Rio Yokota

Le Bronnec Florian, Alexandre Verine, Rio Yokota et al.

LLMs for code generation are commonly evaluated in repeated-sampling settings using Pass@k, where multiple candidate programs are executed against unit tests under a finite sampling budget. While recent verifier-based reinforcement learning (RLVR) methods improve executable correctness, how these objectives affect redundancy among sampled programs remains poorly understood. In this work, we study implementation-level redundancy in code generation using JPlag, a plagiarism-detection system for code. Across models and benchmarks, we show that correctness-only RLVR often concentrates generations around repeated implementations, whereas Pass@k-aware objectives maintain lower redundancy and improve larger-budget performance. Motivated by these observations, we augment RLVR with direct anti-redundancy rewards based on JPlag similarity. Across 3 models and 3 benchmarks, discouraging near-duplicate generations reliably improves finite-budget executable performance, often matching or outperforming specialized Pass@k-aware objectives.

LLM-jp, Akiko Aizawa, Eiji Aramaki et al.

This paper introduces LLM-jp, a cross-organizational project for the research and development of Japanese large language models (LLMs). LLM-jp aims to develop open-source and strong Japanese LLMs, and as of this writing, more than 1,500 participants from academia and industry are working together for this purpose. This paper presents the background of the establishment of LLM-jp, summaries of its activities, and technical reports on the LLMs developed by LLM-jp. For the latest activities, visit https://llm-jp.nii.ac.jp/en/.

Risa Shinoda, Ryo Hayamizu, Kodai Nakashima et al.

Pre-training is a strong strategy for enhancing visual models to efficiently train them with a limited number of labeled images. In semantic segmentation, creating annotation masks requires an intensive amount of labor and time, and therefore, a large-scale pre-training dataset with semantic labels is quite difficult to construct. Moreover, what matters in semantic segmentation pre-training has not been fully investigated. In this paper, we propose the Segmentation Radial Contour DataBase (SegRCDB), which for the first time applies formula-driven supervised learning for semantic segmentation. SegRCDB enables pre-training for semantic segmentation without real images or any manual semantic labels. SegRCDB is based on insights about what is important in pre-training for semantic segmentation and allows efficient pre-training. Pre-training with SegRCDB achieved higher mIoU than the pre-training with COCO-Stuff for fine-tuning on ADE-20k and Cityscapes with the same number of training images. SegRCDB has a high potential to contribute to semantic segmentation pre-training and investigation by enabling the creation of large datasets without manual annotation. The SegRCDB dataset will be released under a license that allows research and commercial use. Code is available at: https://github.com/dahlian00/SegRCDB

Rio Yokota, Huda Ibeid, David Keyes

There has been a large increase in the amount of work on hierarchical low-rank approximation methods, where the interest is shared by multiple communities that previously did not intersect. This objective of this article is two-fold; to provide a thorough review of the recent advancements in this field from both analytical and algebraic perspectives, and to present a comparative benchmark of two highly optimized implementations of contrasting methods for some simple yet representative test cases. We categorize the recent advances in this field from the perspective of compute-memory tradeoff, which has not been considered in much detail in this area. Benchmark tests reveal that there is a large difference in the memory consumption and performance between the different methods.

Huda Ibeid, Rio Yokota, Jennifer Pestana et al.

Among optimal hierarchical algorithms for the computational solution of elliptic problems, the Fast Multipole Method (FMM) stands out for its adaptability to emerging architectures, having high arithmetic intensity, tunable accuracy, and relaxable global synchronization requirements. We demonstrate that, beyond its traditional use as a solver in problems for which explicit free-space kernel representations are available, the FMM has applicability as a preconditioner in finite domain elliptic boundary value problems, by equipping it with boundary integral capability for satisfying conditions at finite boundaries and by wrapping it in a Krylov method for extensibility to more general operators. Here, we do not discuss the well developed applications of FMM to implement matrix-vector multiplications within Krylov solvers of boundary element methods. Instead, we propose using FMM for the volume-to-volume contribution of inhomogeneous Poisson-like problems, where the boundary integral is a small part of the overall computation. Our method may be used to precondition sparse matrices arising from finite difference/element discretizations, and can handle a broader range of scientific applications. Compared with multigrid methods, it is capable of comparable algebraic convergence rates down to the truncation error of the discretized PDE, and it offers potentially superior multicore and distributed memory scalability properties on commodity architecture supercomputers. Compared with other methods exploiting the low rank character of off-diagonal blocks of the dense resolvent operator, FMM-preconditioned Krylov iteration may reduce the amount of communication because it is matrix-free and exploits the tree structure of FMM. We describe our tests in reproducible detail with freely available codes and outline directions for further extensibility.

Hatem Ltaief, Rio Yokota

Fast multipole methods have O(N) complexity, are compute bound, and require very little synchronization, which makes them a favorable algorithm on next-generation supercomputers. Their most common application is to accelerate N-body problems, but they can also be used to solve boundary integral equations. When the particle distribution is irregular and the tree structure is adaptive, load-balancing becomes a non-trivial question. A common strategy for load-balancing FMMs is to use the work load from the previous step as weights to statically repartition the next step. The authors discuss in the paper another approach based on data-driven execution to efficiently tackle this challenging load-balancing problem. The core idea consists of breaking the most time-consuming stages of the FMMs into smaller tasks. The algorithm can then be represented as a Directed Acyclic Graph (DAG) where nodes represent tasks, and edges represent dependencies among them. The execution of the algorithm is performed by asynchronously scheduling the tasks using the QUARK runtime environment, in a way such that data dependencies are not violated for numerical correctness purposes. This asynchronous scheduling results in an out-of-order execution. The performance results of the data-driven FMM execution outperform the previous strategy and show linear speedup on a quad-socket quad-core Intel Xeon system.

