Publications

Publication Archive

Selected and recent papers from TML and Prof. Sangwoo Hong.

Research Taxonomy

Publications organized by AI stack layer and, when applicable, a cross-layer reliability problem.

Layer

Reliable ComputingFoundation ModelsPhysical World

Reliability Problem

UnlearningPrivacyRobustnessAlignmentExplainability

2026

arXiv Foundation Models Robustness

From Token Correction to Claim Admission: Training-Free Hallucination Mitigation for Large Vision-Language Models

Hyeonggeun Han, Taehyun Cho, Sangwoo Hong

arXiv preprint, 2026

Introduces Evidence-Guided Visual Claim Admission, a training-free method that accepts generated claims only when they receive sufficient image-dependent support, reducing visual hallucination while preserving caption quality.

Paper Preprint

arXiv Foundation Models Privacy

Rethinking Memorization Mitigation in Diffusion Models: Reinforcing Text Conditioning

Hyungjun Joo, Sehwan Kim, Hyeonggeun Han, Sangwoo Hong, Jungwoo Lee

arXiv preprint, 2026

Introduces a training-free cross-attention intervention that reinforces text conditioning to reduce training-image reproduction while preserving prompt alignment.

Paper Preprint

arXiv Foundation Models Unlearning

Continual Data Unlearning in Diffusion Models via Transition-based Regularization

Sunbeom Jeong, Sehwan Kim, Sangwoo Hong, Jungwoo Lee

arXiv preprint, 2026

Introduces transition-based regularization that prevents later deletion updates from reversing earlier unlearning while preserving generative utility.

Paper Preprint

arXiv Foundation Models Unlearning

Mitigating Sequential Reappearance in Diffusion Data-Point Unlearning

Donghyun Kim, Taehyuk Lee, Jinyeong Kim, Youngmin Oh, Dohyeong Kim, Jaehyuk Ryu, Sangwoo Hong

arXiv preprint, 2026

Studies sequential reappearance in diffusion data-point unlearning and how deleted instances can remain forgotten across repeated unlearning requests.

Paper Preprint

NeurIPS Foundation Models Explainability

Mechanistic Circuit Identification for Controllable Data Generation

Nakyung Lee, Sangwoo Hong, Jungwoo Lee

The Fortieth Annual Conference on Neural Information Processing Systems, 2026

Identifies model-internal circuits associated with data utility and steers them to generate higher-quality, controllable training data for language models.

Publication Archive

Research Taxonomy

Layer

Reliability Problem

2026

From Token Correction to Claim Admission: Training-Free Hallucination Mitigation for Large Vision-Language Models

Rethinking Memorization Mitigation in Diffusion Models: Reinforcing Text Conditioning

Continual Data Unlearning in Diffusion Models via Transition-based Regularization

Mitigating Sequential Reappearance in Diffusion Data-Point Unlearning

Mechanistic Circuit Identification for Controllable Data Generation

Function-Level Execution Feedback for Code Preference Optimization

Geometry-Preserving Robust Neural Reconstruction via Statistical Reweighting

PredLM: A Sparse Large Language Model Decoding Accelerator with Hardware-Efficient Zero-Activation Prediction

Efficient Process Reward Modeling via Contrastive Mutual Information

Bias Alleviation through Network Pruning for Sparse and Debiased Models

Beyond Prompts: Preserving Semantics in Diffusion-based Communication

MVP-LAM: Learning Action-Centric Latent Action via Cross-Viewpoint Reconstruction

An Adaptive Sampling Framework for Diffusion-based Dataset Distillation with High Fidelity and Diversity

Operator-aware shadow importance sampling for accurate fidelity estimation

2025

Adjusting Initial Noise to Mitigate Memorization in Text-to-Image Diffusion Models

Mix from Failure: Confusion-Pairing Mixup for Long-Tailed Recognition

Prototype-based Explanation for Semantic Gap Reduction with Distributional Embedding

Constructing Fair Latent Space for Intersection of Fairness and Explainability

2024

Mitigating Spurious Correlations via Disagreement Probability

NeRFlex: Flexible Neural Radiance Fields with Diffeomorphic Deformation

Learning Dual Hierarchical Representation for 3D Surface Reconstruction

Group-wise Verifiable Coded Computing under Byzantine Attacks and Stragglers

GBMix: Enhancing Fairness by Group-Balanced Mixup

2023

Straggler-Exploiting Fully Private Distributed Matrix Multiplication With Chebyshev Polynomials

2022

Hierarchical Group Testing for Byzantine Attack Identification in Distributed Matrix Multiplication

2021

Securely Straggler-Exploiting Coded Computation for Distributed Matrix Multiplication

Chebyshev Polynomial Codes: Task Entanglement-based Coding for Distributed Matrix Multiplication

2020

Squeezed Polynomial Codes: Communication-Efficient Coded Computation in Straggler-Exploiting Distributed Matrix Multiplication