Masafumi Oyamada

Masafumi Enomoto, Ryoma Obara, Haochen Zhang et al.

HTML observations in LLM-based web agents are extremely long, and while many reduction methods have been proposed, it remains unclear which methods reduce overall agent latency while maintaining performance. The main obstacle is the high cost of end-to-end evaluation: in our experiments, evaluating 11 methods across 32 configurations on 33 tasks of WorkArena L1 required 232.4 cumulative hours. To address this, we propose a lightweight evaluation framework based on the Minimal Failure Set (MFS), the minimal set of HTML elements whose removal causes task failure. We define coverage as the fraction of instances in which a reduction method fully retains the MFS, which serves as a proxy metric that requires neither web access nor LLM inference. We validate that coverage strongly correlates with end-to-end success rate, with over 100$\times$ speedup in cumulative evaluation time on both benchmarks. Using this framework, we find that extractive HTML reduction methods require either high computation cost or domain-specific optimization to reduce agent latency while maintaining performance. Building on this, we optimize a pruning program on MFS training data, achieving 2.2$\times$ faster per-step latency on WorkArena L1 while retaining 84\% of the original success rate, and 3.1$\times$ faster on WebLinx while retaining 89\%.

Haochen Zhang, Yuyang Dong, Chuan Xiao et al.

Large Language Models (LLMs), typified by OpenAI's GPT, have marked a significant advancement in artificial intelligence. Trained on vast amounts of text data, LLMs are capable of understanding and generating human-like text across a diverse range of topics. This study expands on the applications of LLMs, exploring their potential in data preprocessing, a critical stage in data mining and analytics applications. Aiming at tabular data, we delve into the applicability of state-of-the-art LLMs such as GPT-4 and GPT-4o for a series of preprocessing tasks, including error detection, data imputation, schema matching, and entity matching. Alongside showcasing the inherent capabilities of LLMs, we highlight their limitations, particularly in terms of computational expense and inefficiency. We propose an LLM-based framework for data preprocessing, which integrates cutting-edge prompt engineering techniques, coupled with traditional methods like contextualization and feature selection, to improve the performance and efficiency of these models. The effectiveness of LLMs in data preprocessing is evaluated through an experimental study spanning a variety of public datasets. GPT-4 emerged as a standout, achieving 100\% accuracy or F1 score on 4 of these datasets, suggesting LLMs' immense potential in these tasks. Despite certain limitations, our study underscores the promise of LLMs in this domain and anticipates future developments to overcome current hurdles.

Yuyang Dong, Chuan Xiao, Takuma Nozawa et al.

Due to the usefulness in data enrichment for data analysis tasks, joinable table discovery has become an important operation in data lake management. Existing approaches target equi-joins, the most common way of combining tables for creating a unified view, or semantic joins, which tolerate misspellings and different formats to deliver more join results. They are either exact solutions whose running time is linear in the sizes of query column and target table repository or approximate solutions lacking precision. In this paper, we propose Deepjoin, a deep learning model for accurate and efficient joinable table discovery. Our solution is an embedding-based retrieval, which employs a pre-trained language model (PLM) and is designed as one framework serving both equi- and semantic joins. We propose a set of contextualization options to transform column contents to a text sequence. The PLM reads the sequence and is fine-tuned to embed columns to vectors such that columns are expected to be joinable if they are close to each other in the vector space. Since the output of the PLM is fixed in length, the subsequent search procedure becomes independent of the column size. With a state-of-the-art approximate nearest neighbor search algorithm, the search time is logarithmic in the repository size. To train the model, we devise the techniques for preparing training data as well as data augmentation. The experiments on real datasets demonstrate that by training on a small subset of a corpus, Deepjoin generalizes to large datasets and its precision consistently outperforms other approximate solutions'. Deepjoin is even more accurate than an exact solution to semantic joins when evaluated with labels from experts. Moreover, when equipped with a GPU, Deepjoin is up to two orders of magnitude faster than existing solutions.

Haochen Zhang, Yuyang Dong, Chuan Xiao et al.

