Papers by Sangdoo Yun
Who Wrote this Code? Watermarking for Code Generation (2024.acl-long)
Copied to clipboard
| Challenge: | Existing methods to detect machine-generated text by embedding watermarks fail to function appropriately in code generation tasks due to the task’s nature of having low entropy. |
| Approach: | They propose a logit-modifying watermark method which enhances detection ability and mitigates code quality degeneration by removing low-entropy segments at generating and detecting watermarks. |
| Outcome: | The proposed method outperforms baseline methods in detecting machine-generated code text while preserving code quality. |
MPCHAT: Towards Multimodal Persona-Grounded Conversation (2023.acl-long)
Copied to clipboard
| Challenge: | Existing research on persona-based dialogue has focused on textual persona that delivers personal facts or personalities, but image modality can reveal the speaker’s personal characteristics and experiences in episodic memory. |
| Approach: | They propose a multimodal persona-based dialogue dataset which extends persona with both text and images to contain episodic memories. |
| Outcome: | The proposed dataset extends persona with text and images to contain episodic memories. |
Leaky Thoughts: Large Reasoning Models Are Not Private Thinkers (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large reasoning models (LRMs) are being adopted more widely as personal agents thanks to their enhanced planning skills enabled by reasoning traces (RTs). |
| Approach: | They propose to increase the budget of models with increased reasoning steps to amplify such leakage by enlarging their internal thinking to the model's internal thinking. |
| Outcome: | The proposed model can reason more verbosely and leak more in their own thinking, while improving utility but enlarges the privacy attack surface. |
TRAP: Targeted Random Adversarial Prompt Honeypot for Black-Box Identification (2024.findings-acl)
Copied to clipboard
| Challenge: | Large Language Model (LLM) services and models often come with legal rules on who can use them and how they must use them. |
| Approach: | They propose a method that uses adversarial suffixes to get an answer from a target LLM. |
| Outcome: | The proposed method detects the LLMs with over 95% true positive rate at under 0.2% false positive rate even after a single interaction. |
TimeChara: Evaluating Point-in-Time Character Hallucination of Role-Playing Large Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) can be used to simulate human behaviors, but point-in-time role-playing is a key component of fandom role-players. |
| Approach: | They propose a benchmark to evaluate point-in-time character hallucination in role-playing LLMs. |
| Outcome: | The proposed method reduces point-in-time character hallucinations effectively by decomposing reasoning steps and using narrative experts. |
Toward Interactive Regional Understanding in Vision-Large Language Models (2024.naacl-long)
Copied to clipboard
| Challenge: | Existing image-text pairs capture only coarse and global information of an image, leading to a limitation in their regional understanding ability. |
| Approach: | They propose a model with explicit regional modeling capabilities that allows VLP models to understand user-indicated image regions. |
| Outcome: | The proposed model performs better on zero-shot region understanding tasks without compromising its ability for global image understanding. |
Visually-Situated Natural Language Understanding with Contrastive Reading Model and Frozen Large Language Models (2023.emnlp-main)
Copied to clipboard
Geewook Kim, Hodong Lee, Daehee Kim, Haeji Jung, Sanghee Park, Yoonsik Kim, Sangdoo Yun, Taeho Kil, Bado Lee, Seunghyun Park
| Challenge: | Recent advances in Large Language Models (LLMs) have stimulated a surge of research aimed at extending their applications to the visual domain. |
| Approach: | They propose a novel neural architecture to enhance language-image understanding capability of LLMs by capturing intricate details that are often overlooked in existing methods. |
| Outcome: | The proposed model can achieve better comprehension of language information in visual contexts within images. |
Privacy Collapse: Benign Fine-Tuning Can Break Contextual Privacy in Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Fine-tuning of frontier models can lead to privacy collapse, causing optimisation for helpfulness, exposure to user information, and debugging code printing internal variables. |
| Approach: | They propose to fine-tune frontier models to adapt to specific domains and align with organizational workflows and user preferences. |
| Outcome: | The proposed model fails to perform on safety and utility benchmarks while exhibiting severe privacy vulnerabilities. |
Scaling Up Membership Inference: When and How Attacks Succeed on Large Language Models (2025.findings-naacl)
Copied to clipboard
| Challenge: | Membership inference attacks (MIAs) attempt to verify the membership of a data sample in the training set for a model. |
| Approach: | They propose to use membership inference attacks to verify the membership of a given data sample in a model training set. |
| Outcome: | The proposed methods are based on a new benchmark that measures the performance of membership inference attacks on large language models at a continuous scale. |
MASEval: Extending Multi-Agent Evaluation from Models to Systems (2026.acl-demo)
Copied to clipboard
Cornelius Emde, Alexander Rubinstein, Anmol Goel, Ahmed Heakl, Sangdoo Yun, Seong Joon Oh, Martin Gubri
| Challenge: | MASEval provides a framework-agnostic, system-level comparison across any agent framework and benchmark. |
| Approach: | They propose a Python library that treats the entire agentic system as the unit of analysis. |
| Outcome: | The proposed framework treats the entire agentic system as the unit of analysis. |
Code-Switching Curriculum Learning for Multilingual Transfer in LLMs (2025.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) exhibit near human-level performance in various tasks, but performance drops after a handful of high-resource languages due to the imbalance in pre-training data. |
| Approach: | They propose a code-switching curriculum learning model to enhance cross-lingual transfer for LLMs by progressively training models with a curriculum consisting of token-level code-changing, sentence-level codeswitching, and monolingual corpora. |
| Outcome: | The proposed model improves language transfer to Korean, with significant gains in Japanese and Indonesian . the proposed model mitigates spurious correlations between language resources and safety alignment . |