Can Mixture-of-Experts Surpass Dense LLMs Under Strictly Equal Resources?

Mixture-of-Experts (MoE) language models dramatically expand model capacity and achieve remarkable performance without increasing per-token compute. However, can MoEs surpass dense architectures under strictly equal resource constraints - that is, when the total parameter count, training compute, and data budget are identical? This question remains under-explored despite its significant practical value and potential. In this paper, we propose a novel perspective and methodological framework to study this question thoroughly. First, we comprehensively investigate the architecture of MoEs and achieve an optimal model design that maximizes the performance. Based on this, we subsequently find that an MoE model with activation rate in an optimal region is able to outperform its dense counterpart under the same total parameter, training compute and data resource. More importantly, this optimal region remains consistent across different model sizes. Although additional amount of data turns out to be a trade-off for the enhanced performance, we show that this can be resolved via reusing data. We validate our findings through extensive experiments, training nearly 200 language models at 2B scale and over 50 at 7B scale, cumulatively processing 50 trillion tokens. All models will be released publicly.

Farseer: A Refined Scaling Law in Large Language Models

Training Large Language Models (LLMs) is prohibitively expensive, creating a critical scaling gap where insights from small-scale experiments often fail to transfer to resource-intensive production systems, thereby hindering efficient innovation. To bridge this, we introduce Farseer, a novel and refined scaling law offering enhanced predictive accuracy across scales. By systematically constructing a model loss surface $L(N,D)$, Farseer achieves a significantly better fit to empirical data than prior laws (e.g., Chinchilla's law). Our methodology yields accurate, robust, and highly generalizable predictions, demonstrating excellent extrapolation capabilities, improving upon Chinchilla's law by reducing extrapolation error by 433\%. This allows for the reliable evaluation of competing training strategies across all $(N,D)$ settings, enabling conclusions from small-scale ablation studies to be confidently extrapolated to predict large-scale performance. Furthermore, Farseer provides new insights into optimal compute allocation, better reflecting the nuanced demands of modern LLM training. To validate our approach, we trained an extensive suite of approximately 1,000 LLMs across diverse scales and configurations, consuming roughly 3 million NVIDIA H100 GPU hours. We are comprehensively open-sourcing all models, data, results, and logs at https://github.com/Farseer-Scaling-Law/Farseer to foster further research.

Score-based Generative Modeling for Conditional Independence Testing

Determining conditional independence (CI) relationships between random variables is a fundamental yet challenging task in machine learning and statistics, especially in high-dimensional settings. Existing generative model-based CI testing methods, such as those utilizing generative adversarial networks (GANs), often struggle with undesirable modeling of conditional distributions and training instability, resulting in subpar performance. To address these issues, we propose a novel CI testing method via score-based generative modeling, which achieves precise Type I error control and strong testing power. Concretely, we first employ a sliced conditional score matching scheme to accurately estimate conditional score and use Langevin dynamics conditional sampling to generate null hypothesis samples, ensuring precise Type I error control. Then, we incorporate a goodness-of-fit stage into the method to verify generated samples and enhance interpretability in practice. We theoretically establish the error bound of conditional distributions modeled by score-based generative models and prove the validity of our CI tests. Extensive experiments on both synthetic and real-world datasets show that our method significantly outperforms existing state-of-the-art methods, providing a promising way to revitalize generative model-based CI testing.

Imagination-Limited Q-Learning for Offline Reinforcement Learning

Offline reinforcement learning seeks to derive improved policies entirely from historical data but often struggles with over-optimistic value estimates for out-of-distribution (OOD) actions. This issue is typically mitigated via policy constraint or conservative value regularization methods. However, these approaches may impose overly constraints or biased value estimates, potentially limiting performance improvements. To balance exploitation and restriction, we propose an Imagination-Limited Q-learning (ILQ) method, which aims to maintain the optimism that OOD actions deserve within appropriate limits. Specifically, we utilize the dynamics model to imagine OOD action-values, and then clip the imagined values with the maximum behavior values. Such design maintains reasonable evaluation of OOD actions to the furthest extent, while avoiding its over-optimism. Theoretically, we prove the convergence of the proposed ILQ under tabular Markov decision processes. Particularly, we demonstrate that the error bound between estimated values and optimality values of OOD state-actions possesses the same magnitude as that of in-distribution ones, thereby indicating that the bias in value estimates is effectively mitigated. Empirically, our method achieves state-of-the-art performance on a wide range of tasks in the D4RL benchmark.

