Jun Suzuki

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/.

Sho Takase, Shun Kiyono, Sosuke Kobayashi et al.

From the perspective of the layer normalization (LN) positions, the architectures of Transformers can be categorized into two types: Post-LN and Pre-LN. Recent Transformers tend to be Pre-LN because, in Post-LN with deep Transformers (e.g., those with ten or more layers), the training is often unstable, resulting in useless models. However, Post-LN has consistently achieved better performance than Pre-LN in relatively shallow Transformers (e.g., those with six or fewer layers). This study first investigates the reason for these discrepant observations empirically and theoretically and made the following discoveries: 1, the LN in Post-LN is the main source of the vanishing gradient problem that leads to unstable training, whereas Pre-LN prevents it, and 2, Post-LN tends to preserve larger gradient norms in higher layers during the back-propagation, which may lead to effective training. Exploiting the new findings, we propose a method that can provide both high stability and effective training by a simple modification of Post-LN. We conduct experiments on a wide range of text generation tasks. The experimental results demonstrate that our method outperforms Pre-LN, and enables stable training regardless of the shallow or deep layer settings. Our code is publicly available at https://github.com/takase/b2t_connection.

Masato Mita, Keisuke Sakaguchi, Masato Hagiwara et al.

Natural language processing technology has rapidly improved automated grammatical error correction tasks, and the community begins to explore document-level revision as one of the next challenges. To go beyond sentence-level automated grammatical error correction to NLP-based document-level revision assistant, there are two major obstacles: (1) there are few public corpora with document-level revisions being annotated by professional editors, and (2) it is not feasible to elicit all possible references and evaluate the quality of revision with such references because there are infinite possibilities of revision. This paper tackles these challenges. First, we introduce a new document-revision corpus, TETRA, where professional editors revised academic papers sampled from the ACL anthology which contain few trivial grammatical errors that enable us to focus more on document- and paragraph-level edits such as coherence and consistency. Second, we explore reference-less and interpretable methods for meta-evaluation that can detect quality improvements by document revision. We show the uniqueness of TETRA compared with existing document revision corpora and demonstrate that a fine-tuned pre-trained language model can discriminate the quality of documents after revision even when the difference is subtle. This promising result will encourage the community to further explore automated document revision models and metrics in future.

Hiroaki Funayama, Tasuku Sato, Yuichiroh Matsubayashi et al.

Short answer scoring (SAS) is the task of grading short text written by a learner. In recent years, deep-learning-based approaches have substantially improved the performance of SAS models, but how to guarantee high-quality predictions still remains a critical issue when applying such models to the education field. Towards guaranteeing high-quality predictions, we present the first study of exploring the use of human-in-the-loop framework for minimizing the grading cost while guaranteeing the grading quality by allowing a SAS model to share the grading task with a human grader. Specifically, by introducing a confidence estimation method for indicating the reliability of the model predictions, one can guarantee the scoring quality by utilizing only predictions with high reliability for the scoring results and casting predictions with low reliability to human graders. In our experiments, we investigate the feasibility of the proposed framework using multiple confidence estimation methods and multiple SAS datasets. We find that our human-in-the-loop framework allows automatic scoring models and human graders to achieve the target scoring quality.

Yosuke Kishinami, Reina Akama, Shiki Sato et al.

Prior studies addressing target-oriented conversational tasks lack a crucial notion that has been intensively studied in the context of goal-oriented artificial intelligence agents, namely, planning. In this study, we propose the task of Target-Guided Open-Domain Conversation Planning (TGCP) task to evaluate whether neural conversational agents have goal-oriented conversation planning abilities. Using the TGCP task, we investigate the conversation planning abilities of existing retrieval models and recent strong generative models. The experimental results reveal the challenges facing current technology.

Sosuke Kobayashi, Shun Kiyono, Jun Suzuki et al.

Ensembling is a popular method used to improve performance as a last resort. However, ensembling multiple models finetuned from a single pretrained model has been not very effective; this could be due to the lack of diversity among ensemble members. This paper proposes Multi-Ticket Ensemble, which finetunes different subnetworks of a single pretrained model and ensembles them. We empirically demonstrated that winning-ticket subnetworks produced more diverse predictions than dense networks, and their ensemble outperformed the standard ensemble on some tasks.

Ryo Fujii, Makoto Morishita, Kazuki Yano et al.

