| 片岡 晴彦 | M, 2回目発表 | 自然言語処理学 | 渡辺 太郎, | Sakriani Sakti, | 上垣外 英剛, | 坂井 優介 |
|
title: Tighter Locality: A More Rigorous Assessment of Locality in Knowledge Editing Methods for Large Language Models
abstract: Research on knowledge editing methods for large-scale language models has not only proposed new methods but also introduced various evaluation metrics for them. Among these, “Locality”—which evaluates whether the method can control the impact of edits on unintended knowledge or model behavior— is one of the most important metrics. However, previous Locality evaluations have primarily used knowledge with low relevance to the edited knowledge, resulting in the inability to rigorously assess control capabilities. In this study, with the aim of evaluating Locality more rigorously, we propose a new benchmark for evaluating knowledge editing methods called “Tighter Locality.” Experimental results show that existing knowledge editing methods score significantly lower on the Locality evaluation using Tighter Locality. In the future, we plan to analyze why existing knowledge editing methods score lower on Tighter Locality. language of the presentation: Japanese 発表題目: Tighter Locality : 大規模言語モデルに対する知識編集手法のより厳密な局所性評価 発表概要: 大規模言語モデルの知識編集手法についての研究は、手法の提案とともに、その手法の評価指標についてもさまざまな提案がなされてきた。中でも、編集が意図していない知識やモデルの挙動への影響を 制御できているかどうかを評価する Locality は重要な指標の一つである。しかしこれまでの Locality 評価は、編集した知識と関連性の低い知識を用いたものが主で、制御能力を厳密に評価できていない という問題があった。本研究ではより厳密に Locality を評価することを目的として、知識編集手法評価用の新たなベンチマーク Tighter Locality を提案する。実験の結果、既存の知識編集手法は、Tighter Locality 上での Locality評価でスコアを大きく落とすことがわかった。 今後は既存の知識編集手法が Tighter Locality でスコアを落とすことについての分析を行う予定である。 | ||||||
| 鼎 凌太朗 | M, 2回目発表 | 自然言語処理学 | 渡辺 太郎, | Sakriani Sakti, | 上垣外 英剛, | 坂井 優介 |
|
title: When Can We Trust Diffusion Language Models? A Step-wise Analysis of Confidence Estimation
abstract: Diffusion Language Models generate text through iterative masking and unmasking rather than left-to-right token prediction. Increasing the number of denoising steps is generally considered important for improving generation quality, but it remains unclear whether more steps also lead to more reliable confidence estimates. In this study, I investigate step-wise confidence estimation for pretrained Diffusion Language Models across multiple tasks and models. The main idea is to estimate sentence-level confidence at each generation step by applying multiple random masks to the current intermediate text and measuring how well the model reconstructs the original tokens. This procedure is motivated by the masked reconstruction behavior used during training. Using these confidence estimates, I analyze how reliability changes over generation steps, how many masks are needed for stable estimation, and whether the observed trends are consistent across evaluation metrics, tasks, and models. This study aims to clarify when intermediate states in Diffusion Language Models become trustworthy for final-quality prediction. language of the presentation: Japanese | ||||||
| 門谷 拓能 | M, 2回目発表 | 自然言語処理学 | 渡辺 太郎, | Sakriani Sakti, | 上垣外 英剛, | 坂井 優介 |
|
title: Can Image Descriptions Reconstruct the Original Structure?
