arXiv subjects

Yukinori Yamamoto

Publications and source records attributed to Yukinori Yamamoto.

4 recordsLinked to original sources

arXiv2026

Introspecting Alignment Shifts Beyond Behaviors Implanted Through Fine-Tuning

Fine-tuning enables a source model to acquire desired capabilities and behaviors in a target domain while retaining much of its general-purpose competence. However, this adaptation process can also degrade alignment properties that were present in the source model. Recent work has shown that large language models can be trained using LoRA-based modules known as introspection adapters (IAs) to describe behavioral changes induced by fine-tuning. However, existing studies primarily consider settings in which the model is fine-tuned on datasets explicitly designed to implant a specific behavior and is then asked to explain the implanted behavior. This differs from practical deployment scenarios, where the misalignment is not necessarily deliberately implanted, but arises only as a side effect. To bridge this gap, we formulate a novel problem setting, in which the target of introspection is not necessarily a behavior explicitly implanted through fine-tuning, but rather alignment shifts that may emerge as unintended side effects, and we construct a dataset for this setting. Furthermore, to enhance sensitivity to internal model changes, we propose the Delta-Aware Introspection Adapter (DAIA), a novel mechanism designed to explicitly process both base-model activations and activation differences induced by fine-tuning. Our empirical evaluation shows that introspection learning generalizes to unseen fine-tuned models and safety categories, and that DAIA generally outperforms existing IAs.

cs.LG↗

arXiv2026

FDIF: Formula-Driven supervised Learning with Implicit Functions for 3D Medical Image Segmentation

Deep learning-based 3D medical image segmentation methods relies on large-scale labeled datasets, yet acquiring such data is difficult due to privacy constraints and the high cost of expert annotation. Formula-Driven Supervised Learning (FDSL) offers an appealing alternative by generating training data and labels directly from mathematical formulas. However, existing voxel-based approaches are limited in geometric expressiveness and cannot synthesize realistic textures. We introduce Formula-Driven supervised learning with Implicit Functions (FDIF), a framework that enables scalable pre-training without using any real data and medical expert annotations. FDIF introduces an implicit-function representation based on signed distance functions (SDFs), enabling compact modeling of complex geometries while exploiting the surface representation of SDFs to support controllable synthesis of both geometric and intensity textures. Across three medical image segmentation benchmarks (AMOS, ACDC, and KiTS) and three architectures (SwinUNETR, nnUNet ResEnc-L, and nnUNet Primus-M), FDIF consistently improves over a formula-driven method, and achieves performance comparable to self-supervised approaches pre-trained on large-scale real datasets. We further show that FDIF pre-training also benefits 3D classification tasks, highlighting implicit-function-based formula supervision as a promising paradigm for data-free representation learning. Code is available at https://github.com/yamanoko/FDIF.

cs.CV↗

arXiv2020

Coupling Physics in Machine Learning to Predict Properties of High-temperatures Alloys

High-temperature alloy design requires a concurrent consideration of multiple mechanisms at different length scales. We propose a workflow that couples highly relevant physics into machine learning (ML) to predict properties of complex high-temperature alloys with an example of the 9-12 wt.% Cr steels yield strength. We have incorporated synthetic alloy features that capture microstructure and phase transformations into the dataset. Identified high impact features that affect yield strength of 9Cr from correlation analysis agree well with the generally accepted strengthening mechanism. As part of the verification process, the consistency of sub-datasets has been extensively evaluated with respect to temperature and then refined for the boundary conditions of trained ML models. The predicted yield strength of 9Cr steels using the ML models is in excellent agreement with experiments. The current approach introduces physically meaningful constraints in interrogating the trained ML models to predict properties of hypothetical alloys when applied to data-driven materials.

cond-mat.mtrl-sci↗

arXiv2018

Modern Data Analytics Approach to Predict Creep of High-Temperature Alloys

A breakthrough in alloy design often requires comprehensive understanding in complex multi-component/multi-phase systems to generate novel material hypotheses. We introduce a modern data analytics workflow that leverages high-quality experimental data augmented with advanced features obtained from high-fidelity models. Herein, we use an example of a consistently-measured creep dataset of developmental high-temperature alloy combined with scientific alloy features populated from a high-throughput computational thermodynamic approach. Extensive correlation analyses provide ranking insights for most impactful alloy features for creep resistance, evaluated from a large set of candidate features suggested by domain experts. We also show that we can accurately train machine learning models by integrating high-ranking features obtained from correlation analyses. The demonstrated approach can be extended beyond incorporating thermodynamic features, with input from domain experts used to compile lists of features from other alloy physics, such as diffusion kinetics and microstructure evolution.

cond-mat.mtrl-sci↗