With the advancement of automated software engineering, research focus is increasingly shifting toward practical tasks reflecting the day-to-day work of software engineers. Among these tasks, software migration, a critical process of adapting code to evolving environments, has been largely overlooked. In this study, we introduce TimeMachine-bench, a benchmark designed to evaluate software migration in real-world Python projects. Our benchmark consists of GitHub repositories whose tests begin to fail in response to dependency updates. The construction process is fully automated, enabling live updates of the benchmark. Furthermore, we curated a human-verified subset to ensure problem solvability. We evaluated agent-based baselines built on top of 11 models, including both strong open-weight and state-of-the-art LLMs on this verified subset. Our results indicated that, while LLMs show some promise for migration tasks, they continue to face substantial reliability challenges, including spurious solutions that exploit low test coverage and unnecessary edits stemming from suboptimal tool-use strategies. Our dataset and implementation are available at https://github.com/tohoku-nlp/timemachine-bench.

Yunmeng Li, Jun Suzuki, Makoto Morishita et al.

In this paper, we describe the development of a communication support system that detects erroneous translations to facilitate crosslingual communications due to the limitations of current machine chat translation methods. We trained an error detector as the baseline of the system and constructed a new Japanese-English bilingual chat corpus, BPersona-chat, which comprises multiturn colloquial chats augmented with crowdsourced quality ratings. The error detector can serve as an encouraging foundation for more advanced erroneous translation detection systems.

Atsushi Shirafuji, Yusuke Oda, Jun Suzuki et al.

A less complex and more straightforward program is a crucial factor that enhances its maintainability and makes writing secure and bug-free programs easier. However, due to its heavy workload and the risks of breaking the working programs, programmers are reluctant to do code refactoring, and thus, it also causes the loss of potential learning experiences. To mitigate this, we demonstrate the application of using a large language model (LLM), GPT-3.5, to suggest less complex versions of the user-written Python program, aiming to encourage users to learn how to write better programs. We propose a method to leverage the prompting with few-shot examples of the LLM by selecting the best-suited code refactoring examples for each target programming problem based on the prior evaluation of prompting with the one-shot example. The quantitative evaluation shows that 95.68% of programs can be refactored by generating 10 candidates each, resulting in a 17.35% reduction in the average cyclomatic complexity and a 25.84% decrease in the average number of lines after filtering only generated programs that are semantically correct. Furthermore, the qualitative evaluation shows outstanding capability in code formatting, while unnecessary behaviors such as deleting or translating comments are also observed.

Makoto Morishita, Jun Suzuki, Masaaki Nagata

Although a machine translation model trained with a large in-domain parallel corpus achieves remarkable results, it still works poorly when no in-domain data are available. This situation restricts the applicability of machine translation when the target domain's data are limited. However, there is great demand for high-quality domain-specific machine translation models for many domains. We propose a framework that efficiently and effectively collects parallel sentences in a target domain from the web with the help of crowdworkers. With the collected parallel data, we can quickly adapt a machine translation model to the target domain. Our experiments show that the proposed method can collect target-domain parallel data over a few days at a reasonable cost. We tested it with five domains, and the domain-adapted model improved the BLEU scores to +19.7 by an average of +7.8 points compared to a general-purpose translation model.

Atsushi Shirafuji, Yutaka Watanobe, Takumi Ito et al.

Using large language models (LLMs) for source code has recently gained attention. LLMs, such as Transformer-based models like Codex and ChatGPT, have been shown to be highly capable of solving a wide range of programming problems. However, the extent to which LLMs understand problem descriptions and generate programs accordingly or just retrieve source code from the most relevant problem in training data based on superficial cues has not been discovered yet. To explore this research question, we conduct experiments to understand the robustness of several popular LLMs, CodeGen and GPT-3.5 series models, capable of tackling code generation tasks in introductory programming problems. Our experimental results show that CodeGen and Codex are sensitive to the superficial modifications of problem descriptions and significantly impact code generation performance. Furthermore, we observe that Codex relies on variable names, as randomized variables decrease the solved rate significantly. However, the state-of-the-art (SOTA) models, such as InstructGPT and ChatGPT, show higher robustness to superficial modifications and have an outstanding capability for solving programming problems. This highlights the fact that slight modifications to the prompts given to the LLMs can greatly affect code generation performance, and careful formatting of prompts is essential for high-quality code generation, while the SOTA models are becoming more robust to perturbations.

Shiki Sato, Reina Akama, Hiroki Ouchi et al.

Avoiding the generation of responses that contradict the preceding context is a significant challenge in dialogue response generation. One feasible method is post-processing, such as filtering out contradicting responses from a resulting n-best response list. In this scenario, the quality of the n-best list considerably affects the occurrence of contradictions because the final response is chosen from this n-best list. This study quantitatively analyzes the contextual contradiction-awareness of neural response generation models using the consistency of the n-best lists. Particularly, we used polar questions as stimulus inputs for concise and quantitative analyses. Our tests illustrate the contradiction-awareness of recent neural response generation models and methodologies, followed by a discussion of their properties and limitations.

Shiki Sato, Yosuke Kishinami, Hiroaki Sugiyama et al.

