Research

Tool-augmented Language Models (TaLMs) can invoke external tools to solve problems beyond their parametric capacity. However, it remains unclear whether these tool-enabled gains reflect trustworthy reasoning. Focusing on the Code Interpreter tool, we show that even when tools are selected and executed correctly, TaLMs treat tool outputs as substitutes for reasoning, producing solutions that appear correct but lack coherent justification. We term this failure mode Tool-Induced Myopia (TIM), and study it using PYMATH, a benchmark of 1,679 competition-level mathematical problems for which Python code is helpful but not sufficient. We further develop a multi-dimensional evaluation suite to quantify reasoning degradation in TaLMs relative to their non-tool counterparts. Our findings reveal that while TaLMs achieve up to a 19.3 percentage point gain in final-answer accuracy, their reasoning behavior consistently deteriorates (e.g., non-tool LLMs win up to 41.5% more often in pairwise comparisons of reasoning process). This degradation intensifies with tool use; the more frequently a model invokes tools, the less coherent its reasoning becomes. Moreover, tool use shifts errors from arithmetic mistakes toward global reasoning failures (logic, assumption, creativity); with TIM present in ~55% of high-risk cases. Finally, we propose a preference-optimizationbased framework that realigns TaLMs to use tools as assistive evidence, improving both final-answer accuracy and reasoning depth under tool use. Codes and data are available at: https://github.com/megagonlabs/TIM.

The extraction of structured information from raw text is a fundamental component of many NLP applications, including document retrieval, ranking, and relevance estimation. High-quality extractions often require domain-specific accuracy, up-to-date understanding of specialized taxonomies, and the ability to incorporate emerging jargon and rare outliers. In many domains–such as medical, legal, and HR–the extraction model must also adapt to shifting terminology and benefit from explicit reasoning over structured knowledge. We propose DySECT, a Dynamic Self-Evolving Extraction and Curation Toolkit, which continually improves as it is used. The system incrementally populates a versatile, self-expanding knowledge base (KB) with triples extracted by the LLM. The KB further enriches itself through the integration of probabilistic knowledge and graph-based reasoning, gradually accumulating domain concepts and relationships. The enriched KB then feeds back into the LLM extractor via prompt tuning, sampling of relevant few-shot examples, or fine-tuning using KB-derived synthetic data. As a result, the system forms a symbiotic closed-loop cycle in which extraction continuously improves knowledge, and knowledge continuously improves extraction.

omnes flores is an NLP framework based on Universal Dependencies (UD) that utilizes multilingual Large Language Models (LLMs), and its default model is trained on data from 40 UD languages comprising 40 treebanks. For the EvaLatin 2026 Dependency Parsing Tasks, we extended the training data of omnes flores by incorporating six public Latin treebanks from UD and trained a dependency parsing model using the extended training data. The dependency parser of omnes flores normally takes a list of word FORM values as input. However, since the EvaLatin 2026 test data includes an UPOS column, we investigated whether incorporating both FORM and UPOS during both training and inference could improve parsing accuracy. Our experiments show that training using both FORM and UPOS improves performance by 0.5-1.0 LAS points on Prose compared with training using only FORM, but decreases performance by 5 points on Poetry.

In this research, we introduce LoRA probing, a lightweight approach for observing how core syntactic abilities emergeduring LLM pretraining. Leveraging OLMo-2’s public intermediate checkpoints, we trace learning curves across 24 pretraining stages on 33 Universal Dependencies languages by fine-tuning LoRA with step-by-step parsing instructions and a simple tabular output. To fit the relatively short context length of the OLMo-2, we design a compact 2-step-no-form prompt template and this matches the baseline in average accuracy while halving the context length and substantially increasing throughput, enabling efficient large-scale evaluation. Token Recall surpasses 0.9 within the first 1–2K pretraining steps, indicating that stable output formatting emerges early. Despite OLMo-2-7B’s English-centric pretraining, LAS exceeds 80 points in 29 of 33 languages; however, relations such as iobj and csubj show delayed onset and instability across many languages. LoRA probing thus provides a practical, reproducible lens on the cross-lingual dynamics of syntactic acquisition during LLM pretraining.

大規模言語モデルは差別的な社会的バイアスを含む情報を生成するリスクがあり，その評価が必要である．しかし，何を「差別的な社会的バイアス」とみなすかは社会的文脈に依存するため，普遍的な価値基準を定義することは難しく，個々の社会的文脈において合意可能な価値基準に基づいた安全性担保を行う必要がある．社会的バイアスのベンチマーク構築に関する先行研究では，社会的文脈の一つである国による価値基準の相違に対応するため，当該国のアノテーターを用いてローカライゼーションを行っているが，この手法はアノテーターの主観に強く依存しており，安全性の判断が当該国において合意可能なものであることを明確には担保していない．本研究では，各国において合意された「差別的な社会的バイアス」の最低限の基準として，法令とその判例等を根拠とする安全性担保のローカライゼーションを提案し，日本の雇用関連領域および医療提供関連領域の法令において差別と判断された事例を収集して，社会的バイアスデータセット – JLawBias を構築し，6 つの日本語対応 LLM に対して簡易な評価を実施して手法の有効性を確認した． ここに掲載した著作物の利用に関する注意 本著作物の著作権は人工知能学会に帰属します。本著作物は著作権者である人工知能学会の許可のもとに掲載するものです。ご利用に当たっては「著作権法」に従うことをお願いいたします。 Notice for the use of this material. The copyright of this material is retained by the Japanese Society for Artificial Intelligence (JSAI). This material is published here with the agreement of JSAI. Please be complied with Copyright Law of Japan if any users wish to reproduce, make derivative work, distribute or make available to the public any part or whole thereof. All Rights Reserved, Copyright (C) The Japanese Society for Artificial Intelligence.

Data lakes in modern enterprises are massive, heterogeneous, and noisy, often preventing non-experts from effectively extracting value. Bridging the semantic gap between ambiguous user intent and explicit data requires orchestrating multiple tools under an open-world assumption. However, reliably executing these compound AI workflows to solve complex knowledge extraction tasks, while also providing the transparency needed for evaluation and debugging, remains a significant bottleneck. We propose L.A.K.E. (Logic Agent for Knowledge Extraction), an agentic data planning framework designed to map natural language questions to executable workflows over diverse data sources. Rather than relying on a brittle “one-size-fits-all” approach, L.A.K.E. dynamically generates a declarative plan comprised of modular operators—spanning relational and semantic functions over heterogeneous data sources. Within this framework, we introduce and benchmark three distinct planning regimes: Iterative Planning, Single-Shot Tree Planning, and Cascade Planning. We present an interactive demonstration platform that enables users to visually compare the latency and robustness trade-offs of these planners. By rendering execution paths as interactive Directed Acyclic Graphs (DAGs) with step-level provenance, L.A.K.E. provides the critical observability needed to establish trust, debug failures, and optimize data planning for enterprise-scale lakes.