arXiv subjects

Wonseok Lee

Publications and source records attributed to Wonseok Lee.

18 recordsLinked to original sources

arXiv2026

Beyond On-Policy Exploration: Integrating External Policy Rollouts for Reinforcement Learning in Diffusion Language Models

Recent reinforcement learning methods for diffusion large language models (dLLMs) commonly rely on on-policy rollouts generated by the target dLLM itself. When successful on-policy rollouts are scarce, however, on-policy training may receive little positive reward and make only limited progress. To mitigate this problem, we explore incorporating higher-reward rollouts generated by a stronger external policy alongside on-policy rollouts from the target dLLM. However, directly incorporating these external rollouts introduces two practical challenges: differences in rollout length and instability when jointly processing rewards from on-policy and external rollouts. To address these challenges, we propose External Rollout Integration with Length Control and Source-Specific Processing (ERILS), which controls external-rollout length and processes the rewards of on-policy and external rollouts separately. Experiments on Sudoku, Countdown, and MATH500 under zero-shot evaluation show that ERILS improves multi-sample performance across all three tasks, with the largest gains on Sudoku. On Sudoku, ERILS achieves 98.4% best-of-4 completion accuracy, compared with 40.3% for the strongest baseline. ERILS also maintains approximately 90% deterministic single-completion accuracy on Sudoku across generation lengths of 128, 256, and 512 tokens. Our component analysis further shows that length-controlled external rollouts are more effective than uncontrolled external rollouts, and that source-specific reward processing avoids the training collapse observed with joint reward processing. These results show that rollout construction and reward processing are important design dimensions when integrating external rollouts into dLLM reinforcement learning.

cs.LG↗

arXiv2026

Orthogonal Negative Guidance in Attention Feature Space for Text-to-Image Generation

Text-to-image (T2I) models have become increasingly capable of generating high-quality images. Yet, enforcing the explicit absence of a specified object or attribute remains a fundamentally challenging problem. Existing approaches, including prompt negation, post-hoc editing, and negative guidance, remain insufficient for explicit concept suppression, often failing to remove the target concept or degrading overall image quality. To this end, we propose Orthogonal Negative Guidance in attention feature space, a training-free method that operates in the attention output space of MM-DiT-based T2I transformers. Our method orthogonalizes negative-prompt attention features with respect to positive-prompt features and subtracts only the orthogonal component, suppressing unwanted concepts while preserving desired semantics. Experiments on FLUX-dev and FLUX-schnell show that our method achieves favorable trade-offs between concept suppression, prompt alignment, and image quality. In human evaluation, our method outperforms the second-best baseline by 18.78%. We further show that our method supports multi-concept suppression and adjustable concept suppression.

cs.CV↗

arXiv2025

Imaging nanoscale photocarrier traps in solar water-splitting catalysts

Defects trap photocarriers and hinder solar water splitting. The nanoscale photocarrier transport, trapping, and recombination mechanisms are usually inferred from ensemble-averaged measurements and remain elusive. Because an individual high-performing nanoparticle photocatalyst may outperform the ensemble average, design rules that would otherwise enhance catalytic efficiency remain unclear. Here, we introduce photomodulated electron energy-loss spectroscopy (EELS) in an optically coupled scanning transmission electron microscope (STEM) to map photocarrier localization. Using rhodium-doped strontium titanate (SrTiO3:Rh) solar water-splitting nanoparticles, we directly image the carrier densities concentrated at oxygen-vacancy surface trap states. This is achieved by separating photothermal heating from photocarrier populations through experimental and computational analyses of low-loss spectra. Photomodulated STEM-EELS enables angstrom-scale imaging of defect-induced photocarrier traps and their impact on photocatalytic efficiency.

physics.chem-ph↗

arXiv2025

Imaging Nanoscale Carrier, Thermal, and Structural Dynamics with Time-Resolved and Ultrafast Electron Energy-Loss Spectroscopy

