable and mesh-free, exhibits no discretization errors, and is easier to implement for inverse problems. The inverse problem results also indicate the possibility of optimizing chemical mechanisms from measurements of laboratory 1D flames. Furthermore, some proposed strategies, such as hard constraints and thin-layer normalization, are proven to be essential for the robust learning of FlamePINN-1D. The code for this paper is partially available at https://github.com/CAME-THU/FlamePINN-1D.

CARNA: Characterizing Advanced heart failure Risk and hemodyNAmic phenotypes using learned multi-valued decision diagrams

Early identification of high risk heart failure (HF) patients is key to timely allocation of life-saving therapies. Hemodynamic assessments can facilitate risk stratification and enhance understanding of HF trajectories. However, risk assessment for HF is a complex, multi-faceted decision-making process that can be challenging. Previous risk models for HF do not integrate invasive hemodynamics or support missing data, and use statistical methods prone to bias or machine learning methods that are not interpretable. To address these limitations, this paper presents CARNA, a hemodynamic risk stratification and phenotyping framework for advanced HF that takes advantage of the explainability and expressivity of machine learned Multi-Valued Decision Diagrams (MVDDs). This interpretable framework learns risk scores that predict the probability of patient outcomes, and outputs descriptive patient phenotypes (sets of features and thresholds) that characterize each predicted risk score. CARNA incorporates invasive hemodynamics and can make predictions on missing data. The CARNA models were trained and validated using a total of five advanced HF patient cohorts collected from previous trials, and compared with six established HF risk scores and three traditional ML risk models. CARNA provides robust risk stratification, outperforming all previous benchmarks. Although focused on advanced HF, the CARNA framework is general purpose and can be used to learn risk stratifications for other diseases and medical applications.

Dissimilar Nodes Improve Graph Active Learning

Training labels for graph embedding algorithms could be costly to obtain in many practical scenarios. Active learning (AL) algorithms are very helpful to obtain the most useful labels for training while keeping the total number of label queries under a certain budget. The existing Active Graph Embedding framework proposes to use centrality score, density score, and entropy score to evaluate the value of unlabeled nodes, and it has been shown to be capable of bringing some improvement to the node classification tasks of Graph Convolutional Networks. However, when evaluating the importance of unlabeled nodes, it fails to consider the influence of existing labeled nodes on the value of unlabeled nodes. In other words, given the same unlabeled node, the computed informative score is always the same and is agnostic to the labeled node set. With the aim to address this limitation, in this work, we introduce 3 dissimilarity-based information scores for active learning: feature dissimilarity score (FDS), structure dissimilarity score (SDS), and embedding dissimilarity score (EDS). We find out that those three scores are able to take the influence of the labeled set on the value of unlabeled candidates into consideration, boosting our AL performance. According to experiments, our newly proposed scores boost the classification accuracy by 2.1% on average and are capable of generalizing to different Graph Neural Network architectures.

Rethinking "Batch" in BatchNorm

BatchNorm is a critical building block in modern convolutional neural networks. Its unique property of operating on "batches" instead of individual samples introduces significantly different behaviors from most other operations in deep learning. As a result, it leads to many hidden caveats that can negatively impact model's performance in subtle ways. This paper thoroughly reviews such problems in visual recognition tasks, and shows that a key to address them is to rethink different choices in the concept of "batch" in BatchNorm. By presenting these caveats and their mitigations, we hope this review can help researchers use BatchNorm more effectively.

PointRend: Image Segmentation as Rendering

We present a new method for efficient high-quality image segmentation of objects and scenes. By analogizing classical computer graphics methods for efficient rendering with over- and undersampling challenges faced in pixel labeling tasks, we develop a unique perspective of image segmentation as a rendering problem. From this vantage, we present the PointRend (Point-based Rendering) neural network module: a module that performs point-based segmentation predictions at adaptively selected locations based on an iterative subdivision algorithm. PointRend can be flexibly applied to both instance and semantic segmentation tasks by building on top of existing state-of-the-art models. While many concrete implementations of the general idea are possible, we show that a simple design already achieves excellent results. Qualitatively, PointRend outputs crisp object boundaries in regions that are over-smoothed by previous methods. Quantitatively, PointRend yields significant gains on COCO and Cityscapes, for both instance and semantic segmentation. PointRend's efficiency enables output resolutions that are otherwise impractical in terms of memory or computation compared to existing approaches.