Llava Guard: An Open VLM-based Framework for Safeguarding Vision Datasets and Models
Reviewed:
VLM(Vision-Language Model): Text 및 Image 생성. Visual and Textual Inputs.
Posts by Collection
paper_reviews
Attention is All You Need
Reviewed:
기존의 방식은 병렬처리가 어려움. 병렬처리가 가능해도, 단어의 위치 정보가 손실됨.
Claude 3.7 Sonnet Systen Card
Reviewed:
거부 및 정책 위반 분류기뿐만 아니라 응답의 유용성을 측정하는 “유용성” 분류기를 사용해 응답을 평가함.
GPT-4 Technical Report
Reviewed:
이전 GPT 모델과 마찬가지로, 인간 피드백을 통한 강화 학습(RLHF, Reinforcement Learning from Human Feedback)을 사용해 produce response better aligned with user’s intent.
The Llama 3 Herd of Models
Reviewed:
Llama 3(.1) 모델의 Technical Report
Understanding the Statistical Accuracy-Communication Trade-off in Personalized Federated Learning with Minimax Guarantees, Xin Yu et al., ICML 2025
Reviewed:
FL의 중요한 문제점: data heterogeneity(데이터 이질성)
Synthetic Data from Diffusion Models Improves ImageNet Classification, Shekoofeh Azizi et al., TMLR 2023
Reviewed:
최근 denoising diffusion probabilistic models(DDPMs)가 GAN과 품질면애서 비교할 수 있는 이미지를 생성하며 학습 중 더 큰 안정성을 제공함.
Communication-Efficient Federated Data Augmentation on Non-IID Data
Reviewed:
Non-IID Dataset 에서 누락된 sample을 생성하기 위해 Conditional Variational AutoEncoder, CVAE를 채택함.
FRAug: Tackling Federated Learning with Non-IID Features via Representation Augmentation, ICCV 2023
Reviewed:
Federated Representation Augmentation, FRAug
Class-Balanced Loss Based on Effective Number of Samples, CVPR 2019
Reviewed:
long-tail: skewed distribution, 소수의 지배적인 class가 대부분의 예제를 차지하지만, 다른 대부분의 class는 상대적으로 적은 예제 - 데이터 불균형
Using Synthetic Data for Data Augmentation to Improve Classification Accuracy
Reviewed:
요약이 아직 작성되지 않았습니다.
Towards Active Synthetic Data Generation for Fine-tuning Language Models, ICLR 2026
Reviewed:
Towards Active Synthetic Data Generation for Finetuning Language Models
Federated Balanced Learning, CVPR 2026
Reviewed:
기존 연구들은 Non-IID를 최적화 단계(그래디언트/손실함수 수정)에서 문제를 해결하려 노력함 → Model drift 가 발생된 것을 교정하려는 시도, 근본적 문제(샘플의 불균형)을 해결하는 것이 아님
Balancing Cost and Effectiveness of Synthetic Data Generation Strategies for LLMs, NeurIPS ‘24 Workshop on Fine-Tuning in Modern Machine Learning: Principles and Scalability
Reviewed:
Balancing Cost and Effectiveness of Synthetic Data Generation…
Do Generated Data Always Help Contrastive Learning?, ICLR 2024
Reviewed:
인용 이유. Federated Balanced Learning에서 합성 데이터와 실제 데이터 간의 비율 또는 균형에 대한 탐색의 예시로 인용. 기존 연구 동향을 제시
Judging LLM-as-judge with MT-Bench and Chatbot Arena, NeurIPS 2024
Reviewed:
인용 이유. LLM-as-judge 방식을 활용해 합성 데이터의 난이도 및 품질을 평가하고 데이터를 선별하는 기존의 인기있는 방법을 언급.
Data-Free Knowledge Distillation for Heterogeneous Federated Learning, ICML 2021 PMLR 139
Reviewed:
FL의 데이터 이질성 - 일반적으로 비독립적이고 동일하게 분포되지 않은, Non-IID 방식으로 분포되어있어, 본질적으로 편향된 로컬 최적점을 유발함.