Time-resolved and ultrafast electron energy-loss spectroscopy (EELS) is an emerging technique for measuring photoexcited carriers, lattice dynamics, and near-fields across femtosecond to microsecond timescales. When performed in either a specialized scanning transmission electron microscope or ultrafast electron microscope (UEM), time-resolved and ultrafast EELS can directly image charge carriers, lattice vibrations, and heat dissipation following photoexcitation or applied bias. Yet recent advances in theoretical calculations and electron optics are often required to realize the full potential of ultrafast EEL spectrum imaging. In this review, we present a comprehensive overview of the recent progress in the theory and instrumentation of time-resolved and ultrafast EELS. We begin with an introduction to the technique, followed by a physical description of the loss function. We outline approaches for calculating and interpreting ground-state and transient EEL spectra spanning low-loss plasmons to core-level excitations analogous to X-ray absorption. We then survey the current state of time-resolved and ultrafast EELS techniques beyond photon-induced near-field electron microscopy, highlighting abilities to image carrier and thermal dynamics. Finally, we examine future directions enabled by emerging technologies, including electron beam monochromation, in situ and operando cells, laser-free UEM, and high-speed direct electron detectors. These advances position time-resolved and ultrafast EELS as a critical tool for uncovering nanoscale dynamic processes in quantum materials and solar energy conversion devices.

cond-mat.mtrl-sci↗

arXiv2025

Progressive Multimodal Search and Reasoning for Knowledge-Intensive Visual Question Answering

Knowledge-intensive visual question answering (VQA) requires external knowledge beyond image content, demanding precise visual grounding and coherent integration of visual and textual information. Although multimodal retrieval-augmented generation has achieved notable advances by incorporating external knowledge bases, existing approaches largely adopt single-pass frameworks that often fail to acquire sufficient knowledge and lack mechanisms to revise misdirected reasoning. We propose PMSR (Progressive Multimodal Search and Reasoning), a framework that progressively constructs a structured reasoning trajectory to enhance both knowledge acquisition and synthesis. PMSR uses dual-scope queries conditioned on both the latest record and the trajectory to retrieve diverse knowledge from heterogeneous knowledge bases. The retrieved evidence is then synthesized into compact records via compositional reasoning. This design facilitates controlled iterative refinement, which supports more stable reasoning trajectories with reduced error propagation. Extensive experiments across six diverse benchmarks (Encyclopedic-VQA, InfoSeek, MMSearch, LiveVQA, FVQA, and OK-VQA) demonstrate that PMSR consistently improves both retrieval recall and end-to-end answer accuracy.

cs.CV↗

arXiv2025

Resonant Self-Diffraction of Femtosecond Extreme Ultraviolet Pulses in Cobalt

Self-diffraction is a non-collinear four-wave mixing technique well-known in optics. We explore self-diffraction in the extreme ultraviolet (EUV) range, taking advantage of intense femtosecond EUV pulses produced by a free electron laser. Two pulses are crossed in a thin cobalt film and their interference results in a spatially periodic electronic excitation. The diffraction of one of the same pulses by the associated refractive index modulation is measured as a function of the EUV wavelength. A sharp peak in the self-diffraction efficiency is observed at the M$_{2,3}$ absorption edge of cobalt at 59 eV and a fine structure is found above the edge. The results are compared with a theoretical model assuming that the excitation results in an increase of the electronic temperature. EUV self-diffraction offers a potentially useful spectroscopy tool and will be instrumental in studying coherent effects in the EUV range.

physics.optics↗

arXiv2024

Task-Specific Preconditioner for Cross-Domain Few-Shot Learning

Cross-Domain Few-Shot Learning~(CDFSL) methods typically parameterize models with task-agnostic and task-specific parameters. To adapt task-specific parameters, recent approaches have utilized fixed optimization strategies, despite their potential sub-optimality across varying domains or target tasks. To address this issue, we propose a novel adaptation mechanism called Task-Specific Preconditioned gradient descent~(TSP). Our method first meta-learns Domain-Specific Preconditioners~(DSPs) that capture the characteristics of each meta-training domain, which are then linearly combined using task-coefficients to form the Task-Specific Preconditioner. The preconditioner is applied to gradient descent, making the optimization adaptive to the target task. We constrain our preconditioners to be positive definite, guiding the preconditioned gradient toward the direction of steepest descent. Empirical evaluations on the Meta-Dataset show that TSP achieves state-of-the-art performance across diverse experimental scenarios.

