Tatsuya Hiraoka

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/.

Kohei Tsuji, Tatsuya Hiraoka, Yuchang Cheng et al.

NLP datasets may still contain annotation errors, even when they are manually annotated. Researchers have attempted to develop methods to automatically reduce the adverse effect of errors in datasets. However, existing methods are time-consuming because they require many trained models to detect errors. This paper proposes a time-saving method that utilizes a tokenization technique called subword regularization to simulate multiple error detection models for detecting errors. Our proposed method, SubRegWeigh, can perform annotation weighting four to five times faster than the existing method. Additionally, SubRegWeigh improved performance in document classification and named entity recognition tasks. In experiments with pseudo-incorrect labels, SubRegWeigh clearly identifies pseudo-incorrect labels as annotation errors. Our code is available at https://github.com/4ldk/SubRegWeigh .

Sho Takase, Tatsuya Hiraoka, Naoaki Okazaki

Subword regularizations use multiple subword segmentations during training to improve the robustness of neural machine translation models. In previous subword regularizations, we use multiple segmentations in the training process but use only one segmentation in the inference. In this study, we propose an inference strategy to address this discrepancy. The proposed strategy approximates the marginalized likelihood by using multiple segmentations including the most plausible segmentation and several sampled segmentations. Because the proposed strategy aggregates predictions from several segmentations, we can regard it as a single model ensemble that does not require any additional cost for training. Experimental results show that the proposed strategy improves the performance of models trained with subword regularization in low-resource machine translation tasks.

Tatsuya Hiraoka

We present a subword regularization method for WordPiece, which uses a maximum matching algorithm for tokenization. The proposed method, MaxMatch-Dropout, randomly drops words in a search using the maximum matching algorithm. It realizes finetuning with subword regularization for popular pretrained language models such as BERT-base. The experimental results demonstrate that MaxMatch-Dropout improves the performance of text classification and machine translation tasks as well as other subword regularization methods. Moreover, we provide a comparative analysis of subword regularization methods: subword regularization with SentencePiece (Unigram), BPE-Dropout, and MaxMatch-Dropout.

Tatsuya Hiraoka, Tomoya Iwakura

Is preferred tokenization for humans also preferred for machine-learning (ML) models? This study examines the relations between preferred tokenization for humans (appropriateness and readability) and one for ML models (performance on an NLP task). The question texts of the Japanese commonsense question-answering dataset are tokenized with six different tokenizers, and the performances of human annotators and ML models were compared. Furthermore, we analyze relations among performance of answers by human and ML model, the appropriateness of tokenization for human, and response time to questions by human. This study provides a quantitative investigation result that shows that preferred tokenizations for humans and ML models are not necessarily always the same. The result also implies that existing methods using language models for tokenization could be a good compromise both for human and ML models.

Tatsuya Hiraoka, Tomoya Iwakura

This paper proposes a method to optimize tokenization for the performance improvement of already trained downstream models. Our method generates tokenization results attaining lower loss values of a given downstream model on the training data for restricting vocabularies and trains a tokenizer reproducing the tokenization results. Therefore, our method can be applied to variety of tokenization methods, while existing work cannot due to the simultaneous learning of the tokenizer and the downstream model. This paper proposes an example of the BiLSTM-based tokenizer with vocabulary restriction, which can capture wider contextual information for the tokenization process than non-neural-based tokenization methods used in existing work. Experimental results on text classification in Japanese, Chinese, and English text classification tasks show that the proposed method improves performance compared to the existing methods for tokenization optimization.

Jesse Atuhurra, Iqra Ali, Tomoya Iwakura et al.

Vision Language Models (VLMs) are pivotal for advancing perception in intelligent agents. Yet, evaluation of VLMs remains limited to predominantly English-centric benchmarks in which the image-text pairs comprise short texts. To evaluate VLM fine-grained abilities, in four languages under long-text settings, we introduce a novel multilingual benchmark VLURes featuring eight vision-and-language tasks, and a pioneering unrelatedness task, to probe the fine-grained Visual and Linguistic Understanding capabilities of VLMs across English, Japanese, and low-resource languages, Swahili, and Urdu. Our datasets, curated from web resources in the target language, encompass ten diverse image categories and rich textual context, introducing valuable vision-language resources for Swahili and Urdu. By prompting VLMs to generate responses and rationales, evaluated automatically and by native speakers, we uncover performance disparities across languages and tasks critical to intelligent agents, such as object recognition, scene understanding, and relationship understanding. We conducted evaluations of ten VLMs with VLURes. The best performing model, GPT-4o, achieves an overall accuracy of 90.8% and lags human performance by 6.7%, though the gap is larger for open-source models. The gap highlights VLURes' critical role in developing intelligent agents to tackle multi-modal visual reasoning.