Is Compression Really Linear with Code Intelligence?

Understanding the relationship between data compression and the capabilities of Large Language Models (LLMs) is crucial, especially in specialized domains like code intelligence. Prior work posited a linear relationship between compression and general intelligence. However, it overlooked the multifaceted nature of code that encompasses diverse programming languages and tasks, and struggled with fair evaluation of modern Code LLMs. We address this by evaluating a diverse array of open-source Code LLMs on comprehensive multi-language, multi-task code benchmarks. To address the challenge of efficient and fair evaluation of pre-trained LLMs' code intelligence, we introduce \textit{Format Annealing}, a lightweight, transparent training methodology designed to assess the intrinsic capabilities of these pre-trained models equitably. Compression efficacy, measured as bits-per-character (BPC), is determined using a novel, large-scale, and previously unseen code validation set derived from GitHub. Our empirical results reveal a fundamental logarithmic relationship between measured code intelligence and BPC. This finding refines prior hypotheses of linearity, which we suggest are likely observations of the logarithmic curve's tail under specific, limited conditions. Our work provides a more nuanced understanding of compression's role in developing code intelligence and contributes a robust evaluation framework in the code domain.

Data Synthesis with Diverse Styles for Face Recognition via 3DMM-Guided Diffusion

Identity-preserving face synthesis aims to generate synthetic face images of virtual subjects that can substitute real-world data for training face recognition models. While prior arts strive to create images with consistent identities and diverse styles, they face a trade-off between them. Identifying their limitation of treating style variation as subject-agnostic and observing that real-world persons actually have distinct, subject-specific styles, this paper introduces MorphFace, a diffusion-based face generator. The generator learns fine-grained facial styles, e.g., shape, pose and expression, from the renderings of a 3D morphable model (3DMM). It also learns identities from an off-the-shelf recognition model. To create virtual faces, the generator is conditioned on novel identities of unlabeled synthetic faces, and novel styles that are statistically sampled from a real-world prior distribution. The sampling especially accounts for both intra-subject variation and subject distinctiveness. A context blending strategy is employed to enhance the generator's responsiveness to identity and style conditions. Extensive experiments show that MorphFace outperforms the best prior arts in face recognition efficacy.

Effective Cloud Removal for Remote Sensing Images by an Improved Mean-Reverting Denoising Model with Elucidated Design Space

Cloud removal (CR) remains a challenging task in remote sensing image processing. Although diffusion models (DM) exhibit strong generative capabilities, their direct applications to CR are suboptimal, as they generate cloudless images from random noise, ignoring inherent information in cloudy inputs. To overcome this drawback, we develop a new CR model EMRDM based on mean-reverting diffusion models (MRDMs) to establish a direct diffusion process between cloudy and cloudless images. Compared to current MRDMs, EMRDM offers a modular framework with updatable modules and an elucidated design space, based on a reformulated forward process and a new ordinary differential equation (ODE)-based backward process. Leveraging our framework, we redesign key MRDM modules to boost CR performance, including restructuring the denoiser via a preconditioning technique, reorganizing the training process, and improving the sampling process by introducing deterministic and stochastic samplers. To achieve multi-temporal CR, we further develop a denoising network for simultaneously denoising sequential images. Experiments on mono-temporal and multi-temporal datasets demonstrate the superior performance of EMRDM. Our code is available at https://github.com/Ly403/EMRDM.