This paper explores the utilization of LLMs for data preprocessing (DP), a crucial step in the data mining pipeline that transforms raw data into a clean format conducive to easy processing. Whereas the use of LLMs has sparked interest in devising universal solutions to DP, recent initiatives in this domain typically rely on GPT APIs, raising inevitable data breach concerns. Unlike these approaches, we consider instruction-tuning local LLMs (7 -- 13B models) as universal DP task solvers that operate on a local, single, and low-priced GPU, ensuring data security and enabling further customization. We select a collection of datasets across four representative DP tasks and construct instruction tuning data using data configuration, knowledge injection, and reasoning data distillation techniques tailored to DP. By tuning Mistral-7B, Llama 3-8B, and OpenOrca-Platypus2-13B, our models, namely, Jellyfish-7B/8B/13B, deliver competitiveness compared to GPT-3.5/4 models and strong generalizability to unseen tasks while barely compromising the base models' abilities in NLP tasks. Meanwhile, Jellyfish offers enhanced reasoning capabilities compared to GPT-3.5. Our models are available at: https://huggingface.co/NECOUDBFM/Jellyfish . Our instruction dataset is available at: https://huggingface.co/datasets/NECOUDBFM/Jellyfish-Instruct .

Yoichi Ishibashi, Taro Yano, Masafumi Oyamada

AlphaEvolve and FunSearch have demonstrated the potential of combining large language models (LLMs) with evolutionary search for automated algorithm discovery. However, discovery success is shaped not only by model capability but also significantly by the design of the execution infrastructure, i.e., the harness. This paper investigates effective harness design through three questions: under a fixed token budget, is it better to produce many algorithms with brief thought or fewer algorithms with deeper thought? How should the harness handle evaluation hacks, where generated programs exploit the scoring function? And how can agents that require full filesystem access execute safely in parallel? Using Vesper, an algorithm discovery framework that incorporates harness improvements addressing these questions, we evaluate on Circle Packing under the same token budget. Interestingly, generating fewer algorithms while thinking more deeply about each one achieved higher scores. That is, scaling the quality of each individual is more budget-efficient than scaling the number of evolutionary generations. Surprisingly, more capable models produced evaluation hacks at higher rates, making hack detection increasingly necessary as models scale.

Masafumi Enomoto, Ryoma Obara, Haochen Zhang et al.

Web agents based on large language models (LLMs) rely on observations of web pages -- commonly represented as HTML -- as the basis for identifying available actions and planning subsequent steps. Prior work has treated the verbosity of HTML as an obstacle to performance and adopted observation reduction as a standard practice. We revisit this trend and demonstrate that the optimal observation representation depends on model capability and thinking token budget: (1) compact observations (accessibility trees) are preferable for lower-capability models, while detailed observations (HTML) are advantageous for higher-capability models; moreover, increasing thinking tokens further amplifies the benefit of HTML. (2) Our error analysis suggests that higher-capability models exploit layout information in HTML for better action grounding, while lower-capability models suffer from increased hallucination under longer inputs. We also find that incorporating observation history improves performance across most models and settings, and a diff-based representation offers a token-efficient alternative. Based on these findings, we suggest practical guidelines: adaptively select observation representations based on model capability and thinking token budget, and incorporate observation history using diff-based representations.

Masafumi Oyamada, Kunihiro Takeoka, Kosuke Akimoto et al.

What if a browser agent could learn your work simply by watching you do it? We present cotomi Act, a browser-based computer-using agent that combines reliable multi-step task execution with persistent organizational knowledge learned from user behavior. For execution, an agent scaffold with adaptive lazy observation, verbal-diff-based history compression, coarse-grained actions, and test-time scaling via best-of-N action selection achieves 80.4% on the 179-task WebArena human-evaluation subset, exceeding the reported 78.2% human baseline. For organizational knowledge, a behavior-to-knowledge pipeline passively observes the user's browsing and progressively abstracts it into artifacts (task boards, wiki) exposed through a shared workspace editable by both user and agent. A controlled proxy evaluation confirms that task success improves as behavior-derived knowledge accumulates. In our live demonstration, attendees interact with the system in a real browser, issuing tasks and observing end-to-end autonomous execution and shared knowledge management.