Rio Yokota, Lorena Barba

Among the algorithms that are likely to play a major role in future exascale computing, the fast multipole method (FMM) appears as a rising star. Our previous recent work showed scaling of an FMM on GPU clusters, with problem sizes in the order of billions of unknowns. That work led to an extremely parallel FMM, scaling to thousands of GPUs or tens of thousands of CPUs. This paper reports on a a campaign of performance tuning and scalability studies using multi-core CPUs, on the Kraken supercomputer. All kernels in the FMM were parallelized using OpenMP, and a test using 10^7 particles randomly distributed in a cube showed 78% efficiency on 8 threads. Tuning of the particle-to-particle kernel using SIMD instructions resulted in 4x speed-up of the overall algorithm on single-core tests with 10^3 - 10^7 particles. Parallel scalability was studied in both strong and weak scaling. The strong scaling test used 10^8 particles and resulted in 93% parallel efficiency on 2048 processes for the non-SIMD code and 54% for the SIMD-optimized code (which was still 2x faster). The weak scaling test used 10^6 particles per process, and resulted in 72% efficiency on 32,768 processes, with the largest calculation taking about 40 seconds to evaluate more than 32 billion unknowns. This work builds up evidence for our view that FMM is poised to play a leading role in exascale computing, and we end the paper with a discussion of the features that make it a particularly favorable algorithm for the emerging heterogeneous and massively parallel architectural landscape.

Rio Yokota

The present work attempts to integrate the independent efforts in the fast N-body community to create the fastest N-body library for many-core and heterogenous architectures. Focus is placed on low accuracy optimizations, in response to the recent interest to use FMM as a preconditioner for sparse linear solvers. A direct comparison with other state-of-the-art fast N-body codes demonstrates that orders of magnitude increase in performance can be achieved by careful selection of the optimal algorithm and low-level optimization of the code. The current N-body solver uses a fast multipole method with an efficient strategy for finding the list of cell-cell interactions by a dual tree traversal. A task-based threading model is used to maximize thread-level parallelism and intra-node load-balancing. In order to extract the full potential of the SIMD units on the latest CPUs, the inner kernels are optimized using AVX instructions. Our code -- exaFMM -- is an order of magnitude faster than the current state-of-the-art FMM codes, which are themselves an order of magnitude faster than the average FMM code.

Rio Yokota, Lorena A. Barba

With the current hybridization of treecodes and FMMs, combined with auto-tuning capabilities on heterogeneous architectures, the flexibility of fast N-body methods has been greatly enhanced. These features are a requirement to developing a black-box software library for fast N-body algorithms on heterogeneous systems, which is our immediate goal.

Rio Yokota, L. A. Barba

The Lagrangian vortex method offers an alternative numerical approach for direct numerical simulation of turbulence. The fact that it uses the fast multipole method (FMM)--a hierarchical algorithm for N-body problems with highly scalable parallel implementations--as numerical engine makes it a potentially good candidate for exascale systems. However, there have been few validation studies of Lagrangian vortex simulations and the insufficient comparisons against standard DNS codes has left ample room for skepticism. This paper presents a comparison between a Lagrangian vortex method and a pseudo-spectral method for the simulation of decaying homogeneous isotropic turbulence. This flow field is chosen despite the fact that it is not the most favorable flow problem for particle methods (which shine in wake flows or where vorticity is compact), due to the fact that it is ideal for the quantitative validation of DNS codes. We use a 256^3 grid with Re_lambda=50 and 100 and look at the turbulence statistics, including high-order moments. The focus is on the effect of the various parameters in the vortex method, e.g., order of FMM series expansion, frequency of reinitialization, overlap ratio and time step. The vortex method uses an FMM code (exaFMM) that runs on GPU hardware using CUDA, while the spectral code (hit3d) runs on CPU only. Results indicate that, for this application (and with the current code implementations), the spectral method is an order of magnitude faster than the vortex method when using a single GPU for the FMM and six CPU cores for the FFT.

Hirokatsu Kataoka, Sora Takashima, Ryo Hayamizu et al.

In the present work, we show that the performance of formula-driven supervised learning (FDSL) can match or even exceed that of ImageNet-21k and can approach that of the JFT-300M dataset without the use of real images, human supervision, or self-supervision during the pre-training of vision transformers (ViTs). For example, ViT-Base pre-trained on ImageNet-21k and JFT-300M showed 83.0 and 84.1% top-1 accuracy when fine-tuned on ImageNet-1k, and FDSL showed 83.8% top-1 accuracy when pre-trained under comparable conditions (hyperparameters and number of epochs). Especially, the ExFractalDB-21k pre-training was calculated with x14.2 fewer images compared with JFT-300M. Images generated by formulas avoid privacy and copyright issues, labeling costs and errors, and biases that real images suffer from, and thus have tremendous potential for pre-training general models. To understand the performance of the synthetic images, we tested two hypotheses, namely (i) object contours are what matter in FDSL datasets and (ii) an increased number of parameters for label creation improves performance in FDSL pre-training. To test the former hypothesis, we constructed a dataset that consisted of simple object contour combinations. We found that this dataset matched the performance of fractal databases. For the latter hypothesis, we found that increasing the difficulty of the pre-training task generally leads to better fine-tuning accuracy.

Sora Takashima, Ryo Hayamizu, Nakamasa Inoue et al.

Formula-driven supervised learning (FDSL) has been shown to be an effective method for pre-training vision transformers, where ExFractalDB-21k was shown to exceed the pre-training effect of ImageNet-21k. These studies also indicate that contours mattered more than textures when pre-training vision transformers. However, the lack of a systematic investigation as to why these contour-oriented synthetic datasets can achieve the same accuracy as real datasets leaves much room for skepticism. In the present work, we develop a novel methodology based on circular harmonics for systematically investigating the design space of contour-oriented synthetic datasets. This allows us to efficiently search the optimal range of FDSL parameters and maximize the variety of synthetic images in the dataset, which we found to be a critical factor. When the resulting new dataset VisualAtom-21k is used for pre-training ViT-Base, the top-1 accuracy reached 83.7% when fine-tuning on ImageNet-1k. This is close to the top-1 accuracy (84.2%) achieved by JFT-300M pre-training, while the number of images is 1/14. Unlike JFT-300M which is a static dataset, the quality of synthetic datasets will continue to improve, and the current work is a testament to this possibility. FDSL is also free of the common issues associated with real images, e.g. privacy/copyright issues, labeling costs/errors, and ethical biases.

Ryo Nakamura, Ryu Tadokoro, Ryosuke Yamada et al.

