arXiv subjects

Gregory D. Lyng

Publications and source records attributed to Gregory D. Lyng.

7 recordsLinked to original sources

arXiv2026

MedicalBench: Evaluating Large Language Models Toward Improved Medical Concept Extraction

Medical concept extraction from electronic health records underpins many downstream applications, yet remains challenging because medically meaningful concepts are frequently implied rather than explicitly stated in medical narratives. Existing benchmarks with human-annotated evidence spans underscore the importance of grounding extracted concepts in medical text. However, they predominantly focus on explicitly stated concepts instead of implicit concepts. We present MedicalBench, a benchmark for medical concept extraction with evidence grounding that evaluates implicit medical reasoning. MedicalBench formulates medical concept extraction as a verification task over medical note-concept pairs, coupled with sentence-level evidence identification. Built from MIMIC-IV discharge summaries and human-verified ICD-10 codes, the dataset is curated through a multi-stage large language model (LLM) triage pipeline followed by medical annotation and expert review. It deliberately includes implicit positives, semantically confusable negatives, and cases where LLM judgments disagree with medical expert assessments. We define two complementary evaluation tasks: (1) medical concept extraction and (2) sentence-level evidence retrieval, enabling assessment of both correctness and interpretability. Benchmarking state-of-the-art LLMs reveals that performance remains modest, highlighting the difficulty of extracting implicitly expressed concepts. We further show that performance is largely invariant to note length, indicating that MedicalBench isolates reasoning difficulty rather than superficial confounders. MedicalBench provides the first systematic benchmark for implicit, evidence-grounded medical concept extraction, offering a foundation for developing medical language models that can both identify medically relevant concepts and justify their predictions in a transparent and medically faithful manner.

cs.CL↗

arXiv2026

Fast and Effective On-policy Distillation from Reasoning Prefixes

On-policy distillation (OPD), which samples trajectories from the student model and supervises them with a teacher at the token level, avoids relying solely on verifiable terminal rewards and can yield better generalization than off-policy distillation. However, OPD requires expensive on-the-fly sampling of the student policy during training, which substantially increases training cost, especially for long responses. Our initial analysis shows that, during OPD, training signals are often concentrated in the prefix of each output, and that even a short teacher-generated prefix can significantly help the student produce the correct answer. Motivated by these observations, we propose a simple yet effective modification of OPD: we apply the distillation objective only to prefixes of student-generated outputs and terminate each sampling early during distillation. Experiments on a suite of AI-for-Math and out-of-domain benchmarks show that on-policy prefix distillation matches the performance of full OPD while reducing training FLOP by 2x-47x.

cs.LG↗

arXiv2026

Health-SCORE: Towards Scalable Rubrics for Improving Health-LLMs

Rubrics are essential for evaluating open-ended LLM responses, especially in safety-critical domains such as healthcare. However, creating high-quality and domain-specific rubrics typically requires significant human expertise time and development cost, making rubric-based evaluation and training difficult to scale. In this work, we introduce Health-SCORE, a generalizable and scalable rubric-based training and evaluation framework that substantially reduces rubric development costs without sacrificing performance. We show that Health-SCORE provides two practical benefits beyond standalone evaluation: it can be used as a structured reward signal to guide reinforcement learning with safety-aware supervision, and it can be incorporated directly into prompts to improve response quality through in-context learning. Across open-ended healthcare tasks, Health-SCORE achieves evaluation quality comparable to human-created rubrics while significantly lowering development effort, making rubric-based evaluation and training more scalable.