Evaluating the Reconstructability of Multimodal Large Language Models Using Minecraft
abstract: Multimodal large languagemodels can generate fluent descriptions of visual content, but it remains unclear whether these descriptions contain sufficient information to accurately reconstruct the original object. This study introduces Image-to-Text-to-Build (I2T2B), a task for evaluating image descriptions through the reconstruction of Minecraft structures. Using a dataset of 200 structures, we generate descriptions with multiple multimodal large language models and compare two reconstruction approaches: direct reconstruction from natural-language descriptions and reconstruction through a structured intermediate representation. The generated descriptions and reconstructed structures are evaluated using linguistic quality, voxel overlap, material consistency, exact placement accuracy, and the amount of repair required. The results show that descriptions judged to be natural and detailed do not necessarily preserve the dimensions, spatial arrangements, and materials required for faithful reconstruction. A structured intermediate representation improves reconstruction accuracy for some models, although its effectiveness varies across models. These findings demonstrate the importance of evaluating image descriptions not only by linguistic quality but also by their ability to support accurate reconstruction. language of the presentation: Japanese 発表題目: 画像説明文は元の建築を再構成できるか— Minecraftを用いたマルチモーダル大規模言語モデルの再建築可能性評価 — 発表概要: マルチモーダル大規模言語モデルは、画像の内容を自然な文章として説明できる一方、その説明文が元の対象を正確に再構成できるほど十分な情報を含んでいるかは明らかではない。本研究では、Minecraft建築を対象として、建築画像から説明文を生成し、その説明文のみを用いて建築を再構成するImage-to-Text-to-Build（I2T2B）タスクを提案する。200件の建築データに対して複数のマルチモーダル大規模言語モデルで説明文を生成し、自然言語の説明文から直接再建築する方法と、構造化中間表現を介して再建築する方法を比較する。生成された説明文および再建築結果を、言語的品質、ボクセルの重なり、材質一致率、正確配置率、修正量によって評価する。実験の結果、自然で詳細な説明文であっても、建築物の寸法、空間配置、材質を正確に再現できるとは限らないことが確認された。また、構造化中間表現は一部のモデルで再建築精度を改善したが、その効果はモデルによって異なった。本研究は、画像説明文を言語的な品質だけでなく、説明された対象を正確に再構成できるかという観点から評価する枠組みを示す。 | ||||||
| 松永 立樹 | M, 2回目発表 | ロボットラーニング | 松原 崇充, | 和田 隆広, | 柴田 一騎, | 佐々木 光 | |
|
title: Curriculum Reinforcement Learning for Robust Wheeled-Legged Robot Locomotion on Unstable Terrain with Collapse RisK
abstract: table terrain with collapse risk is difficult to traverse because ground failure is inevitably triggered through robot–terrain contact and can occur within a fraction of a second. Previous studies either verify every foothold in advance at the cost of locomotion speed or learn robust locomotion while assuming that the terrain always provides sufficient support. To address this gap, we propose a curriculum reinforcement learning framework for acquiring a collapse-survival skill. A single proprioceptive policy learns to retreat from sudden support loss through training on cliff-like terrain environments. The learned policy is evaluated zero-shot on unseen collapsing floors in a unified benchmark and demonstrated on a real wheeled-legged quadruped. The results suggest that the proposed framework can help reconcile the trade-off between locomotion speed and safety on collapse-risk terrain. language of the presentation: Japanese | |||||||
| 山田 大夢 | M, 2回目発表 | ロボットラーニング | 松原 崇充, | 和田 隆広, | 柴田 一騎, | 佐々木 光 | |
|
title: Imitation Learning Framework for Screw Removal in the Disassembly of Recycled Products
abstract: In the disassembly of recycled products, the automation of screw removal using robots has attracted significant attention. However, there is a challenge in that screws cannot be successfully loosened due to tool misalignment caused by camera recognition errors. Previous studies have attempted to resolve this by formalizing the amount of misalignment or introducing specialized tools, but high implementation costs have been a problem. Therefore, in this study, I propose a framework that utilizes existing systems to learn movements that correct positional misalignment. Specifically, I focused on “imitation learning,” which involves learning movements from human demonstration data. However, information critical for learning the control strategy—such as the exact amount of misalignment and contact force—is not included in the actual demonstration data. Therefore, I