Automation of dialogue system evaluation is a driving force for the efficient development of dialogue systems. This paper introduces the bipartite-play method, a dialogue collection method for automating dialogue system evaluation. It addresses the limitations of existing dialogue collection methods: (i) inability to compare with systems that are not publicly available, and (ii) vulnerability to cheating by intentionally selecting systems to be compared. Experimental results show that the automatic evaluation using the bipartite-play method mitigates these two drawbacks and correlates as strongly with human subjectivity as existing methods.

Mengyu Ye, Tatsuki Kuribayashi, Jun Suzuki et al.

Large language models (LLMs) take advantage of step-by-step reasoning instructions, e.g., chain-of-thought (CoT) prompting. Building on this, their ability to perform CoT-style reasoning robustly is of interest from a probing perspective. In this study, we inspect the step-by-step reasoning ability of LLMs with a focus on negation, which is a core linguistic phenomenon that is difficult to process. In particular, we introduce several controlled settings (e.g., reasoning in case of fictional entities) to evaluate the logical reasoning abilities of the models. We observed that dozens of modern LLMs were not robust against lexical negation (e.g., plausible ->implausible) when performing CoT-style reasoning, and the results highlight unique limitations in each LLM family.

Kazuki Yano, Shun Kiyono, Sosuke Kobayashi et al.

We investigate the role of learning rate scheduling in the large-scale pre-training of large language models, focusing on its influence on downstream performance after supervised fine-tuning (SFT). Decay-based learning rate schedulers are widely used to minimize pre-training loss. However, despite their widespread use, how these schedulers affect performance after SFT remains underexplored. In this paper, we examine Warmup-Stable-Only (WSO), which maintains a constant learning rate after warmup without any decay. Through experiments with 1B and 8B parameter models, we show that WSO consistently outperforms decay-based schedulers in terms of performance after SFT, even though decay-based schedulers may exhibit better performance after pre-training. The result also holds across different regimes with mid-training and over-training. Loss landscape analysis further reveals that decay-based schedulers lead models into sharper minima, whereas WSO preserves flatter minima that support adaptability. These findings indicate that applying LR decay to improve pre-training metrics may compromise downstream adaptability. Our work also provides practical guidance for training and model release strategies, highlighting that pre-training models with WSO enhances their adaptability for downstream tasks.

Yunmeng Li, Jun Suzuki, Makoto Morishita et al.

The complexities of chats pose significant challenges for machine translation models. Recognizing the need for a precise evaluation metric to address the issues of chat translation, this study introduces Multidimensional Quality Metrics for Chat Translation (MQM-Chat). Through the experiments of five models using MQM-Chat, we observed that all models generated certain fundamental errors, while each of them has different shortcomings, such as omission, overly correcting ambiguous source content, and buzzword issues, resulting in the loss of stylized information. Our findings underscore the effectiveness of MQM-Chat in evaluating chat translation, emphasizing the importance of stylized content and dialogue consistency for future studies.

Ikuya Yamada, Wataru Ikeda, Ko Yoshida et al.

We present an open deep research system for long-form question answering, selected as a winning system in the text-to-text track of the MMU-RAG competition at NeurIPS 2025. The system combines an open-source large language model (LLM) with an open web search API to perform iterative retrieval, reasoning, and synthesis in real-world open-domain settings. To enhance reasoning quality, we apply preference tuning based on LLM-as-a-judge feedback that evaluates multiple aspects, including clarity, insightfulness, and factuality. Our experimental results show that the proposed method consistently improves answer quality across all three aspects. Our source code is publicly available at https://github.com/efficient-deep-research/efficient-deep-research.

Yunmeng Li, Jun Suzuki, Makoto Morishita et al.

Machine translation models are still inappropriate for translating chats, despite the popularity of translation software and plug-in applications. The complexity of dialogues poses significant challenges and can hinder crosslingual communication. Instead of pursuing a flawless translation system, a more practical approach would be to issue warning messages about potential mistranslations to reduce confusion. However, it is still unclear how individuals perceive these warning messages and whether they benefit the crowd. This paper tackles to investigate this question and demonstrates the warning messages' contribution to making chat translation systems effective.

Subaru Kimura, Ryota Tanaka, Shumpei Miyawaki et al.

We explore visual prompt injection (VPI) that maliciously exploits the ability of large vision-language models (LVLMs) to follow instructions drawn onto the input image. We propose a new VPI method, "goal hijacking via visual prompt injection" (GHVPI), that swaps the execution task of LVLMs from an original task to an alternative task designated by an attacker. The quantitative analysis indicates that GPT-4V is vulnerable to the GHVPI and demonstrates a notable attack success rate of 15.8%, which is an unignorable security risk. Our analysis also shows that successful GHVPI requires high character recognition capability and instruction-following ability in LVLMs.