Pre-training and transfer learning are an important building block of current computer vision systems. While pre-training is usually performed on large real-world image datasets, in this paper we ask whether this is truly necessary. To this end, we search for a minimal, purely synthetic pre-training dataset that allows us to achieve performance similar to the 1 million images of ImageNet-1k. We construct such a dataset from a single fractal with perturbations. With this, we contribute three main findings. (i) We show that pre-training is effective even with minimal synthetic images, with performance on par with large-scale pre-training datasets like ImageNet-1k for full fine-tuning. (ii) We investigate the single parameter with which we construct artificial categories for our dataset. We find that while the shape differences can be indistinguishable to humans, they are crucial for obtaining strong performances. (iii) Finally, we investigate the minimal requirements for successful pre-training. Surprisingly, we find that a substantial reduction of synthetic images from 1k to 1 can even lead to an increase in pre-training performance, a motivation to further investigate ''scaling backwards''. Finally, we extend our method from synthetic images to real images to see if a single real image can show similar pre-training effect through shape augmentation. We find that the use of grayscale images and affine transformations allows even real images to ''scale backwards''.

Ryo Nakamura, Hirokatsu Kataoka, Sora Takashima et al.

Formula-driven supervised learning (FDSL) is a pre-training method that relies on synthetic images generated from mathematical formulae such as fractals. Prior work on FDSL has shown that pre-training vision transformers on such synthetic datasets can yield competitive accuracy on a wide range of downstream tasks. These synthetic images are categorized according to the parameters in the mathematical formula that generate them. In the present work, we hypothesize that the process for generating different instances for the same category in FDSL, can be viewed as a form of data augmentation. We validate this hypothesis by replacing the instances with data augmentation, which means we only need a single image per category. Our experiments shows that this one-instance fractal database (OFDB) performs better than the original dataset where instances were explicitly generated. We further scale up OFDB to 21,000 categories and show that it matches, or even surpasses, the model pre-trained on ImageNet-21k in ImageNet-1k fine-tuning. The number of images in OFDB is 21k, whereas ImageNet-21k has 14M. This opens new possibilities for pre-training vision transformers with much smaller datasets.

Hiroki Naganuma, Kartik Ahuja, Shiro Takagi et al.

Modern deep learning systems do not generalize well when the test data distribution is slightly different to the training data distribution. While much promising work has been accomplished to address this fragility, a systematic study of the role of optimizers and their out-of-distribution generalization performance has not been undertaken. In this study, we examine the performance of popular first-order optimizers for different classes of distributional shift under empirical risk minimization and invariant risk minimization. We address this question for image and text classification using DomainBed, WILDS, and Backgrounds Challenge as testbeds for studying different types of shifts -- namely correlation and diversity shift. We search over a wide range of hyperparameters and examine classification accuracy (in-distribution and out-of-distribution) for over 20,000 models. We arrive at the following findings, which we expect to be helpful for practitioners: i) adaptive optimizers (e.g., Adam) perform worse than non-adaptive optimizers (e.g., SGD, momentum SGD) on out-of-distribution performance. In particular, even though there is no significant difference in in-distribution performance, we show a measurable difference in out-of-distribution performance. ii) in-distribution performance and out-of-distribution performance exhibit three types of behavior depending on the dataset -- linear returns, increasing returns, and diminishing returns. For example, in the training of natural language data using Adam, fine-tuning the performance of in-distribution performance does not significantly contribute to the out-of-distribution generalization performance.

Go Ohtani, Ryu Tadokoro, Ryosuke Yamada et al.

In this work, we investigate the understudied effect of the training data used for image super-resolution (SR). Most commonly, novel SR methods are developed and benchmarked on common training datasets such as DIV2K and DF2K. However, we investigate and rethink the training data from the perspectives of diversity and quality, {thereby addressing the question of ``How important is SR training for SR models?''}. To this end, we propose an automated image evaluation pipeline. With this, we stratify existing high-resolution image datasets and larger-scale image datasets such as ImageNet and PASS to compare their performances. We find that datasets with (i) low compression artifacts, (ii) high within-image diversity as judged by the number of different objects, and (iii) a large number of images from ImageNet or PASS all positively affect SR performance. We hope that the proposed simple-yet-effective dataset curation pipeline will inform the construction of SR datasets in the future and yield overall better models.

Aoyu Li, Ikuro Sato, Kohta Ishikawa et al.

Among various supervised deep metric learning methods proxy-based approaches have achieved high retrieval accuracies. Proxies, which are class-representative points in an embedding space, receive updates based on proxy-sample similarities in a similar manner to sample representations. In existing methods, a relatively small number of samples can produce large gradient magnitudes (ie, hard samples), and a relatively large number of samples can produce small gradient magnitudes (ie, easy samples); these can play a major part in updates. Assuming that acquiring too much sensitivity to such extreme sets of samples would deteriorate the generalizability of a method, we propose a novel proxy-based method called Informative Sample-Aware Proxy (Proxy-ISA), which directly modifies a gradient weighting factor for each sample using a scheduled threshold function, so that the model is more sensitive to the informative samples. Extensive experiments on the CUB-200-2011, Cars-196, Stanford Online Products and In-shop Clothes Retrieval datasets demonstrate the superiority of Proxy-ISA compared with the state-of-the-art methods.

Ryosuke Yamada, Kensho Hara, Hirokatsu Kataoka et al.

Throughout the history of computer vision, while research has explored the integration of images (visual) and point clouds (geometric), many advancements in image and 3D object recognition have tended to process these modalities separately. We aim to bridge this divide by integrating images and point clouds on a unified transformer model. This approach integrates the modality-specific properties of images and point clouds and achieves fundamental downstream tasks in image and 3D object recognition on a unified transformer model by learning visual-geometric representations. In this work, we introduce Formula-Supervised Visual-Geometric Pre-training (FSVGP), a novel synthetic pre-training method that automatically generates aligned synthetic images and point clouds from mathematical formulas. Through cross-modality supervision, we enable supervised pre-training between visual and geometric modalities. FSVGP also reduces reliance on real data collection, cross-modality alignment, and human annotation. Our experimental results show that FSVGP pre-trains more effectively than VisualAtom and PC-FractalDB across six tasks: image and 3D object classification, detection, and segmentation. These achievements demonstrate FSVGP's superior generalization in image and 3D object recognition and underscore the potential of synthetic pre-training in visual-geometric representation learning. Our project website is available at https://ryosuke-yamada.github.io/fdsl-fsvgp/.

Hiro Ishii, Kenta Niwa, Hiroshi Sawada et al.