Shintaro Ozaki, Yuta Kato, Siyuan Feng et al.

Retrieval Augmented Generation (RAG) complements the knowledge of Large Language Models (LLMs) by leveraging external information to enhance response accuracy for queries. This approach is widely applied in several fields by taking its advantage of injecting the most up-to-date information, and researchers are focusing on understanding and improving this aspect to unlock the full potential of RAG in such high-stakes applications. However, despite the potential of RAG to address these needs, the mechanisms behind the confidence levels of its outputs remain underexplored. Our study focuses on the impact of RAG, specifically examining whether RAG improves the confidence of LLM outputs in the medical domain. We conduct this analysis across various configurations and models. We evaluate confidence by treating the model's predicted probability as its output and calculating several evaluation metrics which include calibration error method, entropy, the best probability, and accuracy. Experimental results across multiple datasets confirmed that certain models possess the capability to judge for themselves whether an inserted document relates to the correct answer. These results suggest that evaluating models based on their output probabilities determine whether they function as generators in the RAG framework. Our approach allows us to evaluate whether the models handle retrieved documents.

Jonathan Light, Wei Cheng, Benjamin Riviere et al.

Inference scaling methods for LLMs often rely on decomposing problems into steps (or groups of tokens), followed by sampling and selecting the best next steps. However, these steps and their sizes are often predetermined or manually designed based on domain knowledge. We propose dynamic decomposition, a method that adaptively and automatically partitions solution and reasoning traces into manageable steps during inference. By more effectively allocating compute -- particularly through subdividing challenging steps and prioritizing their sampling -- dynamic decomposition significantly improves inference efficiency. Experiments on benchmarks such as APPS, MATH, and LiveCodeBench demonstrate that dynamic decomposition outperforms static approaches, including token-level, sentence-level, and single-step decompositions, reducing the pass@10 error rate by 5.0%, 6.7%, and 10.5% respectively. These findings highlight the potential of dynamic decomposition to improve a wide range of inference scaling techniques.

Yusuke Yamauchi, Taro Yano, Masafumi Oyamada

As large language models (LLMs) continue to advance, reliable evaluation methods are essential particularly for open-ended, instruction-following tasks. LLM-as-a-Judge enables automatic evaluation using LLMs as evaluators, but its reliability remains uncertain. In this work, we analyze key factors affecting its trustworthiness, focusing on alignment with human judgments and evaluation consistency. Using BIGGENBench and EvalBiasBench, we study the effects of evaluation design, decoding strategies, and Chain-of-Tought (CoT) reasoning in evaluation. Our results show that evaluation criteria are critical for reliability, non-deterministic sampling improves alignment with human preferences over deterministic evaluation, and CoT reasoning offers minimal gains when clear evaluation criteria are present.

Yuyang Dong, Nobuhiro Ueda, Krisztián Boros et al.

With the increasing adoption of Large Language Models (LLMs) and Vision-Language Models (VLMs), rich document analysis technologies for applications like Retrieval-Augmented Generation (RAG) and visual RAG are gaining significant attention. Recent research indicates that using VLMs can achieve better RAG performance, but processing rich documents still remains a challenge since a single page contains large amounts of information. In this paper, we present SCAN (\textbf{S}emanti\textbf{C} Document Layout \textbf{AN}alysis), a novel approach enhancing both textual and visual Retrieval-Augmented Generation (RAG) systems working with visually rich documents. It is a VLM-friendly approach that identifies document components with appropriate semantic granularity, balancing context preservation with processing efficiency. SCAN uses a coarse-grained semantic approach that divides documents into coherent regions covering continuous components. We trained the SCAN model by fine-tuning object detection models with sophisticated annotation datasets. Our experimental results across English and Japanese datasets demonstrate that applying SCAN improves end-to-end textual RAG performance by up to 9.0\% and visual RAG performance by up to 6.4\%, outperforming conventional approaches and even commercial document processing solutions.