Munachiso Nwadike, Zangir Iklassov, Toluwani Aremu et al.

We introduce the concept of the self-referencing causal cycle (abbreviated RECALL) - a mechanism that enables large language models (LLMs) to bypass the limitations of unidirectional causality, which underlies a phenomenon known as the reversal curse. When an LLM is prompted with sequential data, it often fails to recall preceding context. For example, when we ask an LLM to recall the line preceding "O say does that star-spangled banner yet wave" in the U.S. National Anthem, it often fails to correctly return "Gave proof through the night that our flag was still there" - this is due to the reversal curse. It occurs because language models such as ChatGPT and Llama generate text based on preceding tokens, requiring facts to be learned and reproduced in a consistent token order. While the reversal curse is often viewed as a limitation, we offer evidence of an alternative view: it is not always an obstacle in practice. We find that RECALL is driven by what we designate as cycle tokens - sequences that connect different parts of the training data, enabling recall of preceding tokens from succeeding ones. Through rigorous probabilistic formalization and controlled experiments, we demonstrate how the cycles they induce influence a model's ability to reproduce information. To facilitate reproducibility, we provide our code and experimental details at https://anonymous.4open.science/r/remember-B0B8/.

Rifo Genadi, Munachiso Nwadike, Nurdaulet Mukhituly et al.

We find that correct-to-incorrect sycophancy signals are most linearly separable within multi-head attention activations. Motivated by the linear representation hypothesis, we train linear probes across the residual stream, multilayer perceptron (MLP), and attention layers to analyze where these signals emerge. Although separability appears in the residual stream and MLPs, steering using these probes is most effective in a sparse subset of middle-layer attention heads. Using TruthfulQA as the base dataset, we find that probes trained on it transfer effectively to other factual QA benchmarks. Furthermore, comparing our discovered direction to previously identified "truthful" directions reveals limited overlap, suggesting that factual accuracy, and deference resistance, arise from related but distinct mechanisms. Attention-pattern analysis further indicates that the influential heads attend disproportionately to expressions of user doubt, contributing to sycophantic shifts. Overall, these findings suggest that sycophancy can be mitigated through simple, targeted linear interventions that exploit the internal geometry of attention activations.

Tatsuya Hiraoka, Kentaro Inui

This paper introduces repetition neurons, regarded as skill neurons responsible for the repetition problem in text generation tasks. These neurons are progressively activated more strongly as repetition continues, indicating that they perceive repetition as a task to copy the previous context repeatedly, similar to in-context learning. We identify these repetition neurons by comparing activation values before and after the onset of repetition in texts generated by recent pre-trained language models. We analyze the repetition neurons in three English and one Japanese pre-trained language models and observe similar patterns across them.

Tatsuya Hiraoka, Naoaki Okazaki

Do pretrained language models have knowledge regarding the surface information of tokens? We examined the surface information stored in word or subword embeddings acquired by pretrained language models from the perspectives of token length, substrings, and token constitution. Additionally, we evaluated the ability of models to generate knowledge regarding token surfaces. We focused on 12 pretrained language models that were mainly trained on English and Japanese corpora. Experimental results demonstrate that pretrained language models have knowledge regarding token length and substrings but not token constitution. Additionally, the results imply that there is a bottleneck on the decoder side in terms of effectively utilizing acquired knowledge.

Ahmed Oumar El-Shangiti, Tatsuya Hiraoka, Hilal AlQuabeh et al.

This paper investigates whether large language models (LLMs) utilize numerical attributes encoded in a low-dimensional subspace of the embedding space when answering questions involving numeric comparisons, e.g., Was Cristiano born before Messi? We first identified, using partial least squares regression, these subspaces, which effectively encode the numerical attributes associated with the entities in comparison prompts. Further, we demonstrate causality, by intervening in these subspaces to manipulate hidden states, thereby altering the LLM's comparison outcomes. Experiments conducted on three different LLMs showed that our results hold across different numerical attributes, indicating that LLMs utilize the linearly encoded information for numerical reasoning.