Hwichan Kim, Jun Suzuki, Tosho Hirasawa et al.

Multilingual large language models (MLLMs), trained on multilingual balanced data, demonstrate better zero-shot learning performance in non-English languages compared to large language models trained on English-dominant data. However, the disparity in performance between English and non-English languages remains a challenge yet to be fully addressed. A distinctive characteristic of MLLMs is their high-quality translation capabilities, indicating an acquired proficiency in aligning between languages. This study explores how to enhance the zero-shot performance of MLLMs in non-English languages by leveraging their alignment capability between English and non-English languages. To achieve this, we first analyze the behavior of MLLMs when performing translation and reveal that there are large magnitude features that play a critical role in the translation process. Inspired by these findings, we retain the weights associated with operations involving the large magnitude features and prune other weights to force MLLMs to rely on these features for tasks beyond translation. We empirically demonstrate that this pruning strategy can enhance the MLLMs' performance in non-English language.

Shinnosuke Nozue, Yuto Nakano, Yotaro Watanabe et al.

Current approaches to developing persuasive dialogue agents often rely on a limited set of predefined persuasive strategies that fail to capture the complexity of real-world interactions. We applied a cross-disciplinary approach to develop a framework for designing persuasive dialogue agents that draws on proven strategies from social psychology, behavioral economics, and communication theory. We validated our proposed framework through experiments on two distinct datasets: the Persuasion for Good dataset, which represents a specific in-domain scenario, and the DailyPersuasion dataset, which encompasses a wide range of scenarios. The proposed framework achieved strong results for both datasets and demonstrated notable improvement in the persuasion success rate as well as promising generalizability. Notably, the proposed framework also excelled at persuading individuals with initially low intent, which addresses a critical challenge for persuasive dialogue agents.

Ryosuke Matsuda, Keito Kudo, Haruto Yoshida et al.

This paper proposes the synthetic long-video meta-evaluation (SLVMEval), a benchmark for meta-evaluating text-to-video (T2V) evaluation systems. The proposed SLVMEval benchmark focuses on assessing these systems on videos of up to 10,486 s (approximately 3 h). The benchmark targets a fundamental requirement, namely, whether the systems can accurately assess video quality in settings that are easy for humans to assess. We adopt a pairwise comparison-based meta-evaluation framework. Building on dense video-captioning datasets, we synthetically degrade source videos to create controlled "high-quality versus low-quality" pairs across 10 distinct aspects. Then, we employ crowdsourcing to filter and retain only those pairs in which the degradation is clearly perceptible, thereby establishing an effective final testbed. Using this testbed, we assess the reliability of existing evaluation systems in ranking these pairs. Experimental results demonstrate that human evaluators can identify the better long video with 84.7%-96.8% accuracy, and in nine of the 10 aspects, the accuracy of these systems falls short of human assessment, revealing weaknesses in text-to-long-video evaluation.

Mengyu Ye, Ryosuke Takahashi, Keito Kudo et al.

Masked diffusion language models (MDLMs) have emerged as a promising alternative to dominant autoregressive approaches. Although they achieve competitive performance on several tasks, a substantial gap remains in open-ended text generation. We hypothesize that one cause of this gap is that strict positional prediction makes MDLM decoding highly sensitive to token misalignment, and we show through controlled interventions that a one-position shift can severely disrupt semantics. This observation suggests that enforcing strict positional supervision during training is misaligned with the irreversible denoising dynamics of MDLM decoding. Motivated by this mismatch, we adopt an alignment-flexible supervision strategy during fine-tuning. Specifically, we introduce a special token <slack> via the connectionist temporal classification objective. We apply this approach to the widely used MDLM model and conduct experiments on five open-ended text generation benchmarks. Our method consistently outperforms the original model and improves robustness to positional shifts, indicating that relaxing strict positional supervision is an important factor in improving generation quality in MDLMs.

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.

Masahiro Kaneko, Masato Mita, Shun Kiyono et al.

This paper investigates how to effectively incorporate a pre-trained masked language model (MLM), such as BERT, into an encoder-decoder (EncDec) model for grammatical error correction (GEC). The answer to this question is not as straightforward as one might expect because the previous common methods for incorporating a MLM into an EncDec model have potential drawbacks when applied to GEC. For example, the distribution of the inputs to a GEC model can be considerably different (erroneous, clumsy, etc.) from that of the corpora used for pre-training MLMs; however, this issue is not addressed in the previous methods. Our experiments show that our proposed method, where we first fine-tune a MLM with a given GEC corpus and then use the output of the fine-tuned MLM as additional features in the GEC model, maximizes the benefit of the MLM. The best-performing model achieves state-of-the-art performances on the BEA-2019 and CoNLL-2014 benchmarks. Our code is publicly available at: https://github.com/kanekomasahiro/bert-gec.