cs.LG↗

arXiv2024

Nanoscale and Element-Specific Lattice Temperature Measurements using Core-Loss Electron Energy-Loss Spectroscopy

Measuring nanoscale local temperatures, particularly in vertically integrated and multi-component systems, remains challenging. Spectroscopic techniques like X-ray absorption and core-loss electron energy-loss spectroscopy (EELS) are sensitive to lattice temperature, but understanding thermal effects is nontrivial. This work explores the potential for nanoscale and element-specific core-loss thermometry by comparing the Si L2,3 edge's temperature-dependent redshift against plasmon energy expansion thermometry (PEET) in a scanning TEM. Using density functional theory (DFT), time-dependent DFT, and the Bethe-Salpeter equation, we ab initio model both the Si L2,3 and plasmon redshift. We find that the core-loss redshift occurs due to bandgap reduction from electron-phonon renormalization. Our results indicate that despite lower core-loss signal intensity compared to plasmon features, core-loss thermometry has key advantages and can be more accurate through standard spectral denoising. Specifically, we show that the Varshni equation easily interprets the core-loss redshift for semiconductors, which avoids plasmon spectral convolution for PEET in complex junctions and interfaces. We also find that core-loss thermometry is more accurate than PEET at modeling thermal lattice expansion in semiconductors, unless the specimen's temperature-dependent dielectric properties are fully characterized. Furthermore, core-loss thermometry has the potential to measure nanoscale heating in multi-component materials and stacked interfaces with elemental specificity at length scales smaller than the plasmon's wavefunction.

cond-mat.mtrl-sci↗

arXiv2024

SAAS: Solving Ability Amplification Strategy for Enhanced Mathematical Reasoning in Large Language Models

This study presents a novel learning approach designed to enhance both mathematical reasoning and problem-solving abilities of Large Language Models (LLMs). We focus on integrating the Chain-of-Thought (CoT) and the Program-of-Thought (PoT) learning, hypothesizing that prioritizing the learning of mathematical reasoning ability is helpful for the amplification of problem-solving ability. Thus, the initial learning with CoT is essential for solving challenging mathematical problems. To this end, we propose a sequential learning approach, named SAAS (Solving Ability Amplification Strategy), which strategically transitions from CoT learning to PoT learning. Our empirical study, involving an extensive performance comparison using several benchmarks, demonstrates that our SAAS achieves state-of-the-art (SOTA) performance. The results underscore the effectiveness of our sequential learning approach, marking a significant advancement in the field of mathematical reasoning in LLMs.

cs.CL↗

arXiv2024

Harnessing Large Language Model to collect and analyze Metal-organic framework property dataset

This research was focused on the efficient collection of experimental Metal-Organic Framework (MOF) data from scientific literature to address the challenges of accessing hard-to-find data and improving the quality of information available for machine learning studies in materials science. Utilizing a chain of advanced Large Language Models (LLMs), we developed a systematic approach to extract and organize MOF data into a structured format. Our methodology successfully compiled information from more than 40,000 research articles, creating a comprehensive and ready-to-use dataset. The findings highlight the significant advantage of incorporating experimental data over relying solely on simulated data for enhancing the accuracy of machine learning predictions in the field of MOF research.

cond-mat.mtrl-sci↗

arXiv2024

An Image Grid Can Be Worth a Video: Zero-shot Video Question Answering Using a VLM