Marco Cognetta, Tatsuya Hiraoka, Naoaki Okazaki et al.

We explore threshold vocabulary trimming in Byte-Pair Encoding subword tokenization, a postprocessing step that replaces rare subwords with their component subwords. The technique is available in popular tokenization libraries but has not been subjected to rigorous scientific scrutiny. While the removal of rare subwords is suggested as best practice in machine translation implementations, both as a means to reduce model size and for improving model performance through robustness, our experiments indicate that, across a large space of hyperparameter settings, vocabulary trimming fails to improve performance, and is even prone to incurring heavy degradation.

Tatsuya Hiraoka, Kentaro Inui

Large language models (LLMs) can spell out tokens character by character with high accuracy, yet they struggle with more complex character-level tasks, such as identifying compositional subcomponents within tokens. In this work, we investigate how LLMs internally represent and utilize character-level information during the spelling-out process. Our analysis reveals that, although spelling out is a simple task for humans, it is not handled in a straightforward manner by LLMs. Specifically, we show that the embedding layer does not fully encode character-level information, particularly beyond the first character. As a result, LLMs rely on intermediate and higher Transformer layers to reconstruct character-level knowledge, where we observe a distinct "breakthrough" in their spelling behavior. We validate this mechanism through three complementary analyses: probing classifiers, identification of knowledge neurons, and inspection of attention weights.

Jesse Atuhurra, Iqra Ali, Tatsuya Hiraoka et al.

Large language models (LLMs) have increased interest in vision language models (VLMs), which process image-text pairs as input. Studies investigating the visual understanding ability of VLMs have been proposed, but such studies are still preliminary because existing datasets do not permit a comprehensive evaluation of the fine-grained visual linguistic abilities of VLMs across multiple languages. To further explore the strengths of VLMs, such as GPT-4V \cite{openai2023GPT4}, we developed new datasets for the systematic and qualitative analysis of VLMs. Our contribution is four-fold: 1) we introduced nine vision-and-language (VL) tasks (including object recognition, image-text matching, and more) and constructed multilingual visual-text datasets in four languages: English, Japanese, Swahili, and Urdu through utilizing templates containing \textit{questions} and prompting GPT4-V to generate the \textit{answers} and the \textit{rationales}, 2) introduced a new VL task named \textit{unrelatedness}, 3) introduced rationales to enable human understanding of the VLM reasoning process, and 4) employed human evaluation to measure the suitability of proposed datasets for VL tasks. We show that VLMs can be fine-tuned on our datasets. Our work is the first to conduct such analyses in Swahili and Urdu. Also, it introduces \textit{rationales} in VL analysis, which played a vital role in the evaluation.

Seiya Ishikura, Hiroaki Yamada, Tatsuya Hiraoka et al.

This study proposes augmenting dialog data with think-aloud utterances (TAUs) for modeling individual personalities in text chat by LLM. TAU is a verbalization of a speaker's thought before articulating the utterance. We expect "persona LLMs" trained with TAU-augmented data can mimic the speaker's personality trait better. We tested whether the trained persona LLMs obtain the human personality with respect to Big Five, a framework characterizing human personality traits from five aspects. The results showed that LLMs trained with TAU-augmented data more closely align to the speakers' Agreeableness and Neuroticism of Big Five than those trained with original dialog data. We also found that the quality of TAU-augmentation impacts persona LLM's performance.

Nhi Hoai Doan, Tatsuya Hiraoka, Kentaro Inui

This paper investigates the relationship between large language models' (LLMs) ability to recognize repetitive input patterns and their performance on in-context learning (ICL). In contrast to prior work that has primarily focused on attention heads, we examine this relationship from the perspective of skill neurons, specifically repetition neurons. Our experiments reveal that the impact of these neurons on ICL performance varies depending on the depth of the layer in which they reside. By comparing the effects of repetition neurons and induction heads, we further identify strategies for reducing repetitive outputs while maintaining strong ICL capabilities.

Sangwhan Moon, Tatsuya Hiraoka, Naoaki Okazaki

Byte-level fallbacks for subword tokenization have become a common practice in large language models. In particular, it has been demonstrated to be incredibly effective as a pragmatic solution for preventing OOV, especially in the context of larger models. However, breaking a character down to individual bytes significantly increases the sequence length for long-tail tokens in languages such as Chinese, Japanese, and Korean (CJK) and other character-diverse contexts such as emoji. The increased sequence length results in longer computation during both training and inference. In this work, we propose a simple compression technique that reduces the sequence length losslessly.

