arXiv subjects

Masafumi Oyamada

Publications and source records attributed to Masafumi Oyamada.

At least 19 recordsLinked to original sources

arXiv2026

$M^3$ Scaling Law: Optimizing Multi-Epoch, Multi-Lingual, and Multi-Stage Training for Low-Resource Language Models

In this paper, we study a fundamental design problem in pretraining Large Language Models (LLMs) for low-resource language regimes. Existing works adopt multi-epoch, multi-lingual, and multi-stage training to utilize the limited target-language corpus efficiently, but no prior scaling law can compare recipes spanning these approaches under the same compute budget $C$ and target-language corpus size $D_T$, leaving the optimal training setup unclear. To address this gap, we propose the $M^3$ Scaling Law, a unified predictive model parameterized by the model scale, the number of target-corpus epochs $k$, the average target-language ratio $r$, and the final-stage target-language ratio $r_f$, which places monolingual single-stage, multi-lingual single-stage, and multi-lingual multi-stage recipes on a single target-language loss surface. Across three language pairs, it extrapolates to unseen hyperparameter regions more accurately than existing scaling laws. Using $M^3$ as a surrogate objective, we derive two practical guidelines for low-resource LLM pretraining: (i) as $D_T$ decreases, the optimal recipe shifts directly from monolingual single-stage to multi-lingual two-stage training at a compute-budget-dependent threshold, with multi-lingual single-stage never optimal in our experimental grid; and (ii) the optimal number of epochs collapses onto a single curve in the scarcity variable $D_T/D^*(C)$, where $D^*(C) \propto C^{α/(α+β)}$ is the monolingual compute-optimal corpus size.

cs.CL↗

arXiv2026

Revisiting Observation Reduction for Web Agents: Comprehensive Evaluation with a Lightweight Framework

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\%.

cs.CL↗

arXiv2026

Learning Deliberately, Acting Intuitively: Unlocking Test-Time Reasoning in Multimodal LLMs

Reasoning is essential for large language models (LLMs), especially in complex tasks such as mathematical problem solving. However, multimodal reasoning still faces challenges in modality alignment and training scalability, as many existing methods rely on additional annotations or complex rule-based rewards. To address these issues, we propose the Deliberate-to-Intuitive reasoning framework (D2I), which improves the understanding and reasoning abilities of multimodal LLMs (MLLMs) without extra annotations or complex rewards. During training, D2I uses deliberate reasoning strategies supervised only by rule-based format rewards to enhance modality alignment. During inference, it shifts to intuitive reasoning by removing these explicit strategies, allowing the model to implicitly apply the acquired abilities in its responses. D2I outperforms baselines on both in-domain and out-of-domain benchmarks, highlighting the effectiveness of format rewards in fostering transferable multimodal reasoning skills and suggesting the benefit of decoupling training-time reasoning depth from test-time response flexibility.

cs.CV↗

arXiv2026

Evaluating the Impact of Reviewer Guideline Design on LLM-Based Automated Peer Review

Peer review is an essential process in scientific research, yet the growing workload has made its automation increasingly necessary. In this study, we analyze how different types of reviewer guidelines, such as official conference guidelines and reviewer-imitating ones generated from high-quality human reviews using LLMs, affect automated peer review. Our experiments show that official conference guidelines produce review results most consistent with human judgments, suggesting that evaluation criteria refined through conference practice serve as effective guidance for automated reviewing as well. In contrast, reviewer-imitating guidelines were generally less effective than official conference guidelines. Furthermore, enforcing strict rubric-style scoring consistently degraded performance, highlighting the importance of allowing subjective and holistic scoring.

cs.CL↗

arXiv2026

Effective Harness Engineering for Algorithm Discovery with Coding Agents

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.

cs.SE↗

arXiv2026

cotomi Act: Learning to Automate Work by Watching You

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.

cs.AI↗

arXiv2026

Best-of-$\infty$ -- Asymptotic Performance of Test-Time LLM Ensembling

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 \boinflower. 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.

stat.ML↗

arXiv2026

SCAN: Semantic Document Layout Analysis for Textual and Visual Retrieval-Augmented Generation

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 yields better RAG performance, but processing rich documents remains a challenge since a single page contains large amounts of information. In this paper, we present SCAN (SemantiC Document Layout ANalysis), a novel approach that enhances both textual and visual Retrieval-Augmented Generation (RAG) systems that work 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 contiguous components. We trained the SCAN model by fine-tuning object detection models on an annotated dataset. Our experimental results across English and Japanese datasets demonstrate that applying SCAN improves end-to-end textual RAG performance by up to 9.4 points and visual RAG performance by up to 10.4 points, outperforming conventional approaches and even commercial document processing solutions.

cs.AI↗

arXiv2025

DISC: Dynamic Decomposition Improves LLM Inference Scaling

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.

cs.LG↗

arXiv2025

Understanding the Impact of Confidence in Retrieval Augmented Generation: A Case Study in the Medical Domain

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.

cs.CL↗

arXiv2025

Can a Crow Hatch a Falcon? Lineage Matters in Predicting Large Language Model Performance

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.

cs.CL↗

arXiv2025

On Synthesizing Data for Context Attribution in Question Answering

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.

cs.IR↗

arXiv2025

An Empirical Study of LLM-as-a-Judge: How Design Choices Impact Evaluation Reliability

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.

cs.CL↗

arXiv2025

Can Large Language Models Invent Algorithms to Improve Themselves?: Algorithm Discovery for Recursive Self-Improvement through Reinforcement Learning

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.

cs.CL↗

arXiv2025

LaMDAgent: An Autonomous Framework for Post-Training Pipeline Optimization via LLM Agents

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.

cs.CL↗

arXiv2025

LLM-based Query Expansion Fails for Unfamiliar and Ambiguous Queries

Query expansion (QE) enhances retrieval by incorporating relevant terms, with large language models (LLMs) offering an effective alternative to traditional rule-based and statistical methods. However, LLM-based QE suffers from a fundamental limitation: it often fails to generate relevant knowledge, degrading search performance. Prior studies have focused on hallucination, yet its underlying cause--LLM knowledge deficiencies--remains underexplored. This paper systematically examines two failure cases in LLM-based QE: (1) when the LLM lacks query knowledge, leading to incorrect expansions, and (2) when the query is ambiguous, causing biased refinements that narrow search coverage. We conduct controlled experiments across multiple datasets, evaluating the effects of knowledge and query ambiguity on retrieval performance using sparse and dense retrieval models. Our results reveal that LLM-based QE can significantly degrade the retrieval effectiveness when knowledge in the LLM is insufficient or query ambiguity is high. We introduce a framework for evaluating QE under these conditions, providing insights into the limitations of LLM-based retrieval augmentation.

cs.IR↗

arXiv2025

Mining Hidden Thoughts from Texts: Evaluating Continual Pretraining with Synthetic Data for LLM Reasoning

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.

cs.CL↗