We propose Federated Preconditioned Mixing (FedPM), a novel Federated Learning (FL) method that leverages second-order optimization. Prior methods--such as LocalNewton, LTDA, and FedSophia--have incorporated second-order optimization in FL by performing iterative local updates on clients and applying simple mixing of local parameters on the server. However, these methods often suffer from drift in local preconditioners, which significantly disrupts the convergence of parameter training, particularly in heterogeneous data settings. To overcome this issue, we refine the update rules by decomposing the ideal second-order update--computed using globally preconditioned global gradients--into parameter mixing on the server and local parameter updates on clients. As a result, our FedPM introduces preconditioned mixing of local parameters on the server, effectively mitigating drift in local preconditioners. We provide a theoretical convergence analysis demonstrating a superlinear rate for strongly convex objectives in scenarios involving a single local update. To demonstrate the practical benefits of FedPM, we conducted extensive experiments. The results showed significant improvements with FedPM in the test accuracy compared to conventional methods incorporating simple mixing, fully leveraging the potential of second-order optimization.

Taishi Nakamura, Mayank Mishra, Simone Tedeschi et al. · ibm-research, stanford

Pretrained language models are an integral part of AI applications, but their high computational cost for training limits accessibility. Initiatives such as Bloom and StarCoder aim to democratize access to pretrained models for collaborative community development. Despite these efforts, such models encounter challenges such as limited multilingual capabilities, risks of catastrophic forgetting during continual pretraining, and the high costs of training models from scratch, alongside the need to align with AI safety standards and regulatory frameworks. This paper presents Aurora-M, a 15B parameter multilingual open-source model trained on English, Finnish, Hindi, Japanese, Vietnamese, and code. Continually pretrained from StarCoderPlus on 435B additional tokens, Aurora-M surpasses 2T tokens in total training token count. It is the first open-source multilingual model fine-tuned on human-reviewed safety instructions, thus aligning its development not only with conventional red-teaming considerations, but also with the specific concerns articulated in the Biden-Harris Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence. We evaluate Aurora-M across a wide range of tasks and languages, showcasing its robustness against catastrophic forgetting and its superior performance in multilingual settings, particularly in safety evaluations. We open-source Aurora-M and its variants to encourage responsible open-source development of large language models at https://huggingface.co/aurora-m.

Bai Cong, Nico Daheim, Yuesong Shen et al.

We show that variational learning can significantly improve the accuracy and calibration of Low-Rank Adaptation (LoRA) without a substantial increase in the cost. We replace AdamW by the Improved Variational Online Newton (IVON) algorithm to finetune large language models. For Llama-2 with 7 billion parameters, IVON improves the accuracy over AdamW by 2.8% and expected calibration error by 4.6%. The accuracy is also better than the other Bayesian alternatives, yet the cost is lower and the implementation is easier. Our work provides additional evidence for the effectiveness of IVON for large language models. The code is available at https://github.com/team-approx-bayes/ivon-lora.

Hiroki Naganuma, Taiji Suzuki, Rio Yokota et al.

Generalization measures have been studied extensively in the machine learning community to better characterize generalization gaps. However, establishing a reliable generalization measure for statistically singular models such as deep neural networks (DNNs) is difficult due to their complex nature. This study focuses on Takeuchi's information criterion (TIC) to investigate the conditions under which this classical measure can effectively explain the generalization gaps of DNNs. Importantly, the developed theory indicates the applicability of TIC near the neural tangent kernel (NTK) regime. In a series of experiments, we trained more than 5,000 DNN models with 12 architectures, including large models (e.g., VGG-16), on four datasets, and estimated the corresponding TIC values to examine the relationship between the generalization gap and the TIC estimates. We applied several TIC approximation methods with feasible computational costs and assessed the accuracy trade-off. Our experimental results indicate that the estimated TIC values correlate well with the generalization gap under conditions close to the NTK regime. However, we show both theoretically and empirically that outside the NTK regime such correlation disappears. Finally, we demonstrate that TIC provides better trial pruning ability than existing methods for hyperparameter optimization.

Daisuke Nohara, Taishi Nakamura, Rio Yokota

Reinforcement learning substantially improves reasoning in large language models, but it also tends to lengthen chain of thought outputs and increase computational cost during both training and inference. Though length control methods have been proposed, it remains unclear what the optimal output length is for balancing efficiency and performance. In this work, we compare several length control methods on two models, Qwen3-1.7B Base and DeepSeek-R1-Distill-Qwen-1.5B. Our results indicate that length penalties may hinder reasoning acquisition, while properly tuned length control can improve efficiency for models with strong prior reasoning. By extending prior work to RL trained policies, we identify two failure modes, 1) long outputs increase dispersion, and 2) short outputs lead to under-thinking.

Kazuki Fujii, Taishi Nakamura, Mengsay Loem et al.

Cross-lingual continual pre-training of large language models (LLMs) initially trained on English corpus allows us to leverage the vast amount of English language resources and reduce the pre-training cost. In this study, we constructed Swallow, an LLM with enhanced Japanese capability, by extending the vocabulary of Llama 2 to include Japanese characters and conducting continual pre-training on a large Japanese web corpus. Experimental results confirmed that the performance on Japanese tasks drastically improved through continual pre-training, and the performance monotonically increased with the amount of training data up to 100B tokens. Consequently, Swallow achieved superior performance compared to other LLMs that were trained from scratch in English and Japanese. An analysis of the effects of continual pre-training revealed that it was particularly effective for Japanese question answering tasks. Furthermore, to elucidate effective methodologies for cross-lingual continual pre-training from English to Japanese, we investigated the impact of vocabulary expansion and the effectiveness of incorporating parallel corpora. The results showed that the efficiency gained through vocabulary expansion had no negative impact on performance, except for the summarization task, and that the combined use of parallel corpora enhanced translation ability.

Taishi Nakamura, Satoki Ishikawa, Masaki Kawamura et al.

Empirical scaling laws have driven the evolution of large language models (LLMs), yet their coefficients shift whenever the model architecture or data pipeline changes. Mixture-of-Experts (MoE) models, now standard in state-of-the-art systems, introduce a new sparsity dimension that current dense-model frontiers overlook. We investigate how MoE sparsity influences two distinct capability regimes: memorization skills and reasoning skills. By training MoE families that vary total parameters, active parameters, and top-$k$ routing under fixed compute budgets, we disentangle pre-training loss from downstream accuracy. Our results reveal two principles. First, Active FLOPs: models with identical training loss but greater active compute achieve higher reasoning accuracy. Second, Total tokens per parameter (TPP): memorization tasks improve with more parameters, while reasoning tasks benefit from optimal TPP, indicating that reasoning is data-hungry. Neither reinforcement learning post-training (GRPO) nor increased test-time compute alters these trends. We therefore argue that optimal MoE sparsity must be determined jointly by active FLOPs and TPP, revising the classical picture of compute-optimal scaling. Our model checkpoints, code and logs are open-source at https://github.com/rioyokotalab/optimal-sparsity.

