arXiv subjects

Yining Xie

Publications and source records attributed to Yining Xie.

4 recordsLinked to original sources

arXiv2026

Spreadsheet-RL: Advancing Large Language Model Agents on Realistic Spreadsheet Tasks via Reinforcement Learning

Spreadsheet systems (e.g., Microsoft Excel, Google Sheets) play a central role in modern data-centric workflows. As AI agents grow increasingly capable of automating complex tasks, such as controlling computers and generating presentations, building an AI-driven spreadsheet agent has emerged as a promising research direction. Most existing spreadsheet agents rely on specialized prompting over general-purpose LLMs; while this design has potentials on simple spreadsheet operations, it struggles to manage the complex, multi-step workflows typical of real-world applications. We introduce Spreadsheet-RL, a reinforcement learning (RL) fine-tuning framework designed to train specialized spreadsheet agents within a realistic Microsoft Excel environment. Spreadsheet-RL features an automated pipeline for scalable collection of paired start-goal spreadsheets from online forums, as well as domain-specific evaluation tasks in areas such as finance and supply chain management, which we compile into the new Domain-Spreadsheet benchmark dataset. It also includes a Spreadsheet Gym environment designed for multi-turn RL: Spreadsheet Gym exposes extensive Excel functionality through a Python sandbox, along with a refined harness that incorporates a comprehensive tool set and carefully designed tool-routing rules for spreadsheet tasks. Through comprehensive experiments, we show that Spreadsheet-RL substantially enhances AI agent's performance on both general and domain-specific spreadsheet tasks: it improves Qwen3-4B-Thinking-2507's Pass@1 on SpreadsheetBench from 12.0% to 23.4%, and raises Pass@1 from 8.4% to 17.2% on our curated Domain-Spreadsheet dataset. These results highlight Spreadsheet-RL's strong potential for generalization and real-world adoption in spreadsheet automation, and broadly, its promise for advancing LLM-based interactions with data interfaces in everyday work.

cs.AI↗

arXiv2025

The Microscopic Structure of Stacking Faults in Sr$_2$NaNb$_5$O$_{15}$

Stacking faults and other topological defects in ferroics can have a significant influence on the electronic and mechanical properties of the material. Here, regular stacking faults in the tetragonal tungsten bronze material Sr$_2$NaNb$_5$O$_{15}$ are investigated through transmission electron microscopy, symmetry mode analysis and machine-learned force-field calculations. It is shown that the faults, with a fault vector of $\frac{1}{4}[\bar{2}12]_o$, annihilate in sets of four in the material, owing to the $\frac{1}{4}$ unit cell displacement along the b-axis. The four resulting domains emerge as four possible directions of the S$_3$ order parameter, related to NbO$_6$ octahedral tilts in the material. Force-field calculations reveal that the stacking faults are likely placed at positions where the octahedra in neighbouring domains have similar magnitudes of rotation, and that the estimated stacking fault energy is 46 mJ/m$^2$. The investigation shows that the stacking faults have a significant local effect on the polar modes present in the structure, and therefore could affect the ferroelectric properties.

cond-mat.mtrl-sci↗

arXiv2025

Template masks for 4D-STEM

We present a new analysis method for atomic resolution four-dimensional scanning transmission electron microscopy (4D-STEM, in which a diffraction pattern is collected at each point of a raster scan of a focused electron beam across the specimen). In 4D-STEM, each measured intensity has a dual character, forming a pixel in a diffraction pattern and, equally, forming a pixel in a STEM image. Applying a mask to the data to obtain a "virtual" bright field or dark field image is widely used and understood. However, there is a complementary procedure, in which an image (template) is applied to the data to obtain a mask. This mask shows the correlation between the data and the template and, when applied to atomic resolution 4D-STEM data produces an image optimised for the template. This allows, for example, imaging of specific atom columns and is a significant improvement over user-defined masks such as virtual annular bright field imaging. We demonstrate the capability of the approach, separately imaging Li and O atom columns in LiFePO4 and O, Pb and Ti across a domain wall in PbTiO3.These template masks provide a computationally straightforward and general method to probe 4D-STEM data. They are particularly effective for specimens of moderate thickness where multiple scattering produces strong and specific correlations in diffraction patterns.

physics.ins-det↗

arXiv2025

MFM-DA: Instance-Aware Adaptor and Hierarchical Alignment for Efficient Domain Adaptation in Medical Foundation Models

Medical Foundation Models (MFMs), trained on large-scale datasets, have demonstrated superior performance across various tasks. However, these models still struggle with domain gaps in practical applications. Specifically, even after fine-tuning on source-domain data, task-adapted foundation models often perform poorly in the target domain. To address this challenge, we propose a few-shot unsupervised domain adaptation (UDA) framework for MFMs, named MFM-DA, which only leverages a limited number of unlabeled target-domain images. Our approach begins by training a Denoising Diffusion Probabilistic Model (DDPM), which is then adapted to the target domain using a proposed dynamic instance-aware adaptor and a distribution direction loss, enabling the DDPM to translate source-domain images into the target domain style. The adapted images are subsequently processed through the MFM, where we introduce a designed channel-spatial alignment Low-Rank Adaptation (LoRA) to ensure effective feature alignment. Extensive experiments on optic cup and disc segmentation tasks demonstrate that MFM-DA outperforms state-of-the-art methods. Our work provides a practical solution to the domain gap issue in real-world MFM deployment. Code will be available at here.

cs.CV↗