Predictable Scale: Part I -- Optimal Hyperparameter Scaling Law in Large Language Model Pretraining

The impressive capabilities of Large Language Models (LLMs) across diverse tasks are now well-established, yet their effective deployment necessitates careful hyperparameter optimization. Through extensive empirical studies involving grid searches across diverse configurations, we discover universal scaling laws governing these hyperparameters: optimal learning rate follows a power-law relationship with both model parameters and data sizes, while optimal batch size scales primarily with data sizes. Our analysis reveals a convex optimization landscape for hyperparameters under fixed models and data size conditions. This convexity implies an optimal hyperparameter plateau. We contribute a universal, plug-and-play optimal hyperparameter tool for the community. Its estimated values on the test set are merely 0.07\% away from the globally optimal LLM performance found via an exhaustive search. These laws demonstrate remarkable robustness across variations in model sparsity, training data distribution, and model shape. To our best known, this is the first work that unifies different model shapes and structures, such as Mixture-of-Experts models and dense transformers, as well as establishes optimal hyperparameter scaling laws across diverse data distributions. This exhaustive optimization process demands substantial computational resources, utilizing nearly one million NVIDIA H800 GPU hours to train 3,700 LLMs of varying sizes and hyperparameters from scratch and consuming approximately 100 trillion tokens in total. To facilitate reproducibility and further research, we will progressively release all loss measurements and model checkpoints through our designated repository https://step-law.github.io/

UIFace: Unleashing Inherent Model Capabilities to Enhance Intra-Class Diversity in Synthetic Face Recognition

Face recognition (FR) stands as one of the most crucial applications in computer vision. The accuracy of FR models has significantly improved in recent years due to the availability of large-scale human face datasets. However, directly using these datasets can inevitably lead to privacy and legal problems. Generating synthetic data to train FR models is a feasible solution to circumvent these issues. While existing synthetic-based face recognition methods have made significant progress in generating identity-preserving images, they are severely plagued by context overfitting, resulting in a lack of intra-class diversity of generated images and poor face recognition performance. In this paper, we propose a framework to Unleash Inherent capability of the model to enhance intra-class diversity for synthetic face recognition, shortened as UIFace. Our framework first trains a diffusion model that can perform sampling conditioned on either identity contexts or a learnable empty context. The former generates identity-preserving images but lacks variations, while the latter exploits the model's intrinsic ability to synthesize intra-class-diversified images but with random identities. Then we adopt a novel two-stage sampling strategy during inference to fully leverage the strengths of both types of contexts, resulting in images that are diverse as well as identitypreserving. Moreover, an attention injection module is introduced to further augment the intra-class variations by utilizing attention maps from the empty context to guide the sampling process in ID-conditioned generation. Experiments show that our method significantly surpasses previous approaches with even less training data and half the size of synthetic dataset. The proposed UIFace even achieves comparable performance with FR models trained on real datasets when we further increase the number of synthetic identities.

Generative Regression Based Watch Time Prediction for Video Recommendation: Model and Performance

Watch time prediction (WTP) has emerged as a pivotal task in short video recommendation systems, designed to encapsulate user interests. Predicting users' watch times on videos often encounters challenges, including wide value ranges and imbalanced data distributions, which can lead to significant bias when directly regressing watch time. Recent studies have tried to tackle these issues by converting the continuous watch time estimation into an ordinal classification task. While these methods are somewhat effective, they exhibit notable limitations. Inspired by language modeling, we propose a novel Generative Regression (GR) paradigm for WTP based on sequence generation. This approach employs structural discretization to enable the lossless reconstruction of original values while maintaining prediction fidelity. By formulating the prediction problem as a numerical-to-sequence mapping, and with meticulously designed vocabulary and label encodings, each watch time is transformed into a sequence of tokens. To expedite model training, we introduce the curriculum learning with an embedding mixup strategy which can mitigate training-and-inference inconsistency associated with teacher forcing. We evaluate our method against state-of-the-art approaches on four public datasets and one industrial dataset. We also perform online A/B testing on Kuaishou, a leading video app with about 400 million DAUs, to demonstrate the real-world efficacy of our method. The results conclusively show that GR outperforms existing techniques significantly. Furthermore, we successfully apply GR to another regression task in recommendation systems, i.e., Lifetime Value (LTV) prediction, which highlights its potential as a novel and effective solution to general regression challenges.