cs.AI↗

arXiv2019

On the Dynamics of Traveling Fronts Arising in Nanoscale Pattern Formation

We study the stability and dynamics of traveling-front solutions of a modified Kuramoto--Sivashinsky equation arising in the modeling of nanoscale ripple patterns that form when a nominally flat solid surface is bombarded with a broad ion beam at an oblique angle of incidence. Structurally, the linearized operators associated with these fronts have unstable essential spectrum---corresponding to instability of the spatially asymptotic states---and stable point spectrum---corresponding to stability of the transition profile of the front. We show that these waves are linearly orbitally asymptotically stable in appropriate exponentially weighted spaces. While the technical device of exponential weights allows us to accommodate the unstable essential spectrum of individual waves in our linear analysis, it does not shed light on the long-time pattern formation that is observed experimentally and in numerical simulations. To begin to address this issue, we consider a periodic array of unstable front and back solutions. While not an exact solution of the governing equation, this periodic pattern mimics experimentally observed phenomena. Our numerical experiments suggest that the convecting instabilities associated with each individual wave are damped as they pass through transition layers and that this stabilization mechanism underlies the pattern formation seen in experiments.

math.AP↗

arXiv2013

The WKB approximation of semiclassical eigenvalues of the Zakharov-Shabat problem

We numerically compute eigenvalues of the non-self-adjoint Zakharov--Shabat problem in the semiclassical regime. In particular, we compute the eigenvalues for a Gaussian potential and compare the results to the corresponding (formal) WKB approximations used in the approach to the semiclassical or zero-dispersion limit of the focusing nonlinear Schroedinger equation via semiclassical soliton ensembles. This numerical experiment, taken together with recent numerical experiments [17,18], speaks directly to the viability of this approach; in particular, our experiment suggests a value for the rate of convergence of the WKB eigenvalues to the true eigenvalues in the semiclassical limit. This information provides some hint as to how these approximations might be rigorously incorporated into the asymptotic analysis of the singular limit for the associated nonlinear partial differential equation.

nlin.SI↗

arXiv2012

A second look at the Gaussian semiclassical soliton ensemble for the focusing nonlinear Schrödinger equation

We present the results of a numerical experiment inspired by the semiclassical (zero-dispersion) limit of the focusing nonlinear Schroedinger (NLS) equation. In particular, we focus on the Gaussian semiclassical soliton ensemble, a family of exact multisoliton solutions obtained by repeatedly solving the initial-value problem for a particular sequence of initial data. The sequence of data is generated by adding an asymptotically vanishing sequence of perturbations to pure Gaussian initial data. These perturbations are obtained by applying the inverse-scattering transform to formal WKB approximations of eigenvalues of the associated spectral problem with a Gaussian potential. Recent results [Lee, Lyng, & Vankova, Physica D 24 (2012):1767--1781] suggest that, remarkably, these perturbations---interlaced as they are with the integrable structure of the equation---do not excite the acute modulational instabilities that are known to be present in the semiclassical regime. Here, we provide additional evidence to support the claim that these WKB-induced perturbations indeed have a very special structure. In particular, as a control experiment, we examine the evolution from a family of initial data created by an asymptotically vanishing family of analytic perturbations which are qualitatively indistinguishable from the WKB-induced perturbations that generate the Gaussian semiclassical soliton ensemble. We then compare this evolution to the (numerically computed) true evolution of the Gaussian and also to the evolution of the corresponding members of the semiclassical soliton ensemble. Our results both highlight the exceptional nature of the WKB-induced perturbations used to generate the semiclassical soliton ensemble and provide new insight into the sensitivity properties of the semiclassical limit problem for the focusing NLS equation.

nlin.SI↗

arXiv2012

Computing the refined stability condition

The classical (inviscid) stability analysis of shock waves is based on the Lopatinski determinant, Δ---a function of frequencies whose zeros determine the stability of the underlying shock. A careful analysis of Δ shows that in some cases the stable and unstable regions of parameter space are separated by an open set of parameters. Zumbrun and Serre [Indiana Univ. Math. J., 48 (1999) 937--992] have shown that, by taking account of viscous effects not present in the definition of Δ, it is possible to determine the precise location in the open, neutral set of parameter space at which stability is lost. In particular, they show that the transition to instability under suitably localized perturbations is determined by an "effective viscosity" coefficient. Here, in the simplest possible setting, we propose and implement two new approaches toward the practical computation of this coefficient. Moreover, in a special case, we derive an exact solution of the relevant differential equations.

math.AP↗