Sho Yokoi, Ryo Takahashi, Reina Akama et al.

A key principle in assessing textual similarity is measuring the degree of semantic overlap between two texts by considering the word alignment. Such alignment-based approaches are intuitive and interpretable; however, they are empirically inferior to the simple cosine similarity between general-purpose sentence vectors. To address this issue, we focus on and demonstrate the fact that the norm of word vectors is a good proxy for word importance, and their angle is a good proxy for word similarity. Alignment-based approaches do not distinguish them, whereas sentence-vector approaches automatically use the norm as the word importance. Accordingly, we propose a method that first decouples word vectors into their norm and direction, and then computes alignment-based similarity using earth mover's distance (i.e., optimal transport cost), which we refer to as word rotator's distance. Besides, we find how to grow the norm and direction of word vectors (vector converter), which is a new systematic approach derived from sentence-vector estimation methods. On several textual similarity datasets, the combination of these simple proposed methods outperformed not only alignment-based approaches but also strong baselines. The source code is available at https://github.com/eumesy/wrd

Masato Hagiwara, Takumi Ito, Tatsuki Kuribayashi et al.

Language technologies play a key role in assisting people with their writing. Although there has been steady progress in e.g., grammatical error correction (GEC), human writers are yet to benefit from this progress due to the high development cost of integrating with writing software. We propose TEASPN, a protocol and an open-source framework for achieving integrated writing assistance environments. The protocol standardizes the way writing software communicates with servers that implement such technologies, allowing developers and researchers to integrate the latest developments in natural language processing (NLP) with low cost. As a result, users can enjoy the integrated experience in their favorite writing software. The results from experiments with human participants show that users use a wide range of technologies and rate their writing experience favorably, allowing them to write more fluent text.

Sho Takase, Jun Suzuki, Masaaki Nagata

This paper proposes a state-of-the-art recurrent neural network (RNN) language model that combines probability distributions computed not only from a final RNN layer but also from middle layers. Our proposed method raises the expressive power of a language model based on the matrix factorization interpretation of language modeling introduced by Yang et al. (2018). The proposed method improves the current state-of-the-art language model and achieves the best score on the Penn Treebank and WikiText-2, which are the standard benchmark datasets. Moreover, we indicate our proposed method contributes to two application tasks: machine translation and headline generation. Our code is publicly available at: https://github.com/nttcslab-nlp/doc_lm.

Daiki Shiono, Shumpei Miyawaki, Ryota Tanaka et al.

Following the initial flourishing of large language models (LLMs), there has been a surge in proposed large vision-language models (LVLMs) that integrate LLMs with vision capabilities. However, it has been observed that LVLMs, after tuning to visual instruction using commonly used training datasets, often fail to exhibit the instruction-following ability that was present in the LLM before integration, leading to results in which they do not follow task instructions as expected. This study quantitatively demonstrates that LVLMs' instruction-following ability declines after fine-tuning and analyzes its underlying causes. In particular, we constructed new training datasets highlighting whether the output format is specified. Then, we investigated how explicitly indicating the output format during fine-tuning affects LVLMs' instruction-following ability. Our quantitative evaluation confirmed that LVLMs' instruction-following ability declines after fine-tuning with commonly used datasets. Furthermore, we found that LVLMs trained with datasets, including instructions on output format, tend to follow instructions more accurately than models that do not. These findings suggest that including samples with instructions on output format during (visual) instruction tuning may help mitigate the decline in instruction-following abilities.

Kazuki Yano, Jun Suzuki, Shinji Watanabe

Adapting pre-trained text Large Language Models (LLMs) into Speech Language Models (Speech LMs) via continual pretraining on speech data is promising, but often degrades the original text capabilities. We propose Multimodal Depth Upscaling, an extension of an emerging strategy in continual LLM pre-training, where new transformer layers are inserted into a frozen text LLM and only the added layers are trained on speech data. Experiments with SmolLM2-360M and SmolLM2-1.7B on 48k hours of English Automatic Speech Recognition (ASR) data show that depth up-scaling achieves ASR comparable to full fine-tuning while causing far less text degradation than both full fine-tuning and Low-Rank Adaptation (LoRA). We further show that incorporating E-Branchformer, an architecture designed for speech recognition, as the inserted layers achieves ASR that matches or surpasses full fine-tuning on the larger model while reducing text degradation by over 75% with 60% fewer trainable parameters.

Keigo Shibata, Kazuki Yano, Ryosuke Takahashi et al.