Shun Iwase, Xingyu Liu, Rawal Khirodkar et al.

We present RePOSE, a fast iterative refinement method for 6D object pose estimation. Prior methods perform refinement by feeding zoomed-in input and rendered RGB images into a CNN and directly regressing an update of a refined pose. Their runtime is slow due to the computational cost of CNN, which is especially prominent in multiple-object pose refinement. To overcome this problem, RePOSE leverages image rendering for fast feature extraction using a 3D model with a learnable texture. We call this deep texture rendering, which uses a shallow multi-layer perceptron to directly regress a view-invariant image representation of an object. Furthermore, we utilize differentiable Levenberg-Marquardt (LM) optimization to refine a pose fast and accurately by minimizing the feature-metric error between the input and rendered image representations without the need of zooming in. These image representations are trained such that differentiable LM optimization converges within few iterations. Consequently, RePOSE runs at 92 FPS and achieves state-of-the-art accuracy of 51.6% on the Occlusion LineMOD dataset - a 4.1% absolute improvement over the prior art, and comparable result on the YCB-Video dataset with a much faster runtime. The code is available at https://github.com/sh8/repose.

Kazuki Osawa, Siddharth Swaroop, Anirudh Jain et al.

Bayesian methods promise to fix many shortcomings of deep learning, but they are impractical and rarely match the performance of standard methods, let alone improve them. In this paper, we demonstrate practical training of deep networks with natural-gradient variational inference. By applying techniques such as batch normalisation, data augmentation, and distributed training, we achieve similar performance in about the same number of epochs as the Adam optimiser, even on large datasets such as ImageNet. Importantly, the benefits of Bayesian principles are preserved: predictive probabilities are well-calibrated, uncertainties on out-of-distribution data are improved, and continual-learning performance is boosted. This work enables practical deep learning while preserving benefits of Bayesian principles. A PyTorch implementation is available as a plug-and-play optimiser.

Rio Yokota, L. A. Barba, Matthew G. Knepley

We have developed a parallel algorithm for radial basis function (RBF) interpolation that exhibits O(N) complexity,requires O(N) storage, and scales excellently up to a thousand processes. The algorithm uses a GMRES iterative solver with a restricted additive Schwarz method (RASM) as a preconditioner and a fast matrix-vector algorithm. Previous fast RBF methods, --,achieving at most O(NlogN) complexity,--, were developed using multiquadric and polyharmonic basis functions. In contrast, the present method uses Gaussians with a small variance (a common choice in particle methods for fluid simulation, our main target application). The fast decay of the Gaussian basis function allows rapid convergence of the iterative solver even when the subdomains in the RASM are very small. The present method was implemented in parallel using the PETSc library (developer version). Numerical experiments demonstrate its capability in problems of RBF interpolation with more than 50 million data points, timing at 106 seconds (19 iterations for an error tolerance of 10^-15 on 1024 processors of a Blue Gene/L (700 MHz PowerPC processors). The parallel code is freely available in the open-source model.

Qi Sun, Stefan Nielsen, Rio Yokota et al.

Large Language Models cannot reliably acquire new knowledge post-deployment -- even when relevant text resources exist, models fail to transform them into actionable knowledge without retraining. Retrieval-Augmented Generation attempts to bridge this gap by surfacing relevant documents at inference time, yet similarity-based retrieval often fails to identify context that actually improves task performance. We introduce Evolutionary Context Search (ECS), an evolutionary method that searches context combinations using accuracy on a small development set, requiring only inference calls without weight updates. ECS moves beyond semantic similarity to discover non-obvious context pairings that significantly boost performance. Our empirical results show that ECS improves BackendBench by 27\% and $τ$-bench airline by 7\%. The evolved contexts are model-agnostic, as those evolved with Gemini-3-Flash transfer effectively to Claude Sonnet and DeepSeek. This suggests that ECS opens a path toward automated context discovery for skill acquisition -- an efficient alternative to manual prompt engineering or costly fine-tuning.

Yuesong Shen, Nico Daheim, Bai Cong et al.

We give extensive empirical evidence against the common belief that variational learning is ineffective for large neural networks. We show that an optimizer called Improved Variational Online Newton (IVON) consistently matches or outperforms Adam for training large networks such as GPT-2 and ResNets from scratch. IVON's computational costs are nearly identical to Adam but its predictive uncertainty is better. We show several new use cases of IVON where we improve finetuning and model merging in Large Language Models, accurately predict generalization error, and faithfully estimate sensitivity to data. We find overwhelming evidence that variational learning is effective.

Naoaki Okazaki, Kakeru Hattori, Hirai Shota et al.

Open Japanese large language models (LLMs) have been trained on the Japanese portions of corpora such as CC-100, mC4, and OSCAR. However, these corpora were not created for the quality of Japanese texts. This study builds a large Japanese web corpus by extracting and refining text from the Common Crawl archive (21 snapshots of approximately 63.4 billion pages crawled between 2020 and 2023). This corpus consists of approximately 312.1 billion characters (approximately 173 million pages), which is the largest of all available training corpora for Japanese LLMs, surpassing CC-100 (approximately 25.8 billion characters), mC4 (approximately 239.7 billion characters) and OSCAR 23.10 (approximately 74 billion characters). To confirm the quality of the corpus, we performed continual pre-training on Llama 2 7B, 13B, 70B, Mistral 7B v0.1, and Mixtral 8x7B Instruct as base LLMs and gained consistent (6.6-8.1 points) improvements on Japanese benchmark datasets. We also demonstrate that the improvement on Llama 2 13B brought from the presented corpus was the largest among those from other existing corpora.

Taishi Nakamura, Takuya Akiba, Kazuki Fujii et al.