Yoichi Ishibashi, Taro Yano, Masafumi Oyamada

Large Language Models (LLMs) have demonstrated significant improvements in reasoning capabilities through supervised fine-tuning and reinforcement learning. However, when training reasoning models, these approaches are primarily applicable to specific domains such as mathematics and programming, which imposes fundamental constraints on the breadth and scalability of training data. In contrast, continual pretraining (CPT) offers the advantage of not requiring task-specific signals. Nevertheless, how to effectively synthesize training data for reasoning and how such data affect a wide range of domains remain largely unexplored. This study provides a detailed evaluation of Reasoning CPT, a form of CPT that uses synthetic data to reconstruct the hidden thought processes underlying texts, based on the premise that texts are the result of the author's thinking process. Specifically, we apply Reasoning CPT to Gemma2-9B using synthetic data with hidden thoughts derived from STEM and Law corpora, and compare it to standard CPT on the MMLU benchmark. Our analysis reveals that Reasoning CPT consistently improves performance across all evaluated domains. Notably, reasoning skills acquired in one domain transfer effectively to others; the performance gap with conventional methods widens as problem difficulty increases, with gains of up to 8 points on the most challenging problems. Furthermore, models trained with hidden thoughts learn to adjust the depth of their reasoning according to problem difficulty.

Taro Yano, Yoichi Ishibashi, Masafumi Oyamada

Large Language Models (LLMs) have demonstrated exceptional performance across a wide range of tasks. To further tailor LLMs to specific domains or applications, post-training techniques such as Supervised Fine-Tuning (SFT), Preference Learning, and model merging are commonly employed. While each of these methods has been extensively studied in isolation, the automated construction of complete post-training pipelines remains an underexplored area. Existing approaches typically rely on manual design or focus narrowly on optimizing individual components, such as data ordering or merging strategies. In this work, we introduce LaMDAgent (short for Language Model Developing Agent), a novel framework that autonomously constructs and optimizes full post-training pipelines through the use of LLM-based agents. LaMDAgent systematically explores diverse model generation techniques, datasets, and hyperparameter configurations, leveraging task-based feedback to discover high-performing pipelines with minimal human intervention. Our experiments show that LaMDAgent improves tool-use accuracy by 9.0 points while preserving instruction-following capabilities. Moreover, it uncovers effective post-training strategies that are often overlooked by conventional human-driven exploration. We further analyze the impact of data and model size scaling to reduce computational costs on the exploration, finding that model size scalings introduces new challenges, whereas scaling data size enables cost-effective pipeline discovery.

Kosuke Akimoto, Kunihiro Takeoka, Masafumi Oyamada

Retrieval-augmented generation models augment knowledge encoded in a language model by providing additional relevant external knowledge (context) during generation. Although it has been shown that the quantity and quality of context impact the performance of retrieval-augmented generation models during inference, limited research explores how these characteristics affect model training. This paper explores how context quantity and quality during model training affect the performance of Fusion-in-Decoder (FiD), the state-of-the-art retrieval-augmented generation model, in extractive open-domain question answering tasks. Experimental results suggest that FiD models overfit to context quality during training and show suboptimal performance when evaluated on different context quality. Through the experimental results, we also reveal FiD models trained with different context quality have different cross-attention distribution patterns. Specifically, as context quality during training increases, FiD models tend to attend more uniformly to each passage in context. Finally, based on these observations, we propose a method to mitigate overfitting to specific context quality by introducing bias to the cross-attention distribution, which we demonstrate to be effective in improving the performance of FiD models on different context quality.