Jailbroken: How Does LLM Safety Training Fail? — Wei et al. (2024), NeurIPS 2023(Oral)
Reviewed:
📎 NeurIPS 2023 · arXiv:2307.02483 저자: Alexander Wei, Nika Haghtalab, Jacob Steinhardt (UC Berkeley) 우리 논문과의 관계: Type A/B 분류의 이론적 토대. Competing objectives ↔ Type B, mismatched generalization ↔ Type A로 대응시킬 수 있음.
Logit-Gap Steering: Efficient Short-Suffix Jailbreaks for Aligned Large Language Models — Li & Liu (2025), arXiv
Reviewed:
📎 arXiv:2506.24056 저자: Tung-Ling Li, Hongliang Liu 우리 논문과의 관계: 우리의 $St = \mu{cmp} - \mu{ref}$와 거의 동일한 logit-gap 정의를 공격에 사용. 우리는 진단에 사용. 같은 metric, 반대 목적. “Diagnostic vs. interventional” 구분의 핵심 사례.
Refusal in Language Models Is Mediated by a Single Direction — Arditi et al. (2024), NeurIPS 2024
Reviewed:
📎 NeurIPS 2024 · arXiv:2406.11717 저자: Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, Neel Nanda 우리 논문과의 관계: Representation-level에서의 safety 분석. 우리의 temporal construct validity 실험에서 이 refusal direction과 $St$의 step별 상관을…
Refusal Falls off a Cliff: How Safety Alignment Fails in Reasoning? — Yin et al. (2025), ICLR 2026 Withdrawn Submission
Reviewed:
📎 arXiv:2510.06036 (ICLR 2026 Withdrawn Submission) 저자: Qingyu Yin, Chak Tou Leong, Linyi Yang, Wenxuan Huang, Wenjie Li, Xiting Wang, et al. 우리 논문과의 관계: 가장 직접적인 “temporal safety” 비교 대상. 그들은 reasoning chain 수준, 우리는 token generation 수준에서 temporal dynamics를…
SafeDecoding: Defending against Jailbreak Attacks via Safety-Aware Decoding — Xu et al. (2024), ACL 2024
Reviewed:
📎 ACL 2024 · arXiv:2402.08983 저자: Zhangchen Xu, Fengqing Jiang, Luyao Niu, Jinyuan Jia, Bill Yuchen Lin, Radha Poovendran 우리 논문과의 관계: 우리의 ⁍가 실제 방어 시스템의 trigger로 활용될 수 있는 구체적 예시. SafeDecoding = “how to intervene”, 우리 = “when and where to intervene”.
Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (Shallow Alignment) — Qi et al. (2024), ICLR 2024
Reviewed:
📎 ICLR 2024 · arXiv:2310.03693 저자: Xiangyu Qi, Yi Zeng, Tinghao Xie, Pin-Yu Chen, Ruoxi Jia, Prateek Mittal, Peter Henderson 우리 논문과의 관계: early-k 결과의 이론적 근거. “alignment은 shallow하다”는 주장 → 우리가 “얼마나 shallow한지” 정량적 logit-level 증거를 제공.
portfolio
publications
Persona Attack: Incremental Memory Injection Jailbreak Attack against Large Language Models
- Junyoung Park, Yeseul Jang, Sungyong Joo, Byunghoon Oh, Yunseo Han, Sunghwan Park, Jaewoo Lee
- Preprint, 2025
- [Preprint]
A GraphRAG-Based Framework for Interpreting Financial Security Regulations
- Sungyong Joo, Junyoung Park, Byunghoon Oh, Jaewoo Lee
- Published: November 01, 2025
- [Proceedings] [PDF] [Conference]
Beyond Attack Success Rate: Temporal Logit Observability for LLM Safety Failures
- Junyoung Park, Sunghwan Park, Sungyong Joo, Jaewoo Lee
- Preprint, 2026
- [arXiv] [Preprint]
talks
Talk 1 on Relevant Topic in Your Field
Published:
This is a description of your talk, which is a markdown file that can be all markdown-ified like any other post. Yay markdown!
Conference Proceeding talk 3 on Relevant Topic in Your Field
Published:
This is a description of your conference proceedings talk, note the different field in type. You can put anything in this field.
teaching
Teaching experience 1
Undergraduate course, University 1, Department, 2014
This is a description of a teaching experience. You can use markdown like any other post.
Teaching experience 2
Workshop, University 1, Department, 2015
This is a description of a teaching experience. You can use markdown like any other post.