This paper discusses the internal behavior of Transformer language models. Many recent pre-trained models have been reported to exhibit only slight changes in the angular distance between the input and output hidden state vectors in the middle Transformer layers, despite a disproportionately large ``jump'' in the angular distance occurring in or around the final Transformer layer. To characterize this, we first introduce a quantitative metric for the jump strength around the final layer, and then demonstrate its prevalence across many open-weight models, as well as its amplification throughout pre-training. Assuming such jumps indicate an undesirable property, we propose the jump-suppressing regularizer (JREG) which penalizes this jump during pre-training, thereby encouraging more balanced capability usage across the middle layers. Empirical evaluations of three model sizes of Llama-based models, trained with the proposed JREG method, reveal improved task performance compared to the baseline without altering the model architecture.

Sho Takase, Shun Kiyono, Sosuke Kobayashi et al.

Loss spikes often occur during pre-training of large language models. The spikes degrade the performance of large language models and sometimes ruin the pre-training. Since the pre-training needs a vast computational budget, we should avoid such spikes. Based on the assumption that the loss spike is caused by the sudden growth of the gradient norm, we explore factors to keep the gradient norm small through an analysis of the spectral norms of the Jacobian matrices for the sub-layers. Our findings suggest that stabilizing the pre-training process requires two conditions: small sub-layers and large shortcut. We conduct various experiments to empirically verify our theoretical analyses. Experimental results demonstrate that methods satisfying the conditions effectively prevent loss spikes during pre-training.

Ryota Tanaka, Taichi Iki, Taku Hasegawa et al.

We aim to develop a retrieval-augmented generation (RAG) framework that answers questions over a corpus of visually-rich documents presented in mixed modalities (e.g., charts, tables) and diverse formats (e.g., PDF, PPTX). In this paper, we introduce a new RAG framework, VDocRAG, which can directly understand varied documents and modalities in a unified image format to prevent missing information that occurs by parsing documents to obtain text. To improve the performance, we propose novel self-supervised pre-training tasks that adapt large vision-language models for retrieval by compressing visual information into dense token representations while aligning them with textual content in documents. Furthermore, we introduce OpenDocVQA, the first unified collection of open-domain document visual question answering datasets, encompassing diverse document types and formats. OpenDocVQA provides a comprehensive resource for training and evaluating retrieval and question answering models on visually-rich documents in an open-domain setting. Experiments show that VDocRAG substantially outperforms conventional text-based RAG and has strong generalization capability, highlighting the potential of an effective RAG paradigm for real-world documents.

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.

Takumi Ito, Kees van Deemter, Jun Suzuki

A number of automatic evaluation metrics have been proposed for natural language generation systems. The most common approach to automatic evaluation is the use of a reference-based metric that compares the model's output with gold-standard references written by humans. However, it is expensive to create such references, and for some tasks, such as response generation in dialogue, creating references is not a simple matter. Therefore, various reference-free metrics have been developed in recent years. In this survey, which intends to cover the full breadth of all NLG tasks, we investigate the most commonly used approaches, their application, and their other uses beyond evaluating models. The survey concludes by highlighting some promising directions for future research.

Keito Kudo, Haruki Nagasawa, Jun Suzuki et al.

This paper proposes a practical multimodal video summarization task setting and a dataset to train and evaluate the task. The target task involves summarizing a given video into a predefined number of keyframe-caption pairs and displaying them in a listable format to grasp the video content quickly. This task aims to extract crucial scenes from the video in the form of images (keyframes) and generate corresponding captions explaining each keyframe's situation. This task is useful as a practical application and presents a highly challenging problem worthy of study. Specifically, achieving simultaneous optimization of the keyframe selection performance and caption quality necessitates careful consideration of the mutual dependence on both preceding and subsequent keyframes and captions. To facilitate subsequent research in this field, we also construct a dataset by expanding upon existing datasets and propose an evaluation framework. Furthermore, we develop two baseline systems and report their respective performance.

Kazuki Yano, Takumi Ito, Jun Suzuki

Pre-training large language models (LLMs) faces significant memory challenges due to the large size of model parameters. We introduce STaged parameter-Efficient Pre-training (STEP), which integrates parameter-efficient tuning techniques with model growth. We conduct experiments on pre-training LLMs of various sizes and demonstrate that STEP achieves up to a 53.9% reduction in maximum memory requirements compared to vanilla pre-training while maintaining equivalent performance. Furthermore, we show that the model by STEP performs comparably to vanilla pre-trained models on downstream tasks after instruction tuning.

Mengyu Ye, Tatsuki Kuribayashi, Goro Kobayashi et al.