Junpei Komiyama, Daisuke Oba, Masafumi Oyamada

We study best-of-$N$ for large language models (LLMs) where the selection is based on majority voting. In particular, we analyze the limit $N \to \infty$, which we denote as Best-of-$\infty$. While this approach achieves impressive performance in the limit, it requires an infinite test-time budget. To address this, we propose an adaptive generation scheme that selects $N$ based on answer agreement, thereby efficiently allocating inference-time computation. Beyond adaptivity, we extend the framework to weighted ensembles of multiple LLMs, showing that such mixtures can outperform any individual model. The optimal ensemble weighting is formulated and efficiently computed as a mixed-integer linear program. Extensive experiments demonstrate the effectiveness of our approach.

Yahan Yu, Yuyang Dong, Masafumi Oyamada

Reasoning is a key capability for large language models (LLMs), particularly when applied to complex tasks such as mathematical problem solving. However, multimodal reasoning research still requires further exploration of modality alignment and training costs. Many of these approaches rely on additional data annotation and relevant rule-based rewards to enhance the understanding and reasoning ability, which significantly increases training costs and limits scalability. To address these challenges, we propose the Deliberate-to-Intuitive reasoning framework (D2I) that improves the understanding and reasoning ability of multimodal LLMs (MLLMs) without extra annotations and complex rewards. Specifically, our method sets deliberate reasoning strategies to enhance modality alignment only through the rule-based format reward during training. While evaluating, the reasoning style shifts to intuitive, which removes deliberate reasoning strategies during training and implicitly reflects the model's acquired abilities in the response. D2I outperforms baselines across both in-domain and out-of-domain benchmarks. Our findings highlight the role of format reward in fostering transferable reasoning skills in MLLMs, and inspire directions for decoupling training-time reasoning depth from test-time response flexibility.

Takuya Tamura, Taro Yano, Masafumi Enomoto et al.

Accurately forecasting the performance of Large Language Models (LLMs) before extensive fine-tuning or merging can substantially reduce both computational expense and development time. Although prior approaches like scaling laws account for global factors such as parameter size or training tokens, they often overlook explicit lineage relationships-i.e., which models are derived or merged from which parents. In this work, we propose a novel Lineage-Regularized Matrix Factorization (LRMF) framework that encodes ancestral ties among LLMs via a graph Laplacian regularizer. By leveraging multi-hop parent-child connections, LRMF consistently outperforms conventional matrix factorization and collaborative filtering methods in both instance-level and benchmark-level performance prediction. Our large-scale study includes 2,934 publicly available Hugging Face models and 21,000+ instances across 6 major benchmarks, showing that the introduction of lineage constraints yields up to 0.15-0.30 higher Pearson correlation coefficients with actual performance compared to baseline methods. Moreover, LRMF effectively addresses the cold-start problem, providing accurate estimates for newly derived or merged models even with minimal data. This lineage-guided strategy thus offers a resource-efficient way to inform hyperparameter tuning, data selection, and model combination in modern LLM development.

Gorjan Radevski, Kiril Gashteovski, Shahbaz Syed et al.

Question Answering (QA) accounts for a significant portion of LLM usage "in the wild". However, LLMs sometimes produce false or misleading responses, also known as "hallucinations". Therefore, grounding the generated answers in contextually provided information -- i.e., providing evidence for the generated text -- is paramount for LLMs' trustworthiness. Providing this information is the task of context attribution. In this paper, we systematically study LLM-based approaches for this task, namely we investigate (i) zero-shot inference, (ii) LLM ensembling, and (iii) fine-tuning of small LMs on synthetic data generated by larger LLMs. Our key contribution is SynQA: a novel generative strategy for synthesizing context attribution data. Given selected context sentences, an LLM generates QA pairs that are supported by these sentences. This leverages LLMs' natural strengths in text generation while ensuring clear attribution paths in the synthetic training data. We show that the attribution data synthesized via SynQA is highly effective for fine-tuning small LMs for context attribution in different QA tasks and domains. Finally, with a user study, we validate the usefulness of small LMs (fine-tuned on synthetic data from SynQA) in context attribution for QA.