Stimulated by the sophisticated reasoning capabilities of recent Large Language Models (LLMs), a variety of strategies for bridging video modality have been devised. A prominent strategy involves Video Language Models (VideoLMs), which train a learnable interface with video data to connect advanced vision encoders with LLMs. Recently, an alternative strategy has surfaced, employing readily available foundation models, such as VideoLMs and LLMs, across multiple stages for modality bridging. In this study, we introduce a simple yet novel strategy where only a single Vision Language Model (VLM) is utilized. Our starting point is the plain insight that a video comprises a series of images, or frames, interwoven with temporal information. The essence of video comprehension lies in adeptly managing the temporal aspects along with the spatial details of each frame. Initially, we transform a video into a single composite image by arranging multiple frames in a grid layout. The resulting single image is termed as an image grid. This format, while maintaining the appearance of a solitary image, effectively retains temporal information within the grid structure. Therefore, the image grid approach enables direct application of a single high-performance VLM without necessitating any video-data training. Our extensive experimental analysis across ten zero-shot video question answering benchmarks, including five open-ended and five multiple-choice benchmarks, reveals that the proposed Image Grid Vision Language Model (IG-VLM) surpasses the existing methods in nine out of ten benchmarks.

cs.CV↗

arXiv2024

Improving Forward Compatibility in Class Incremental Learning by Increasing Representation Rank and Feature Richness

Class Incremental Learning (CIL) constitutes a pivotal subfield within continual learning, aimed at enabling models to progressively learn new classification tasks while retaining knowledge obtained from prior tasks. Although previous studies have predominantly focused on backward compatible approaches to mitigate catastrophic forgetting, recent investigations have introduced forward compatible methods to enhance performance on novel tasks and complement existing backward compatible methods. In this study, we introduce an effective-Rank based Feature Richness enhancement (RFR) method, designed for improving forward compatibility. Specifically, this method increases the effective rank of representations during the base session, thereby facilitating the incorporation of more informative features pertinent to unseen novel tasks. Consequently, RFR achieves dual objectives in backward and forward compatibility: minimizing feature extractor modifications and enhancing novel task performance, respectively. To validate the efficacy of our approach, we establish a theoretical connection between effective rank and the Shannon entropy of representations. Subsequently, we conduct comprehensive experiments by integrating RFR into eleven well-known CIL methods. Our results demonstrate the effectiveness of our approach in enhancing novel-task performance while mitigating catastrophic forgetting. Furthermore, our method notably improves the average incremental accuracy across all eleven cases examined.

cs.LG↗

arXiv2024

Determining Quasi-Equilibrium Electron and Hole Distributions of Plasmonic Photocatalysts using Photomodulated X-ray Absorption Spectroscopy

Most photocatalytic and photovoltaic devices operate under broadband, constant illumination. Electron and hole dynamics in these devices, however, are usually measured using ultrafast pulsed lasers in a narrow wavelength range. In this work, we prove that steady-state, photomodulated X-ray spectra from a non-time-resolved synchrotron beamline can be used to estimate electron and hole distributions. A set of plasmonic metal core-shell nanoparticles is designed to systematically isolate photothermal, hot electron, and thermalized electron-hole pairs in a TiO2 shell. Steady-state changes in the Ti L2,3 edge are measured with and without continuous-wave illumination of the nanoparticle's localized surface plasmon resonance. Ab initio excited-state X-ray theory developed for transient X-ray measurements is then applied to model the experimental spectra in an attempt to extract the resultant steady-state carrier distributions. The results suggest that, within error, the quasi-equilibrium carrier distribution can be determined even from relatively noisy data with mixed excited-state phenomena.

physics.chem-ph↗

arXiv2023

Graph-Network-Based Predictive Modeling for Highly Cross-Linked Polymer Systems

In this study, a versatile methodology for initiating polymerization from monomers in highly cross-linked materials is investigated. As polymerization progresses, force-field parameters undergo continuous modification due to the formation of new chemical bonds. This dynamic process not only impacts the atoms directly involved in bonding, but also influences the neighboring atomic environment. Monitoring these complex changes in highly cross-linked structures poses a challenge. To address this issue, we introduce a graph-network-based algorithm that offers both rapid and accurate predictions. The algorithm merges polymer construction protocols with LAMMPS, a large-scale molecular dynamics simulation software. The adaptability of this code has been demonstrated by its successful application to various amorphous polymers, including porous polymer networks (PPNs), and epoxy-resins, while the algorithm has been employed for additional tasks, such as implementing pore-piercing deformations and calculating material properties.