The Mixture of Experts (MoE) architecture reduces the training and inference cost significantly compared to a dense model of equivalent capacity. Upcycling is an approach that initializes and trains an MoE model using a pre-trained dense model. While upcycling leads to initial performance gains, the training progresses slower than when trained from scratch, leading to suboptimal performance in the long term. We propose Drop-Upcycling - a method that effectively addresses this problem. Drop-Upcycling combines two seemingly contradictory approaches: utilizing the knowledge of pre-trained dense models while statistically re-initializing some parts of the weights. This approach strategically promotes expert specialization, significantly enhancing the MoE model's efficiency in knowledge acquisition. Extensive large-scale experiments demonstrate that Drop-Upcycling significantly outperforms previous MoE construction methods in the long term, specifically when training on hundreds of billions of tokens or more. As a result, our MoE model with 5.9B active parameters achieves comparable performance to a 13B dense model in the same model family, while requiring approximately 1/4 of the training FLOPs. All experimental resources, including source code, training data, model checkpoints and logs, are publicly available to promote reproducibility and future research on MoE.

Kazuki Fujii, Yukito Tajima, Sakae Mizuki et al.

The performance of large language models (LLMs) in program synthesis and mathematical reasoning is fundamentally limited by the quality of their pre-training corpora. We introduce two openly licensed datasets, released under the Llama 3.3 Community License, that significantly enhance LLM performance by systematically rewriting public data. SwallowCode (approximately 16.1 billion tokens) refines Python snippets from The-Stack-v2 through a novel four-stage pipeline: syntax validation, pylint-based style filtering, and a two-stage LLM rewriting process that enforces style conformity and transforms snippets into self-contained, algorithmically efficient examples. Unlike prior methods that rely on exclusionary filtering or limited transformations, our transform-and-retain approach upgrades low-quality code, maximizing data utility. SwallowMath (approximately 2.3 billion tokens) enhances Finemath-4+ by removing boilerplate, restoring context, and reformatting solutions into concise, step-by-step explanations. Within a fixed 50 billion token training budget, continual pre-training of Llama-3.1-8B with SwallowCode boosts pass@1 by +17.0 on HumanEval and +17.7 on HumanEval+ compared to Stack-Edu, surpassing the baseline model's code generation capabilities. Similarly, substituting SwallowMath yields +12.4 accuracy on GSM8K and +7.6 on MATH. Ablation studies confirm that each pipeline stage contributes incrementally, with rewriting delivering the largest gains. All datasets, prompts, and checkpoints are publicly available, enabling reproducible research and advancing LLM pre-training for specialized domains.

Kazuki Fujii, Kohei Watanabe, Rio Yokota

In large language model (LLM) training, several parallelization strategies, including Tensor Parallelism (TP), Pipeline Parallelism (PP), Data Parallelism (DP), as well as Sequence Parallelism (SP) and Context Parallelism (CP), are employed to distribute model parameters, activations, and optimizer states across devices. Identifying the optimal parallelization configuration for each environment while avoiding GPU memory overflow remains a challenging task. In this study, we provide precise formulas to estimate the memory consumed by parameters, gradients, optimizer states, and activations for 4D parallel training (DP, TP, PP, CP) in the Llama architecture. We conducted 454 experiments on A100 and H100 GPUs, incorporating often neglected factors such as temporary buffers and memory fragmentation into our analysis. Results indicate that when the estimated memory usage is below 80\% of the available GPU memory, the training never encounters out-of-memory errors. This simple yet effective formula allows us to identify parallelization configurations that could lead to memory overflow in advance, significantly reducing the configuration search space. Additionally, through a comprehensive exploration of optimal configurations in 4D parallelism, our analysis of the 454 experimental results provides empirical insights into optimal 4D parallelism configurations.

Bai Cong, Nico Daheim, Yuesong Shen et al.

Bayesian methods have recently been used to improve LoRA finetuning and, although they improve calibration, their effect on other metrics (such as accuracy) is marginal and can sometimes even be detrimental. Moreover, Bayesian methods also increase computational overheads and require additional tricks for them to work well. Here, we fix these issues by using a recently proposed variational algorithm called IVON. We show that IVON is easy to implement and has similar costs to AdamW, and yet it can also drastically improve many metrics by using a simple posterior pruning technique. We present extensive results on billion-scale LLMs (Llama and Qwen series) going way beyond the scale of existing applications of IVON. For example, we finetune a Llama-3.2-3B model on a set of commonsense reasoning tasks and improve accuracy over AdamW by 1.3% and reduce ECE by 5.4%, outperforming AdamW and other recent Bayesian methods like Laplace-LoRA and BLoB. Overall, our results show that variational learning with IVON can effectively improve LoRA finetuning.

Koshiro Saito, Sakae Mizuki, Masanari Ohi et al.

Why do we build local large language models (LLMs)? What should a local LLM learn from the target language? Which abilities can be transferred from other languages? Do language-specific scaling laws exist? To explore these research questions, we evaluated 35 Japanese, English, and multilingual LLMs on 19 evaluation benchmarks for Japanese and English, taking Japanese as a local language. Adopting an observational approach, we analyzed correlations of benchmark scores, and conducted principal component analysis (PCA) on the scores to derive \textit{ability factors} of local LLMs. We found that training on English text can improve the scores of academic subjects in Japanese (JMMLU). In addition, it is unnecessary to specifically train on Japanese text to enhance abilities for solving Japanese code generation, arithmetic reasoning, commonsense, and reading comprehension tasks. In contrast, training on Japanese text could improve question-answering tasks about Japanese knowledge and English-Japanese translation, which indicates that abilities for solving these two tasks can be regarded as \textit{Japanese abilities} for LLMs. Furthermore, we confirmed that the Japanese abilities scale with the computational budget for Japanese text.

Satoki Ishikawa, Tal Ben-Nun, Brian Van Essen et al.

Communication overhead is a key challenge in distributed deep learning, especially on slower Ethernet interconnects, and given current hardware trends, communication is likely to become a major bottleneck. While gradient compression techniques have been explored for SGD and Adam, the Lion optimizer has the distinct advantage that its update vectors are the output of a sign operation, enabling straightforward quantization. However, simply compressing updates for communication and using techniques like majority voting fails to lead to end-to-end speedups due to inefficient communication algorithms and reduced convergence. We analyze three factors critical to distributed learning with Lion: optimizing communication methods, identifying effective quantization methods, and assessing the necessity of momentum synchronization. Our findings show that quantization techniques adapted to Lion and selective momentum synchronization can significantly reduce communication costs while maintaining convergence. We combine these into Lion Cub, which enables up to 5x speedups in end-to-end training compared to Lion. This highlights Lion's potential as a communication-efficient solution for distributed training.

