arXiv subjects

Timothy A. Miller

Publications and source records attributed to Timothy A. Miller.

13 recordsLinked to original sources

arXiv2026

LogosKG: Hardware-Optimized Scalable and Interpretable Knowledge Graph Retrieval

Knowledge graphs (KGs) are increasingly integrated with large language models (LLMs) to provide structured, verifiable reasoning. A core operation in this integration is multi-hop retrieval, yet existing systems struggle to balance efficiency, scalability, and interpretability. We introduce LogosKG, a novel, hardware-aligned framework that enables scalable and interpretable k-hop retrieval on large KGs by building on symbolic KG formulations and executing traversal as hardware-efficient operations over decomposed subject, object, and relation representations. To scale to billion-edge graphs, LogosKG integrates degree-aware partitioning, cross-graph routing, and on-demand caching. Experiments show substantial efficiency gains over CPU and GPU baselines without loss of retrieval fidelity. With proven performance in KG retrieval, a downstream two-round KG-LLM interaction demonstrates how LogosKG enables large-scale, evidence-grounded analysis of how KG topology, such as hop distribution and connectivity, shapes the alignment between structured biomedical knowledge and LLM diagnostic reasoning, thereby opening the door for next-generation KG-LLM integration. The source code is publicly available at https://github.com/LARK-NLP-Lab/LogosKG, and an online demo is available at https://lark-nlp-lab-logoskg.hf.space/.

cs.CL↗

arXiv2026

CLSGen: A Dual-Head Fine-Tuning Framework for Joint Probabilistic Classification and Verbalized Explanation

With the recent progress of Large Language Models (LLMs), there is a growing interest in applying these models to solve complex and challenging problems. Modern LLMs, capable of processing long contexts and generating verbalized explanations, offer significant potential in addressing real-world applications. However, a critical hurdle in deploying LLMs for practical decision-making is their inability to provide reliable, quantitative probabilities. While task-specific fine-tuning of LLMs using traditional discriminative objectives (similar to encoder-only models) can yield probability estimates, this often leads to catastrophic forgetting and linguistic collapse. Consequently, the model loses its ability to generate explanations, severely undermining its interpretability and usability. To address this challenge, we propose CLSGen, a novel LLM fine-tuning framework designed for binary classification tasks. The CLSGen framework encompasses a new model architecture, training methodology, and data construction strategy to enable robust probability estimation without sacrificing the model's inherent explanation-generation capabilities. Experimental results across multiple benchmark datasets demonstrate that models fine-tuned with CLSGen outperform existing baselines in classification metrics (AUROC and F1-score). Regarding explanation, the results showed strong alignment between predicted labels and generated justifications, as well as high readability.

cs.CL↗

arXiv2025

Evaluating Retrieval-Augmented Generation vs. Long-Context Input for Clinical Reasoning over EHRs

Objective: To evaluate whether retrieval-augmented generation (RAG) can serve as an efficient alternative to long-context prompting for clinical reasoning over electronic health records (EHRs). Methods: We defined three EHR-based tasks that are replicable across health systems and vary in reasoning complexity: 1) extracting imaging procedures (modality, date, and anatomic site), 2) generating timelines of therapeutic antibiotic use, and 3) identifying the key diagnoses for a hospitalization. Using real inpatient clinical notes from a US academic health system, we evaluated three large language models (GPT-5.4-mini, Mistral Medium 3, DeepSeek V3.1) with varying amounts of provided context, comparing targeted retrieval to using the most recent clinical notes. Results: For Imaging Procedures, RAG strongly outperformed recent-note inputs and exceeded long-context performance (by 0.17-9.83 F1 across all models) using fewer than 8K tokens. Similar benefits were observed for Antibiotic Timelines, where <8K of retrieved tokens matched long-context recent-notes performance (between -3.26 to +3.24 Jaccard). Error analysis revealed that missing information in the clinical notes--often due to inter-hospital transfers--limited performance to some extent. However, performance on the Diagnosis Generation task remains largely static across methods and models. Discussion: RAG demonstrated strong token efficiency across tasks, with the clearest and most consistent gains observed for imaging extraction and antibiotic timeline reconstruction. Diagnosis generation proved the most challenging task, suggesting ceiling effects imposed by documentation variability and evaluation constraints. Conclusion: Our results suggest that RAG remains a competitive and efficient approach for clinical tasks over large amounts of EHR, even as newer models become capable of handling increasingly longer amounts of text.

cs.CL↗

arXiv2025

Using tournaments to calculate AUROC for zero-shot classification with LLMs

Large language models perform surprisingly well on many zero-shot classification tasks, but are difficult to fairly compare to supervised classifiers due to the lack of a modifiable decision boundary. In this work, we propose and evaluate a method that transforms binary classification tasks into pairwise comparisons between instances within a dataset, using LLMs to produce relative rankings of those instances. Repeated pairwise comparisons can be used to score instances using the Elo rating system (used in chess and other competitions), inducing a confidence ordering over instances in a dataset. We evaluate scheduling algorithms for their ability to minimize comparisons, and show that our proposed algorithm leads to improved classification performance, while also providing more information than traditional zero-shot classification.