utilize this information as “privileged information” obtainable only within a simulation environment. This allows the robot to simultaneously learn to estimate misalignment and perform corrective movements. In our experiments, I focused on the pushing motion in a simulation environment, which forms the basis of screw removal. The results confirmed that the proposed method enables the robot to imitate a pushing motion at a constant force. Going forward, I aim to apply this technology to the automation of actual screw removal tasks. language of the presentation: Japanese 発表題目: リサイクル製品の分解におけるネジ外しに向けた模倣学習フレームワーク 発表概要: リサイクル製品の分解において、ロボットによるネジ外しの自動化が注目されています。しかし、カメラの認識誤差による工具の位置ズレが原因で、ネジをうまく緩められないという課題があります。先行研究ではズレ量の定式化や専用ツールの導入で解決を図っていますが、高い導入コストが問題でした。 そこで本研究では、既存のシステムを活用し、位置ズレを修正する動作を学習する枠組みを提案します。具体的には、人間の教示データから動作を学ぶ「模倣学習」に着目しました。しかし、方策の学習に重要な正確なズレ量や接触力といった情報は、実際の教示データには含まれません。そこで、シミュレーション環境のみで得られる「特権情報」として活用します。これにより、ロボットにズレの推定と修正動作を同時に学習させます。 実験では、ネジ外しの基礎となるシミュレーション環境での押し付け動作を対象としました。その結果、提案手法によって、一定の力での押し付け動作を模倣できることを確認しました。今後はこの技術を実際のネジ外し自動化へ応用することを目指します | |||||||
| YANG YUQI | M, 2回目発表 | ロボットラーニング | 松原 崇充, | 和田 隆広, | 柴田 一騎, | 佐々木 光, | 角川 勇貴 |
|
title: Recovery-Aware Hierarchical Policy Learning Framework through Online Terrain Safety Estimation for Navigation in Unknown Hazardous Terrain
abstract: Goal-directed navigation in disaster environments is particularly challenging because terrain safety often remains unknown until the robot contacts the ground, when failure may already be imminent. Such navigation problems are commonly addressed using hierarchical frameworks that separate high-level route planning from low-level locomotion. However, existing hierarchical methods rely on prior terrain observations and neither update safety from contact feedback nor consider the recovery capability of the low-level policy. We propose a hierarchical framework integrating online safety estimation with recovery-aware high-level policy learning. Contact and proprioceptive feedback update a terrain safety belief, while the high-level policy replans routes and selects actions recoverable by the low-level controller. Built upon a perceptive locomotion policy using an onboard terrain map, the framework is validated through sim-to-real transfer on a wheeled-legged quadruped. Preliminary simulations demonstrate online safety-map refinement and adaptive route replanning, enabling safer navigation under unknown terrain hazards. language of the presentation: English | |||||||
| 荒木 駿佑 | D, 中間発表 | ロボットラーニング | 松原 崇充, | 和田 隆広, | 柴田 一騎, | 佐々木 光 | |
|
title: DecompGrind: A Decomposition Framework for Robotic Grinding via Cutting-Surface Planning and Contact-Force Adaptation
abstract: Robotic grinding is widely used for shaping workpieces in manufacturing, but it remains difficult to automate this process efficiently. In particular, efficiently grinding workpieces with different shapes and density-dependent removal resistance is challenging because removal resistance varies with local contact conditions. Moreover, it is difficult to achieve accurate estimation of removal resistance and analytical modeling of shape transition, and learning-based approaches often require large amounts of training data to cover diverse processing conditions. To address these challenges, we decompose robotic grinding into two components: removal-shape planning and contact-force adaptation. Based on this formulation, we propose DecompGrind, a framework that combines Global Cutting-Surface Planning (GCSP) and Local Contact-Force Adaptation (LCFA). GCSP determines removal shapes through geometric analysis of the current and target shapes without learning, while LCFA learns a contact-force adaptation policy using bilateral control-based imitation learning during the grinding of each removal shape. This decomposition restricts learning to local contact-force adaptation, allowing the policy to be learned from a small number of demonstrations, while handling global shape transition geometrically. Experiments using a robotic grinding system on 3D-printed workpieces with different shapes and densities under fixed belt conditions demonstrate reduced grinding time while maintaining safe levels of contact force. language of the presentation: Japanese | |||||||