Kazuki Fujii, Taishi Nakamura, Rio Yokota

Large Language Models (LLMs) have attracted significant attention due to their human-like language understanding and generation capabilities, as well as their applicability across various domains. These models, characterized by their massive scale and extensive training data, continue to push the boundaries of what is possible in natural language processing. The Llama 3 series, for instance, exemplifies this trend with its flagship model boasting 405 billion parameters trained on 15.6 trillion tokens. The immense computational demands associated with training such models have spurred ongoing research into optimizing the efficiency of the training process, particularly through the use of lower-precision formats. NVIDIA's H100 GPU, which introduces support for FP8 in addition to the more conventional FP16 and BF16 formats, has emerged as a focal point in this optimization effort. Preliminary studies suggest that FP8 could offer substantial reductions in training time without sacrificing model performance when compared to BF16, making it a promising candidate for large-scale model training. However, the broader implications of adopting FP8, particularly in terms of training stability and downstream task performance, have yet to be fully understood. In this study, we delve into the practical trade-offs involved in adopting FP8 over BF16 for training LLMs.

Marlon Tobaben, Mohamed Ali Souibgui, Rubèn Tito et al.

The Privacy Preserving Federated Learning Document VQA (PFL-DocVQA) competition challenged the community to develop provably private and communication-efficient solutions in a federated setting for a real-life use case: invoice processing. The competition introduced a dataset of real invoice documents, along with associated questions and answers requiring information extraction and reasoning over the document images. Thereby, it brings together researchers and expertise from the document analysis, privacy, and federated learning communities. Participants fine-tuned a pre-trained, state-of-the-art Document Visual Question Answering model provided by the organizers for this new domain, mimicking a typical federated invoice processing setup. The base model is a multi-modal generative language model, and sensitive information could be exposed through either the visual or textual input modality. Participants proposed elegant solutions to reduce communication costs while maintaining a minimum utility threshold in track 1 and to protect all information from each document provider using differential privacy in track 2. The competition served as a new testbed for developing and testing private federated learning methods, simultaneously raising awareness about privacy within the document image analysis and recognition community. Ultimately, the competition analysis provides best practices and recommendations for successfully running privacy-focused federated learning challenges in the future.

Masaki Kawamura, Nakamasa Inoue, Rintaro Yanagi et al.

Contrastive vision-language pre-training frameworks such as CLIP have demonstrated impressive zero-shot performance across a range of vision-language tasks. Recent studies have shown that aligning individual text tokens with specific image patches or regions enhances fine-grained compositional understanding. However, it remains challenging to capture compositional semantics that span multiple image regions. To address this limitation, we propose PowerCLIP, a novel contrastive pre-training framework enhanced by powerset alignment, which exhaustively optimizes region-to-phrase alignments by minimizing the loss defined between powersets of image regions and textual parse trees. Since the naive powerset construction incurs exponential computational cost due to the combinatorial explosion in the number of region subsets, we introduce efficient non-linear aggregators (NLAs) that reduce complexity from O(2^M) to O(M) with respect to the number of regions M, while approximating the exact loss value with arbitrary precision. Our extensive experiments demonstrate that PowerCLIP outperforms state-of-the-art methods in zero-shot classification and retrieval tasks, underscoring the compositionality and robustness of our approach. Our code will be made publicly available.

Yukito Tajima, Nakamasa Inoue, Yusuke Sekikawa et al.

Gated Linear Units (GLUs) have become essential components in the feed-forward networks of state-of-the-art Large Language Models (LLMs). However, they require twice as many memory reads compared to feed-forward layers without gating, due to the use of separate weight matrices for the gate and value streams. To address this bottleneck, we introduce Masked Gated Linear Units (MGLUs), a novel family of GLUs with an efficient kernel implementation. The core contribution of MGLUs include: (1) the Mixture of Element-wise Gating (MoEG) architecture that learns multiple binary masks, each determining gate or value assignments at the element level on a single shared weight matrix resulting in reduced memory transfer, and (2) FlashMGLU, a hardware-friendly kernel that yields up to a 19.7 $\times$ inference-time speed-up over a naive PyTorch MGLU and is 47% more memory-efficient and 34% faster than standard GLUs despite added architectural complexity on an RTX5090 GPU. In LLM experiments, the Swish-activated variant SwiMGLU preserves its memory advantages while matching - or even surpassing - the downstream accuracy of the SwiGLU baseline.

Avrajit Ghosh, Bai Cong, Rio Yokota et al.

Variational Learning (VL) has recently gained popularity for training deep neural networks. Part of its empirical success can be explained by theories such as PAC-Bayes bounds, minimum description length and marginal likelihood, but little has been done to unravel the implicit regularization in play. Here, we analyze the implicit regularization of VL through the Edge of Stability (EoS) framework. EoS has previously been used to show that gradient descent can find flat solutions and we extend this result to show that VL can find even flatter solutions. This result is obtained by controlling the shape of the variational posterior as well as the number of posterior samples used during training. The derivation follows in a similar fashion as in the standard EoS literature for deep learning, by first deriving a result for a quadratic problem and then extending it to deep neural networks. We empirically validate these findings on a wide variety of large networks, such as ResNet and ViT, to find that the theoretical results closely match the empirical ones. Ours is the first work to analyze the EoS dynamics of VL.

Youmi Ma, Sakae Mizuki, Kazuki Fujii et al.

Instruction tuning is crucial for enabling Large Language Models (LLMs) to solve real-world tasks. Prior work has shown the effectiveness of instruction-tuning data synthesized solely from LLMs, raising a fundamental question: Do we still need human-originated signals for instruction tuning? This work answers the question affirmatively: we build state-of-the-art instruction-tuning datasets sourced from human-written instructions, by simply pairing them with LLM-generated responses. LLMs fine-tuned on our datasets consistently outperform those fine-tuned on existing ones. Our data construction approach can be easily adapted to other languages; we build datasets for Japanese and confirm that LLMs tuned with our data reach state-of-the-art performance. Analyses suggest that instruction-tuning in a new language allows LLMs to follow instructions, while the tuned models exhibit a notable lack of culture-specific knowledge in that language. The datasets and fine-tuned models will be publicly available. Our datasets, synthesized with open-weight LLMs, are openly distributed under permissive licenses, allowing for diverse use cases.

