Institution profile

University of Maryland

Academic institutionnorthamerica · us
Official website
Research library1,808linked papers
Opportunities0open roles
Selected work

Representative Papers

Federated Large Language Models: Current Progress and Future Directions

Sep 24, 2024arXiv.org

To address the convergence difficulties and high communication overhead of large language models (LLMs) in federated learning (FL) caused by data heterogeneity, this paper introduces FedLLM—the first unified analytical framework for LLMs in FL. It systematically surveys two dominant paradigms: federated fine-tuning and federated prompt learning, while rigorously analyzing core challenges including data heterogeneity, communication efficiency, and privacy preservation. The work identifies promising future directions—namely, federated pre-training and LLM-augmented FL—and fills a critical gap in systematic literature review. A multidimensional taxonomy and evaluation framework is established to clarify key technical bottlenecks. Integrating insights from FL, LLM adaptation, prompt engineering, distributed optimization, and privacy-preserving computation, this study delivers a practical, robust, and privacy-aware methodology for deploying LLMs in real-world federated settings. (149 words)

16 citations1 influentialRead paper

AdvPrefix: An Objective for Nuanced LLM Jailbreaks

Dec 13, 2024arXiv.org

Existing LLM jailbreaking attacks suffer from weak controllability, incomplete response generation, and rigid, non-adaptive optimization formats. To address these limitations, we propose AdvPrefix—a novel prefix-based objective function designed for fine-grained jailbreaking. AdvPrefix introduces the first model-adaptive prefix selection mechanism, which automatically identifies high-quality prefixes during the prefilling stage using a dual criterion: attack success rate and negative log-likelihood. It further enables multi-prefix collaborative optimization, departing from conventional fixed-prefix paradigms and exposing alignment models’ generalization vulnerabilities to unseen prefixes. AdvPrefix is fully compatible with mainstream optimization frameworks (e.g., GCG) and requires no model modification or retraining. Evaluated on Llama-3, AdvPrefix boosts GCG’s fine-grained jailbreaking success rate from 14% to 80%, demonstrating the critical impact of objective function design on jailbreaking efficacy.

10 citations3 influentialRead paper

Task Adaptation from Skills: Information Geometry, Disentanglement, and New Objectives for Unsupervised Reinforcement Learning

Jun 12, 2025International Conference on Learning Representations

Unsupervised reinforcement learning (URL) aims to acquire transferable skills for unknown downstream tasks, yet existing mutual information-based skill learning (MISL) lacks theoretical characterization of skill transferability. We identify that skill diversity and separability are essential for efficient downstream policy initialization—properties not guaranteed by MISL. To address this, we propose two novel objectives—WSEP and PWSEP—grounded in Wasserstein geometry, along with a decoupling-aware metric, LSEPIN. Crucially, we establish the first theoretical link between Wasserstein distance and downstream adaptation cost, rigorously proving that our framework ensures complete discovery of optimal initial policies. Experiments demonstrate significant improvements in zero-shot transfer performance across multiple benchmarks, consistently outperforming MISL. Our method yields more disentangled skill representations and superior policy pretraining, enabling more effective downstream adaptation.

7 citationsRead paper

MoE Lens -- An Expert Is All You Need

Mar 06, 2026

This work investigates the opaque expert specialization mechanism in Mixture-of-Experts (MoE) models, which limits inference and memory efficiency. By analyzing domain-specific routing patterns and employing an early-decoding framework, the study systematically examines how individual experts contribute to model outputs. Through comprehensive analyses—including routing distribution statistics, cosine similarity of hidden states, comparisons between single-expert and ensemble outputs, and perplexity evaluation—the authors find that a small subset of experts handles over 50% of all requests. Remarkably, outputs from a single dominant expert exhibit high consistency with the full model (cosine similarity up to 0.95), with only a 5% increase in perplexity. These findings suggest that precise expert pruning can substantially enhance inference efficiency without compromising performance, offering a promising avenue for efficient MoE deployment and knowledge localization.

5 citations1 influentialRead paper

On the origin of neural scaling laws: from random graphs to natural language

Jan 15, 2026

This study investigates whether neural scaling laws depend on power-law structures in data. By training a minimal Transformer—comprising only two layers and a context length of 50—to predict random walk sequences on synthetic graphs of tunable complexity (including Erdős–Rényi and Barabási–Albert models), the authors systematically simplify the language modeling task. They demonstrate, for the first time, that neural scaling laws emerge even in data devoid of power-law correlations, revealing a monotonic relationship between linguistic complexity and the scaling exponent. This finding establishes that power-law structures in data are not necessary for the emergence of scaling laws. Furthermore, the work introduces an alternative method for constructing compute-optimal curves and shows that the Maximal Update Parametrization (μP) achieves superior parameter efficiency.

3 citationsRead paper
Recent publications

Latest Papers