arXiv subjects

Seiya Tokui

Publications and source records attributed to Seiya Tokui.

8 recordsLinked to original sources

arXiv2026

Coarse Semantic Blocks for Efficient Training-Free Sparse Attention

We introduce a training-free sparse attention method that builds a prompt-specific key index to accelerate both prefill and decoding. The method clusters the keys into large semantic blocks that group tokens from noncontiguous positions. Each query position selects blocks by scoring their centroids and attends to the original keys and values within them. Queries that select the same block are processed together, so prefill stays fast even though each query chooses its own blocks. Large blocks make centroid scoring and this grouped processing cheap, so each query can attend to more tokens in the same time. Larger blocks with a larger token budget can thus be both faster and more accurate on long-context tasks than smaller blocks with a smaller budget. We evaluate two hybrid LLMs, Qwen3.5-4B and 35B-A3B, on one NVIDIA H200. Including the cost of building and using the index, whole-model wall-clock speedups over dense attention are 2.3-2.4x for prefill and 1.7-2.0x for decoding at 512K tokens, and 2.9-3.0x for prefill at 1M. The 35B-A3B model stays close to dense attention on five RULER tasks through 1M, on LongBench-v2 overall, and on MRCR through 256K. The 4B model loses accuracy mainly on multi-key retrieval, yet outperforms dense Qwen3.5-2B, whose latency is similar at 512K, on all three benchmarks.

cs.LG↗

arXiv2026

Head Dimension Matters in Language Models for Long-Context Retrieval

Multi-head attention retrieves and combines context representations through query--key interactions and value aggregation, with head dimension constraining the per-head representation space. In long-context processing, it remains unclear whether allocating a fixed parameter budget to fewer, wider heads improves information retrieval. In this paper, we investigate how head dimension affects the performance of long-context retrieval tasks. Our empirical studies reveal that pretraining loss and short-context non-retrieval task performance are not affected much by head dimension, while long-context retrieval performance improves with wider heads at fixed model sizes. We find that models with wider heads have a higher proportion of retrieval heads, associating more parameters with retrieval. Our ablation studies also show that the retrieval performance improves with wider heads even when the projection weights in attention modules are frozen during long-context continual pretraining, suggesting that wider heads not only benefit updates in attention modules but lead to more effective adaptation of components such as embeddings and MLPs.

cs.LG↗

arXiv2025

When Does Metadata Conditioning (NOT) Work for Language Model Pre-Training? A Study with Context-Free Grammars

The ability to acquire latent semantics is one of the key properties that determines the performance of language models. One convenient approach to invoke this ability is to prepend metadata (e.g. URLs, domains, and styles) at the beginning of texts in the pre-training data, making it easier for the model to access latent semantics before observing the entire text. Previous studies have reported that this technique actually improves the performance of trained models in downstream tasks; however, this improvement has been observed only in specific downstream tasks, without consistent enhancement in average next-token prediction loss. To understand this phenomenon, we closely investigate how prepending metadata during pre-training affects model performance by examining its behavior using artificial data. Interestingly, we found that this approach produces both positive and negative effects on the downstream tasks. We demonstrate that the effectiveness of the approach depends on whether latent semantics can be inferred from the downstream task's prompt. Specifically, through investigations using data generated by probabilistic context-free grammars, we show that training with metadata helps improve model's performance when the given context is long enough to infer the latent semantics. In contrast, the technique negatively impacts performance when the context lacks the necessary information to make an accurate posterior inference.

cs.CL↗

arXiv2022

Disentanglement Analysis with Partial Information Decomposition

We propose a framework to analyze how multivariate representations disentangle ground-truth generative factors. A quantitative analysis of disentanglement has been based on metrics designed to compare how one variable explains each generative factor. Current metrics, however, may fail to detect entanglement that involves more than two variables, e.g., representations that duplicate and rotate generative factors in high dimensional spaces. In this work, we establish a framework to analyze information sharing in a multivariate representation with Partial Information Decomposition and propose a new disentanglement metric. This framework enables us to understand disentanglement in terms of uniqueness, redundancy, and synergy. We develop an experimental protocol to assess how increasingly entangled representations are evaluated with each metric and confirm that the proposed metric correctly responds to entanglement. Through experiments on variational autoencoders, we find that models with similar disentanglement scores have a variety of characteristics in entanglement, for each of which a distinct strategy may be required to obtain a disentangled representation.

stat.ML↗

arXiv2019

Chainer: A Deep Learning Framework for Accelerating the Research Cycle

Software frameworks for neural networks play a key role in the development and application of deep learning methods. In this paper, we introduce the Chainer framework, which intends to provide a flexible, intuitive, and high performance means of implementing the full range of deep learning models needed by researchers and practitioners. Chainer provides acceleration using Graphics Processing Units with a familiar NumPy-like API through CuPy, supports general and dynamic models in Python through Define-by-Run, and also provides add-on packages for state-of-the-art computer vision models as well as distributed training.

cs.LG↗

arXiv2018

Adversarial Attacks and Defences Competition

To accelerate research on adversarial examples and robustness of machine learning classifiers, Google Brain organized a NIPS 2017 competition that encouraged researchers to develop new methods to generate adversarial examples as well as to develop new ways to defend against them. In this chapter, we describe the structure and organization of the competition and the solutions developed by several of the top-placing teams.

cs.CV↗

arXiv2017

Learning Discrete Representations via Information Maximizing Self-Augmented Training

Learning discrete representations of data is a central machine learning task because of the compactness of the representations and ease of interpretation. The task includes clustering and hash learning as special cases. Deep neural networks are promising to be used because they can model the non-linearity of data and scale to large datasets. However, their model complexity is huge, and therefore, we need to carefully regularize the networks in order to learn useful representations that exhibit intended invariance for applications of interest. To this end, we propose a method called Information Maximizing Self-Augmented Training (IMSAT). In IMSAT, we use data augmentation to impose the invariance on discrete representations. More specifically, we encourage the predicted representations of augmented data points to be close to those of the original data points in an end-to-end fashion. At the same time, we maximize the information-theoretic dependency between data and their predicted discrete representations. Extensive experiments on benchmark datasets show that IMSAT produces state-of-the-art results for both clustering and unsupervised hash learning.

stat.ML↗

arXiv2016

Reparameterization trick for discrete variables

Low-variance gradient estimation is crucial for learning directed graphical models parameterized by neural networks, where the reparameterization trick is widely used for those with continuous variables. While this technique gives low-variance gradient estimates, it has not been directly applicable to discrete variables, the sampling of which inherently requires discontinuous operations. We argue that the discontinuity can be bypassed by marginalizing out the variable of interest, which results in a new reparameterization trick for discrete variables. This reparameterization greatly reduces the variance, which is understood by regarding the method as an application of common random numbers to the estimation. The resulting estimator is theoretically guaranteed to have a variance not larger than that of the likelihood-ratio method with the optimal input-dependent baseline. We give empirical results for variational learning of sigmoid belief networks.

stat.ML↗