Institution profile

University of Maryland, Baltimore County

Academic institutionnorthamerica · us
Official website
Research library309linked papers
Opportunities0open roles
Selected work

Representative Papers

SCRIBE: Structured Mid-Level Supervision for Tool-Using Language Models

Jan 07, 2026arXiv.org

This work proposes SCRIBE, a framework addressing the training instability of tool-augmented language models in multi-step reasoning caused by credit assignment difficulties. SCRIBE constructs a structured skill prototype library to validate subgoals at an intermediate level of abstraction, transforming open-ended LLM evaluation into a constrained verification task that yields precise, low-variance process-level reward signals. It introduces structured supervision at the skill-abstraction level for the first time, decoupling high-level planning from low-level execution and revealing a co-evolutionary mechanism wherein mid-level skill mastery precedes the emergence of high-level planning capabilities. Experiments demonstrate state-of-the-art performance across multiple reasoning and tool-use benchmarks: Qwen3-4B achieves 63.3% accuracy on AIME25, a substantial improvement from 43.3%, significantly boosting success rates in complex, multi-turn tool interactions and showing strong compatibility with underlying tool optimization methods.

3 citationsRead paper

Differentially Private Synthetic Data Generation Using Context-Aware GANs

Dec 15, 2024BigData Congress [Services Society]

This work addresses the dual challenge of generating synthetic data that simultaneously satisfies differential privacy (DP) requirements—ensuring compliance with GDPR and HIPAA—and preserves domain-specific rule consistency (e.g., drug contraindications, regulatory logic)—critical in healthcare, finance, and security. We propose a constraint-matrix-driven, context-aware discriminator for DP-GANs, the first to jointly encode explicit statistical constraints and implicit domain knowledge into the DP-GAN framework. Built upon the WGAN architecture, our method integrates: (i) constraint-matrix-based modeling of domain rules; (ii) a differentially private discriminator employing gradient clipping and Gaussian noise injection; and (iii) domain-knowledge-guided adversarial loss. Experiments demonstrate robust resistance to membership inference attacks under ε ≤ 2.0, a 38% improvement in clinical plausibility (as measured by expert validation rate), and a 22% average gain in downstream task F1 scores—achieving synergistic optimization of privacy preservation and domain-rule fidelity.

3 citationsRead paper

Bayesian Meta-Analyses Could Be More: A Case Study in Trial of Labor After a Cesarean-section Outcomes and Complications

Jan 15, 2026

Traditional meta-analysis struggles to reliably estimate intervention effects when key covariates are missing, limiting its applicability in clinical decision-making contexts such as trial of labor after cesarean (TOLAC). To address this challenge, this work proposes a Bayesian meta-analytic approach that integrates prior knowledge with limited observed data to enable robust inference of treatment effects under missing covariate conditions while quantifying associated uncertainty. The method substantially enhances the reliability and practical utility of effect estimates, providing strong evidence to support individualized clinical decisions in TOLAC scenarios. By offering a principled framework for drawing valid conclusions even with incomplete data, it bolsters clinician confidence and strengthens the evidence base—particularly in settings where intervention options are constrained.

1 citationsRead paper

In Context Learning and Reasoning for Symbolic Regression with Large Language Models

Oct 22, 2024arXiv.org

Symbolic regression—discovering concise, accurate, and interpretable mathematical equations from data—remains challenging for large language models (LLMs) due to their limited grounding in scientific reasoning and symbolic manipulation. Method: We propose a closed-loop iterative reasoning framework that integrates scientific context modeling and chain-of-thought prompting within a scratchpad mechanism, enabling GPT-4 to collaboratively generate candidate expressions using data, historical formulas, and physical priors. Candidates undergo symbolic verification via SymPy and numerical optimization via SciPy, with feedback-driven iterative refinement. Contribution/Results: This work achieves the first end-to-end integration of natural-language scientific knowledge with verifiable symbolic optimization. Our method successfully recovers five canonical physics equations and generates plausible expressions on unseen datasets. Incorporating contextual grounding and scratchpad reasoning significantly improves accuracy, demonstrating that LLMs can serve as effective auxiliary reasoning engines for interpretable scientific modeling.

1 citationsRead paper
Recent publications

Latest Papers