Shun Iwase, Shuya Takahashi, Nakamasa Inoue et al.

The double descent phenomenon, which deviates from the traditional bias-variance trade-off theory, attracts considerable research attention; however, the mechanism of its occurrence is not fully understood. On the other hand, in the study of convolutional neural networks (CNNs) for image recognition, methods are proposed to quantify the bias on shape features versus texture features in images, determining which features the CNN focuses on more. In this work, we hypothesize that there is a relationship between the shape/texture bias in the learning process of CNNs and epoch-wise double descent, and we conduct verification. As a result, we discover double descent/ascent of shape/texture bias synchronized with double descent of test error under conditions where epoch-wise double descent is observed. Quantitative evaluations confirm this correlation between the test errors and the bias values from the initial decrease to the full increase in test error. Interestingly, double descent/ascent of shape/texture bias is observed in some cases even in conditions without label noise, where double descent is thought not to occur. These experimental results are considered to contribute to the understanding of the mechanisms behind the double descent phenomenon and the learning process of CNNs in image recognition.

Satoki Ishikawa, Rio Yokota, Ryo Karakida

Local learning, which trains a network through layer-wise local targets and losses, has been studied as an alternative to backpropagation (BP) in neural computation. However, its algorithms often become more complex or require additional hyperparameters because of the locality, making it challenging to identify desirable settings in which the algorithm progresses in a stable manner. To provide theoretical and quantitative insights, we introduce the maximal update parameterization ($μ$P) in the infinite-width limit for two representative designs of local targets: predictive coding (PC) and target propagation (TP). We verified that $μ$P enables hyperparameter transfer across models of different widths. Furthermore, our analysis revealed unique and intriguing properties of $μ$P that are not present in conventional BP. By analyzing deep linear networks, we found that PC's gradients interpolate between first-order and Gauss-Newton-like gradients, depending on the parameterization. We demonstrate that, in specific standard settings, PC in the infinite-width limit behaves more similarly to the first-order gradient. For TP, even with the standard scaling of the last layer, which differs from classical $μ$P, its local loss optimization favors the feature learning regime over the kernel regime.

Kazuki Osawa, Satoki Ishikawa, Rio Yokota et al.

Gradient preconditioning is a key technique to integrate the second-order information into gradients for improving and extending gradient-based learning algorithms. In deep learning, stochasticity, nonconvexity, and high dimensionality lead to a wide variety of gradient preconditioning methods, with implementation complexity and inconsistent performance and feasibility. We propose the Automatic Second-order Differentiation Library (ASDL), an extension library for PyTorch, which offers various implementations and a plug-and-play unified interface for gradient preconditioning. ASDL enables the study and structured comparison of a range of gradient preconditioning methods.

Hana Hoshino, Kei Ota, Asako Kanezaki et al.

Inverse Reinforcement Learning (IRL) is attractive in scenarios where reward engineering can be tedious. However, prior IRL algorithms use on-policy transitions, which require intensive sampling from the current policy for stable and optimal performance. This limits IRL applications in the real world, where environment interactions can become highly expensive. To tackle this problem, we present Off-Policy Inverse Reinforcement Learning (OPIRL), which (1) adopts off-policy data distribution instead of on-policy and enables significant reduction of the number of interactions with the environment, (2) learns a stationary reward function that is transferable with high generalization capabilities on changing dynamics, and (3) leverages mode-covering behavior for faster convergence. We demonstrate that our method is considerably more sample efficient and generalizes to novel environments through the experiments. Our method achieves better or comparable results on policy performance baselines with significantly fewer interactions. Furthermore, we empirically show that the recovered reward function generalizes to different tasks where prior arts are prone to fail.

Kento Doi, Ryuhei Hamaguchi, Shun Iwase et al.

This paper describes a viewpoint-robust object-based change detection network (OBJ-CDNet). Mobile cameras such as drive recorders capture images from different viewpoints each time due to differences in camera trajectory and shutter timing. However, previous methods for pixel-wise change detection are vulnerable to the viewpoint differences because they assume aligned image pairs as inputs. To cope with the difficulty, we introduce a deep graph matching network that establishes object correspondence between an image pair. The introduction enables us to detect object-wise scene changes without precise image alignment. For more accurate object matching, we propose an epipolar-guided deep graph matching network (EGMNet), which incorporates the epipolar constraint into the deep graph matching layer used in OBJCDNet. To evaluate our network's robustness against viewpoint differences, we created synthetic and real datasets for scene change detection from an image pair. The experimental results verified the effectiveness of our network.

Kazuki Osawa, Yohei Tsuji, Yuichiro Ueno et al.

Large-scale distributed training of deep neural networks results in models with worse generalization performance as a result of the increase in the effective mini-batch size. Previous approaches attempt to address this problem by varying the learning rate and batch size over epochs and layers, or ad hoc modifications of batch normalization. We propose Scalable and Practical Natural Gradient Descent (SP-NGD), a principled approach for training models that allows them to attain similar generalization performance to models trained with first-order optimization methods, but with accelerated convergence. Furthermore, SP-NGD scales to large mini-batch sizes with a negligible computational overhead as compared to first-order methods. We evaluated SP-NGD on a benchmark task where highly optimized first-order methods are available as references: training a ResNet-50 model for image classification on ImageNet. We demonstrate convergence to a top-1 validation accuracy of 75.4% in 5.5 minutes using a mini-batch size of 32,768 with 1,024 GPUs, as well as an accuracy of 74.9% with an extremely large mini-batch size of 131,072 in 873 steps of SP-NGD.

Kazuki Osawa, Yohei Tsuji, Yuichiro Ueno et al.

Large-scale distributed training of deep neural networks suffer from the generalization gap caused by the increase in the effective mini-batch size. Previous approaches try to solve this problem by varying the learning rate and batch size over epochs and layers, or some ad hoc modification of the batch normalization. We propose an alternative approach using a second-order optimization method that shows similar generalization capability to first-order methods, but converges faster and can handle larger mini-batches. To test our method on a benchmark where highly optimized first-order methods are available as references, we train ResNet-50 on ImageNet. We converged to 75% Top-1 validation accuracy in 35 epochs for mini-batch sizes under 16,384, and achieved 75% even with a mini-batch size of 131,072, which took only 978 iterations.