Interpreting the internal process of neural models has long been a challenge. This challenge remains relevant in the era of large language models (LLMs) and in-context learning (ICL); for example, ICL poses a new issue of interpreting which example in the few-shot examples contributed to identifying/solving the task. To this end, in this paper, we design synthetic diagnostic tasks of inductive reasoning, inspired by the generalization tests typically adopted in psycholinguistics. Here, most in-context examples are ambiguous w.r.t. their underlying rule, and one critical example disambiguates it. The question is whether conventional input attribution (IA) methods can track such a reasoning process, i.e., identify the influential example, in ICL. Our experiments provide several practical findings; for example, a certain simple IA method works the best, and the larger the model, the generally harder it is to interpret the ICL with gradient-based IA methods.

Kazuki Yano, Sho Takase, Sosuke Kobayashi et al.

As Large Language Models (LLMs) gain widespread practical application, providing the model family of different parameter sizes has become standard practice to address diverse computational requirements. Conventionally, each model in a family is trained independently, resulting in computational costs that scale additively with the number of models. We propose an efficient method for constructing the model family through progressive training, where smaller models are incrementally expanded to larger sizes to create a complete model family. Through extensive experiments with a model family ranging from 1B to 8B parameters, we demonstrate that our method reduces computational costs by approximately 25% while maintaining comparable performance to independently trained models. Furthermore, by strategically adjusting maximum learning rates based on model size, our method outperforms the independent training across various metrics. Beyond performance gains, our approach offers an additional advantage: models in our family tend to yield more consistent behavior across different model sizes.

Mengyu Ye, Jun Suzuki, Tatsuro Inaba et al.

Recent interpretability work on large language models (LLMs) has been increasingly dominated by a feature-discovery approach with the help of proxy modules. Then, the quality of features learned by, e.g., sparse auto-encoders (SAEs), is evaluated. This paradigm naturally raises a critical question: do such learned features have better properties than those already represented within the original model parameters, and unfortunately, only a few studies have made such comparisons systematically so far. In this work, we revisit the interpretability of feature vectors stored in feed-forward (FF) layers, given the perspective of FF as key-value memories, with modern interpretability benchmarks. Our extensive evaluation revealed that SAE and FFs exhibits a similar range of interpretability, although SAEs displayed an observable but minimal improvement in some aspects. Furthermore, in certain aspects, surprisingly, even vanilla FFs yielded better interpretability than the SAEs, and features discovered in SAEs and FFs diverged. These bring questions about the advantage of SAEs from both perspectives of feature quality and faithfulness, compared to directly interpreting FF feature vectors, and FF key-value parameters serve as a strong baseline in modern interpretability research.

Wataru Ikeda, Kazuki Yano, Ryosuke Takahashi et al.

This study investigates the layerwise importance of feed-forward networks (FFNs) in Transformer-based language models during pretraining. We introduce an experimental approach that, while maintaining the total parameter count, increases the FFN dimensions in some layers and completely removes the FFNs from other layers. Furthermore, since our focus is on the importance of FFNs during pretraining, we train models from scratch to examine whether the importance of FFNs varies depending on their layer positions, rather than using publicly available pretrained models as is frequently done. Through comprehensive evaluations of models with varying sizes (285M, 570M, and 1.2B parameters) and layer counts (12, 24, and 40 layers), we demonstrate that concentrating FFNs in 70% of the consecutive middle layers consistently outperforms standard configurations for multiple downstream tasks.

Ikumi Numaya, Shoji Moriya, Shiki Sato et al.

Recent advancements in dialogue generation have broadened the scope of human-bot interactions, enabling not only contextually appropriate responses but also the analysis of human affect and sensitivity. While prior work has suggested that stylistic similarity between user and system may enhance user impressions, the distinction between subjective and objective similarity is often overlooked. To investigate this issue, we introduce a novel dataset that includes users' preferences, subjective stylistic similarity based on users' own perceptions, and objective stylistic similarity annotated by third party evaluators in open-domain dialogue settings. Analysis using the constructed dataset reveals a strong positive correlation between subjective stylistic similarity and user preference. Furthermore, our analysis suggests an important finding: users' subjective stylistic similarity differs from third party objective similarity. This underscores the importance of distinguishing between subjective and objective evaluations and understanding the distinct aspects each captures when analyzing the relationship between stylistic similarity and user preferences. The dataset presented in this paper is available online.

Ryohei Kamei, Daiki Shiono, Reina Akama et al.

With the remarkable development of large language models (LLMs), ensuring the factuality of output has become a challenge. However, having all the contents of the response with given knowledge or facts is not necessarily a good thing in dialogues. This study aimed to achieve both attractiveness and factuality in a dialogue response for which a task was set to predict sentences that do not require factual correctness judgment such as agreeing, or personal opinions/feelings. We created a dataset, dialogue dataset annotated with fact-check-needed label (DDFC), for this task via crowdsourcing, and classification tasks were performed on several models using this dataset. The model with the highest classification accuracy could yield about 88% accurate classification results.