| SIVASUNDARAM SUHARNAN | D, 中間発表 | 計算システムズ生物学 | 金谷 重彦, | 松本 健一, | MD.Altaf-Ul-Amin, | 小野 直亮 | |
|
title: Biomass Yield, Growth Kinetics, and Wet-to-Dry Conversion of Kappaphycus alvarezii in Northern Sri Lanka Across Monsoonal Time Series and Environmental Dynamics
abstract: Uniform 100 g vegetative propagules were monitored over standard 45-day cycles using floating long-line systems (N = 45 replicate lines per block), with growth sampled at Days 0, 10, 25, and 45. Results revealed pronounced seasonal divergence: the June–December 2024 block delivered the highest final wet weight (1,670 g), a Biomass Multiplication Factor (BMF) of 16.7, and a Daily Growth Rate (DGR) of 6.26% day⁻¹. The September–January 2025–2026 block achieved 1,580 g (DGR = 6.13% day⁻¹), while the January–August 2025 block suffered severe thermal retardation, yielding only 1,040 g (DGR = 5.20% day⁻¹)—a 37.7% reduction in commercial dry output. A universal tri-phasic growth pattern was confirmed across all blocks: a lag phase (Days 0–10), a hyper-exponential phase (Days 10–25) peaking at 83.6 g day⁻¹ absolute gain, and a deceleration phase (Days 25–45) where DGR declined to 2.2% day⁻¹. Gravimetric conversion profiles revealed dry matter yields of 10.3–12.1% across regional sites. High-frequency environmental monitoring recorded extreme temperatures peaking at 40.0°C (Siruthivu, August 2025), providing a mechanistic context for observed growth collapse and heightened ice-ice disease risk.The DGR values recorded here (5.20–6.26% day⁻¹) substantially exceed the national Sri Lankan benchmark of 3.50 ± 0.50% day⁻¹, warranting independent multi-strain replication. These findings collectively establish that late-year monsoonal cycles are optimal for K. alvareziicultivation in Northern Sri Lanka, and that a strict 25–35 day harvest window maximises carrageenan yield while minimising environmental crop-loss risk. language of the presentation: English | |||||||
| KHAN MUHAMMAD IMRAN | D, 中間発表 | 計算システムズ生物学 | 金谷 重彦, | 松本 健一, | MD.Altaf-Ul-Amin, | 小野 直亮, | Ahmad Kamal Nasution |
|
title: A Challenge-Driven Deep Learning Framework for Antimicrobial Resistance Prediction: When Do Complex Architectures Matter?
abstract: Biomedical datasets share recurring challenges: extreme sparsity, class imbalance, high dimensionality, mixed data types, and complex inter-feature relationships. We propose a challenge-driven design philosophy where each data challenge is addressed by a dedicated architectural component, and evaluate this approach on antimicrobial resistance (AMR) prediction — a critical clinical problem where traditional susceptibility testing requires 48–72 hours. Our framework maps five challenges to architectural solutions: (1) sparse binary features to a Bernoulli encoder providing gradient flow despite >94% zeros; (2) zero-inflated continuous features to a Zero-Inflated MLP encoder; (3) complex feature relationships to a heterogeneous graph neural network with five semantic edge types; (4) multi-faceted information integration to type-specific Graph Attention Networks with attention-based fusion; (5) class imbalance to Dynamic Task Prioritization loss. Evaluated on the AMR-UTI dataset (116,902 cases), all models, including simple baselines, achieved AUROC of 0.68–0.83 across four antibiotics, representing 12.5–17.7% improvement over prior graph-based methods. Critically, comprehensive evaluation across three experimental regimes (full features, dominant feature removal, and temporal distribution shift from 2007–2013 to 2014–2016) revealed that a standard multi-layer perceptron achieves equivalent performance in all regimes, indicating that improvement stems from preprocessing and training methodology rather than architectural complexity. Feature importance analysis identified colonization pressure and prior resistance as dominant predictors that obviate relational modeling. We propose a signal-dominance framework explaining when complex architectures are unnecessary, thus, contributing to the understanding of architectural selection in clinical machine learning. language of the presentation: English | |||||||