cs.CE↗

arXiv2023

MatGD: Materials Graph Digitizer

We have developed MatGD (Material Graph Digitizer), which is a tool for digitizing a data line from scientific graphs. The algorithm behind the tool consists of four steps: (1) identifying graphs within subfigures, (2) separating axes and data sections, (3) discerning the data lines by eliminating irrelevant graph objects and matching with the legend, and (4) data extraction and saving. From the 62,534 papers in the areas of batteries, catalysis, and MOFs, 501,045 figures were mined. Remarkably, our tool showcased performance with over 99% accuracy in legend marker and text detection. Moreover, its capability for data line separation stood at 66%, which is much higher compared to other existing figure mining tools. We believe that this tool will be integral to collecting both past and future data from publications, and these data can be used to train various machine learning models that can enhance material predictions and new materials discovery.

cs.CV↗

arXiv2023

Using Electron Energy-Loss Spectroscopy to Measure Nanoscale Electronic and Vibrational Dynamics in a TEM

Electron energy-loss spectroscopy (EELS) can measure similar information to X-ray, UV-Vis, and IR spectroscopies but with atomic resolution and increased scattering cross sections. Recent advances in electron monochromators have expanded EELS capabilities from chemical identification to the realms of synchrotron-level core-loss measurements and to low-loss, 10-100 meV excitations such as phonons, excitons, and valence structure. EELS measurements are easily correlated with electron diffraction and atomic-scale real-space imaging in a transmission electron microscope (TEM) to provide detailed local pictures of quasiparticle and bonding states. This perspective provides an overview of existing high-resolution EELS (HR-EELS) capabilities while also motivating the powerful next step in the field - ultrafast EELS in a TEM. Ultrafast EELS aims to combine atomic level, element specific, and correlated temporal measurements to better understand spatially specific excited state phenomena. Ultrafast EELS measurements also add to the abilities of steady-state HR-EELS by being able to image the electromagnetic field and use electrons to excite photon-forbidden and momentum-specific transitions. We discuss the technical challenges ultrafast HR-EELS currently faces, as well as how integration with in situ and cryo measurements could expand the technique to new systems of interest, especially molecular and biological samples.

physics.chem-ph↗

arXiv2021

Core-Shell Bimetallic Nanoparticle Trimers for Efficient Light-to-Chemical Energy Conversion

Incorporation of catalytically active materials into plasmonic metal nanostructures can efficiently merge the reactivity and energy harvesting abilities of both types of materials for visible light photocatalysis. Here we explore the influence of electromagnetic hotspots in the ability of plasmonic core-shell colloidal structures to induce chemical transformations. For this study, we developed a synthetic strategy for the fabrication of Au nanoparticle (NP) trimers in aqueous solution through fine controlled galvanic replacement between Ag nanoprisms and Au precursors. Core-shell Au@M NP trimers with catalytically active metals (M = Pd, Pt) were subsequently synthesized using Au NP trimers as templates. Our experimental and computational results highlight the synergy of geometry and composition in plasmonic catalysts for plasmon-driven chemical reactions.

physics.chem-ph↗

arXiv2020

Semantics-Preserving Adversarial Training

Adversarial training is a defense technique that improves adversarial robustness of a deep neural network (DNN) by including adversarial examples in the training data. In this paper, we identify an overlooked problem of adversarial training in that these adversarial examples often have different semantics than the original data, introducing unintended biases into the model. We hypothesize that such non-semantics-preserving (and resultingly ambiguous) adversarial data harm the robustness of the target models. To mitigate such unintended semantic changes of adversarial examples, we propose semantics-preserving adversarial training (SPAT) which encourages perturbation on the pixels that are shared among all classes when generating adversarial examples in the training stage. Experiment results show that SPAT improves adversarial robustness and achieves state-of-the-art results in CIFAR-10 and CIFAR-100.

cs.LG↗