Shiki Sato, Reina Akama, Jun Suzuki et al.

Mitigating the generation of contradictory responses poses a substantial challenge in dialogue response generation. The quality and quantity of available contradictory response data play a vital role in suppressing these contradictions, offering two significant benefits. First, having access to large contradiction data enables a comprehensive examination of their characteristics. Second, data-driven methods to mitigate contradictions may be enhanced with large-scale contradiction data for training. Nevertheless, no attempt has been made to build an extensive collection of model-generated contradictory responses. In this paper, we build a large dataset of response generation models' contradictions for the first time. Then, we acquire valuable insights into the characteristics of model-generated contradictions through an extensive analysis of the collected responses. Lastly, we also demonstrate how this dataset substantially enhances the performance of data-driven contradiction suppression methods.

Ryota Tanaka, Taichi Iki, Kyosuke Nishida et al.

We study the problem of completing various visual document understanding (VDU) tasks, e.g., question answering and information extraction, on real-world documents through human-written instructions. To this end, we propose InstructDoc, the first large-scale collection of 30 publicly available VDU datasets, each with diverse instructions in a unified format, which covers a wide range of 12 tasks and includes open document types/formats. Furthermore, to enhance the generalization performance on VDU tasks, we design a new instruction-based document reading and understanding model, InstructDr, that connects document images, image encoders, and large language models (LLMs) through a trainable bridging module. Experiments demonstrate that InstructDr can effectively adapt to new VDU datasets, tasks, and domains via given instructions and outperforms existing multimodal LLMs and ChatGPT without specific training.

Makoto Morishita, Katsuki Chousa, Jun Suzuki et al.

Most current machine translation models are mainly trained with parallel corpora, and their translation accuracy largely depends on the quality and quantity of the corpora. Although there are billions of parallel sentences for a few language pairs, effectively dealing with most language pairs is difficult due to a lack of publicly available parallel corpora. This paper creates a large parallel corpus for English-Japanese, a language pair for which only limited resources are available, compared to such resource-rich languages as English-German. It introduces a new web-based English-Japanese parallel corpus named JParaCrawl v3.0. Our new corpus contains more than 21 million unique parallel sentence pairs, which is more than twice as many as the previous JParaCrawl v2.0 corpus. Through experiments, we empirically show how our new corpus boosts the accuracy of machine translation models on various domains. The JParaCrawl v3.0 corpus will eventually be publicly available online for research purposes.

Hiroki Ouchi, Jun Suzuki, Sosuke Kobayashi et al.

Interpretable rationales for model predictions are crucial in practical applications. We develop neural models that possess an interpretable inference process for dependency parsing. Our models adopt instance-based inference, where dependency edges are extracted and labeled by comparing them to edges in a training set. The training edges are explicitly used for the predictions; thus, it is easy to grasp the contribution of each edge to the predictions. Our experiments show that our instance-based models achieve competitive accuracy with standard neural models and have the reasonable plausibility of instance-based explanations.

Shun Kiyono, Sosuke Kobayashi, Jun Suzuki et al.

Position representation is crucial for building position-aware representations in Transformers. Existing position representations suffer from a lack of generalization to test data with unseen lengths or high computational cost. We investigate shifted absolute position embedding (SHAPE) to address both issues. The basic idea of SHAPE is to achieve shift invariance, which is a key property of recent successful position representations, by randomly shifting absolute positions during training. We demonstrate that SHAPE is empirically comparable to its counterpart while being simpler and faster.

Yasufumi Taniguchi, Hiroki Nakayama, Kubo Takahiro et al.

Text-to-SQL is a crucial task toward developing methods for understanding natural language by computers. Recent neural approaches deliver excellent performance; however, models that are difficult to interpret inhibit future developments. Hence, this study aims to provide a better approach toward the interpretation of neural models. We hypothesize that the internal behavior of models at hand becomes much easier to analyze if we identify the detailed performance of schema linking simultaneously as the additional information of the text-to-SQL performance. We provide the ground-truth annotation of schema linking information onto the Spider dataset. We demonstrate the usefulness of the annotated data and how to analyze the current state-of-the-art neural models.

Sewon Min, Jordan Boyd-Graber, Chris Alberti et al.

We review the EfficientQA competition from NeurIPS 2020. The competition focused on open-domain question answering (QA), where systems take natural language questions as input and return natural language answers. The aim of the competition was to build systems that can predict correct answers while also satisfying strict on-disk memory budgets. These memory budgets were designed to encourage contestants to explore the trade-off between storing retrieval corpora or the parameters of learned models. In this report, we describe the motivation and organization of the competition, review the best submissions, and analyze system predictions to inform a discussion of evaluation for open-domain QA.