cs.CL↗

arXiv2024

Lessons Learned on Information Retrieval in Electronic Health Records: A Comparison of Embedding Models and Pooling Strategies

Objective: Applying large language models (LLMs) to the clinical domain is challenging due to the context-heavy nature of processing medical records. Retrieval-augmented generation (RAG) offers a solution by facilitating reasoning over large text sources. However, there are many parameters to optimize in just the retrieval system alone. This paper presents an ablation study exploring how different embedding models and pooling methods affect information retrieval for the clinical domain. Methods: Evaluating on three retrieval tasks on two electronic health record (EHR) data sources, we compared seven models, including medical- and general-domain models, specialized encoder embedding models, and off-the-shelf decoder LLMs. We also examine the choice of embedding pooling strategy for each model, independently on the query and the text to retrieve. Results: We found that the choice of embedding model significantly impacts retrieval performance, with BGE, a comparatively small general-domain model, consistently outperforming all others, including medical-specific models. However, our findings also revealed substantial variability across datasets and query text phrasings. We also determined the best pooling methods for each of these models to guide future design of retrieval systems. Discussion: The choice of embedding model, pooling strategy, and query formulation can significantly impact retrieval performance and the performance of these models on other public benchmarks does not necessarily transfer to new domains. Further studies such as this one are vital for guiding empirically-grounded development of retrieval frameworks, such as in the context of RAG, for the clinical domain.

cs.CL↗

arXiv2023

Considerations for health care institutions training large language models on electronic health records

Large language models (LLMs) like ChatGPT have excited scientists across fields; in medicine, one source of excitement is the potential applications of LLMs trained on electronic health record (EHR) data. But there are tough questions we must first answer if health care institutions are interested in having LLMs trained on their own data; should they train an LLM from scratch or fine-tune it from an open-source model? For healthcare institutions with a predefined budget, what are the biggest LLMs they can afford? In this study, we take steps towards answering these questions with an analysis on dataset sizes, model sizes, and costs for LLM training using EHR data. This analysis provides a framework for thinking about these questions in terms of data scale, compute scale, and training budgets.

cs.CY↗

arXiv2023

Natural language processing to automatically extract the presence and severity of esophagitis in notes of patients undergoing radiotherapy

Radiotherapy (RT) toxicities can impair survival and quality-of-life, yet remain under-studied. Real-world evidence holds potential to improve our understanding of toxicities, but toxicity information is often only in clinical notes. We developed natural language processing (NLP) models to identify the presence and severity of esophagitis from notes of patients treated with thoracic RT. We fine-tuned statistical and pre-trained BERT-based models for three esophagitis classification tasks: Task 1) presence of esophagitis, Task 2) severe esophagitis or not, and Task 3) no esophagitis vs. grade 1 vs. grade 2-3. Transferability was tested on 345 notes from patients with esophageal cancer undergoing RT. Fine-tuning PubmedBERT yielded the best performance. The best macro-F1 was 0.92, 0.82, and 0.74 for Task 1, 2, and 3, respectively. Selecting the most informative note sections during fine-tuning improved macro-F1 by over 2% for all tasks. Silver-labeled data improved the macro-F1 by over 3% across all tasks. For the esophageal cancer notes, the best macro-F1 was 0.73, 0.74, and 0.65 for Task 1, 2, and 3, respectively, without additional fine-tuning. To our knowledge, this is the first effort to automatically extract esophagitis toxicity severity according to CTCAE guidelines from clinic notes. The promising performance provides proof-of-concept for NLP-based automated detailed toxicity monitoring in expanded domains.

cs.CL↗

arXiv2017

Nanoscale spectroscopic imaging of phase separation in a correlated material by resonant X-ray holography

Nanoscale phase coexistence and inhomogeneity are ubiquitous in correlated electron materials, existing in doped Mott insulators, manganites, and high-temperature superconductors. The small length scales and lack of contrast mechanisms make it extremely challenging to measure real-space images of the phase coexistence with high resolution. Yet, images aid our understanding of how phase coexistence and domain boundaries dictate the exotic material properties. Here we show that resonant soft-X-ray holography, previously employed to image magnetic domains, can be used to image phase separation on the nanoscale. We observe nucleation and growth of nanometre-sized metallic domains out of the insulating phase of the prototypical correlated material VO2, using linearly polarized coherent synchrotron radiation. By spectrally resolving the holograms, we extract differential soft X-ray absorption spectra with 50 nm spatial resolution. Furthermore, when combined with ultra-bright and ultra-short X-ray sources, X-ray holography could capture both nanoscale spatial variations and temporal fluctuations that occur close to the transition temperatures or are induced by femtosecond light pulses that cannot be observed with alternative imaging methods.

cond-mat.mes-hall↗

arXiv2016

Resonant optical control of the structural distortions that drive ultrafast demagnetization in Cr$_2$O$_3$