Shintaro Ozaki, Tatsuya Hiraoka, Hiroto Otake et al.

Large Language Models (LLMs) are known to process information using a proficient internal language consistently, referred to as latent language, which may differ from the input or output languages. However, how the discrepancy between the latent language and the input and output language affects downstream task performance remains largely unexplored. While many studies research the latent language of LLMs, few address its importance in influencing task performance. In our study, we hypothesize that thinking in latent language consistently enhances downstream task performance. To validate this, our work varies the input prompt languages across multiple downstream tasks and analyzes the correlation between consistency in latent language and task performance. We create datasets consisting of questions from diverse domains such as translation and geo-culture, which are influenced by the choice of latent language. Experimental results across multiple LLMs on translation and geo-culture tasks, which are sensitive to the choice of language, indicate that maintaining consistency in latent language is not always necessary for optimal downstream task performance. This is because these models adapt their internal representations near the final layers to match the target language, reducing the impact of consistency on overall performance.

Kohei Tsuji, Tatsuya Hiraoka, Yuchang Cheng et al.

This paper investigates how LLMs encode inputs with typos. We hypothesize that specific neurons and attention heads recognize typos and fix them internally using local and global contexts. We introduce a method to identify typo neurons and typo heads that work actively when inputs contain typos. Our experimental results suggest the following: 1) LLMs can fix typos with local contexts when the typo neurons in either the early or late layers are activated, even if those in the other are not. 2) Typo neurons in the middle layers are responsible for the core of typo-fixing with global contexts. 3) Typo heads fix typos by widely considering the context not focusing on specific tokens. 4) Typo neurons and typo heads work not only for typo-fixing but also for understanding general contexts.

H. V. AlquBoj, Hilal AlQuabeh, Velibor Bojkovic et al.

Humans are believed to perceive numbers on a logarithmic mental number line, where smaller values are represented with greater resolution than larger ones. This cognitive bias, supported by neuroscience and behavioral studies, suggests that numerical magnitudes are processed in a sublinear fashion rather than on a uniform linear scale. Inspired by this hypothesis, we investigate whether large language models (LLMs) exhibit a similar logarithmic-like structure in their internal numerical representations. By analyzing how numerical values are encoded across different layers of LLMs, we apply dimensionality reduction techniques such as PCA and PLS followed by geometric regression to uncover latent structures in the learned embeddings. Our findings reveal that the model's numerical representations exhibit sublinear spacing, with distances between values aligning with a logarithmic scale. This suggests that LLMs, much like humans, may encode numbers in a compressed, non-uniform manner.

Tatsuya Hiraoka, Sho Takase, Kei Uchiumi et al.

Since traditional tokenizers are isolated from a downstream task and model, they cannot output an appropriate tokenization depending on the task and model, although recent studies imply that the appropriate tokenization improves the performance. In this paper, we propose a novel method to find an appropriate tokenization to a given downstream model by jointly optimizing a tokenizer and the model. The proposed method has no restriction except for using loss values computed by the downstream model to train the tokenizer, and thus, we can apply the proposed method to any NLP task. Moreover, the proposed method can be used to explore the appropriate tokenization for an already trained model as post-processing. Therefore, the proposed method is applicable to various situations. We evaluated whether our method contributes to improving performance on text classification in three languages and machine translation in eight language pairs. Experimental results show that our proposed method improves the performance by determining appropriate tokenizations.

Youmi Ma, Tatsuya Hiraoka, Naoaki Okazaki

In this study, a novel method for extracting named entities and relations from unstructured text based on the table representation is presented. By using contextualized word embeddings, the proposed method computes representations for entity mentions and long-range dependencies without complicated hand-crafted features or neural-network architectures. We also adapt a tensor dot-product to predict relation labels all at once without resorting to history-based predictions or search strategies. These advances significantly simplify the model and algorithm for the extraction of named entities and relations. Despite its simplicity, the experimental results demonstrate that the proposed method outperforms the state-of-the-art methods on the CoNLL04 and ACE05 English datasets. We also confirm that the proposed method achieves a comparable performance with the state-of-the-art NER models on the ACE05 datasets when multiple sentences are provided for context aggregation.