Yoichi Ishibashi, Taro Yano, Masafumi Oyamada

Large Language Models (LLMs) have achieved remarkable capabilities, yet their improvement methods remain fundamentally constrained by human design. We present Self-Developing, a framework that enables LLMs to autonomously discover, implement, and refine their own improvement algorithms. Our approach employs an iterative cycle where a seed model generates algorithmic candidates as executable code, evaluates their effectiveness, and uses Direct Preference Optimization to recursively improve increasingly sophisticated improvement strategies. We demonstrate this framework through model merging, a practical technique for combining specialized models. Self-Developing successfully discovered novel merging algorithms that outperform existing human-designed algorithms. On mathematical reasoning benchmarks, the autonomously discovered algorithms improve the seed model's GSM8k performance by 6\% and exceed human-designed approaches like Task Arithmetic by 4.3\%. Remarkably, these algorithms exhibit strong generalization, achieving 7.4\% gains on out-of-domain models without re-optimization. Our findings demonstrate that LLMs can transcend their training to invent genuinely novel optimization techniques. This capability represents a crucial step toward a new era where LLMs not only solve problems but autonomously develop the methodologies for their own advancement.

Kosuke Akimoto, Masafumi Oyamada

In this paper, we address the challenge of optimizing training setups for Large Language Models (LLMs) of low-resource language with a limited amount of corpus. Existing works adopt multi-epoch, multi-lingual, and two-stage training to utilize the limited target language corpus efficiently. However, there is still a lack of understanding about the optimal hyperparameter setups for combining these three approaches to train LLMs. We exhaustively explore training setups for low-resource language LLM, combining these three approaches, and found the following insights for efficiently reducing the cost of hyperparameter search: (1) As the amount of target language corpus decreases, the optimal training approach shifts from monolingual single-stage training to multi-lingual two-stage training at a compute budget dependent threshold. (2) The optimal model scale remains stable regardless of the amount of target language corpus, allowing the use of the compute-optimal scale of monolingual training. (3) The optimal number of epochs can be extrapolated from smaller-scale experiments to larger scale using our proposed model. Also, we provide evidence that, in single-stage training, the target language validation loss follows a power law with respect to the target language ratio, with an exponent independent of the amount of data, model scale, and language pair.

Masafumi Enomoto, Kunihiro Takeoka, Kosuke Akimoto et al.

Open-Domain Multi-Document Summarization (ODMDS) is the task of generating summaries from large document collections in response to user queries. This task is crucial for efficiently addressing diverse information needs from users. Traditional retrieve-then-summarize approaches fall short for open-ended queries in ODMDS tasks. These queries often require broader context than initially retrieved passages provide, making it challenging to retrieve all relevant information in a single search. While iterative retrieval methods has been explored for multi-hop question answering (MQA), it's impractical for ODMDS due to high latency from repeated LLM inference. Accordingly, we propose LightPAL, a lightweight passage retrieval method for ODMDS. LightPAL leverages an LLM to pre-construct a graph representing passage relationships, then employs random walk during retrieval, avoiding iterative LLM inference. Experiments demonstrate that LightPAL outperforms naive sparse and pre-trained dense retrievers in both retrieval and summarization metrics, while achieving higher efficiency compared to iterative MQA approaches.

Yuyang Dong, Kunihiro Takeoka, Chuan Xiao et al.

Finding joinable tables in data lakes is key procedure in many applications such as data integration, data augmentation, data analysis, and data market. Traditional approaches that find equi-joinable tables are unable to deal with misspellings and different formats, nor do they capture any semantic joins. In this paper, we propose PEXESO, a framework for joinable table discovery in data lakes. We embed textual values as high-dimensional vectors and join columns under similarity predicates on high-dimensional vectors, hence to address the limitations of equi-join approaches and identify more meaningful results. To efficiently find joinable tables with similarity, we propose a block-and-verify method that utilizes pivot-based filtering. A partitioning technique is developed to cope with the case when the data lake is large and the index cannot fit in main memory. An experimental evaluation on real datasets shows that our solution identifies substantially more tables than equi-joins and outperforms other similarity-based options, and the join results are useful in data enrichment for machine learning tasks. The experiments also demonstrate the efficiency of the proposed method.