We study how the color and polarization of ultrashort pulses of visible light can be used to control the demagnetization processes of the antiferromagnetic insulator Cr$_2$O$_3$. We utilize time-resolved second harmonic generation (SHG) to probe how changes in the magnetic and structural state evolve in time. We show that, varying the pump photon-energy to excite either localized transitions within the Cr or charge transfer states, leads to markedly different dynamics. Through a full polarization analysis of the SHG signal, symmetry considerations and density functional theory calculations, we show that, in the non-equilibrium state, SHG is sensitive to {\em both} lattice displacements and changes to the magnetic order, which allows us to conclude that different excited states couple to phonon modes of different symmetries. Furthermore, the spin-scattering rate depends on the induced distortion, enabling us to control the timescale for the demagnetization process. Our results suggest that selective photoexcitation of antiferromagnetic insulators allows fast and efficient manipulation of their magnetic state.

cond-mat.mtrl-sci↗

arXiv2016

The Ultrafast Optical Response of the Amorphous and Crystalline States of the Phase Change Material Ge$_2$Sb$_2$Te$_5$

We examine the ultrafast optical response of the crystalline and amorphous phases of the phase change material Ge$_2$Sb$_2$Te$_5$ below the phase transformation threshold. Simultaneous measurement of the transmissivity and reflectivity of thin film samples yields the time-dependent evolution of the dielectric function for both phases. We then identify how lattice motion and electronic excitation manifest in the dielectric response. The dielectric response of both phases is large but markedly different. At 800 nm, the changes in amorphous GST are well described by the Drude response of the generated photo-carriers, whereas the crystalline phase is better described by the depopulation of resonant bonds. We find that the generated coherent phonons have a greater influence in the amorphous phase than the crystalline phase. Furthermore, coherent phonons do not influence resonant bonding. For fluences up to 50% of the transformation threshold, the structure does not exhibit bond softening in either phase, enabling large changes of the optical properties without structural modification.

cond-mat.mtrl-sci↗

arXiv2015

Ultrafast broadband tuning of resonant optical nanostructures using phase change materials

The functionalities of a wide range of optical and opto-electronic devices are based on resonance effects and active tuning of the amplitude and wavelength response is often essential. Plasmonic nanostructures are an efficient way to create optical resonances, a prominent example is the extraordinary optical transmission (EOT) through arrays of nanoholes patterned in a metallic film. Tuning of resonances by heating, applying electrical or optical signals has proven to be more elusive, due to the lack of materials that can induce modulation over a broad spectral range and/or at high speeds. Here we show that nanopatterned metals combined with phase change materials (PCMs) can overcome this limitation due to the large change in optical constants which can be induced thermally or on an ultrafast timescale. We demonstrate resonance wavelength shifts as large as 385 nm - an order of magnitude higher than previously reported - by combining properly designed Au EOT nanostructures with Ge2Sb2Te5 (GST). Moreover, we show, through pump probe measurements, repeatable and reversible, large amplitude modulations in the resonances, especially at telecommunication wavelengths, over ps time scales and at powers far below those needed to produce a permanent phase transition. Our findings open a pathway to the design of hybrid metal PCM nanostructures with ultrafast and widely tuneable resonance responses, which hold potential impact on active nanophotonic devices such as tuneable optical filters, smart windows, biosensors and reconfigurable memories.

physics.optics↗

arXiv2015

THz Field Control of In-Plane Orbital Order in La0.5Sr1.5MnO4

In-plane anisotropic ground states are ubiquitous in correlated solids such as pnictides, cuprates and manganites. They can arise from doping Mott insulators and compete with phases such as superconductivity, however their origins are debated. Strong coupling between lattice, charge, orbital and spin degrees of freedom results in simultaneous ordering of multiple parameters, masking the mechanism that drives the transition. We demonstrate that the anisotropic orbital domains in a manganite can be oriented by the polarization of a pulsed THz light field. Through the application of the Hubbard model, we show that domain control can be achieved either through field assisted hopping of charges or a field-induced modification of bond angles. Both routes enhance the local Coulomb repulsions which drive domain reorientation and the dominant mechanism is dictated by the equilibrium Mn-O-Mn bond angle. Our results highlight the key role played by the Coulomb interaction in driving orbital order in manganites and demonstrate how THz can be utilized in new ways to understand and manipulate anisotropic phases in a broad range of correlated materials.

cond-mat.str-el↗

arXiv2015

Time-Domain Separation of Optical Properties From Structural Transitions in Resonantly Bonded Materials

The extreme electro-optical contrast between crystalline and amorphous states in phase change materials is routinely exploited in optical data storage and future applications include universal memories, flexible displays, reconfigurable optical circuits, and logic devices. Optical contrast is believed to arise due to a change in crystallinity. Here we show that the connection between optical properties and structure can be broken. Using a unique combination of single-shot femtosecond electron diffraction and optical spectroscopy, we simultaneously follow the lattice dynamics and dielectric function in the phase change material Ge2Sb2Te5 during an irreversible state transformation. The dielectric function changes by 30% within 100 femtoseconds due to a rapid depletion of electrons from resonantly-bonded states. This occurs without perturbing the crystallinity of the lattice, which heats with a 2 ps time constant. The optical changes are an order-of-magnitude larger than those achievable with silicon and present new routes to manipulate light on an ultrafast timescale without structural changes.

cond-mat.mtrl-sci↗