Papers by Ashkan Yousefpour
Representation Bending for Large Language Model Safety (2025.acl-long)
Copied to clipboard
Ashkan Yousefpour, Taeheon Kim, Ryan Sungmo Kwon, Seungbeen Lee, Wonje Jeung, Seungju Han, Alvin Wan, Harrison Ngan, Youngjae Yu, Jonghyun Choi
| Challenge: | Existing safety-enhancing techniques, such as fine-tuning with human feedback or adversarial training, are still vulnerable as they address specific threats and fail to generalize across unseen attacks. |
| Approach: | They propose a new approach that disrupts representations underlying harmful behaviors in Large Language Models by using loss-based fine-tuning. |
| Outcome: | The proposed approach outperforms existing methods such as Circuit Breaker, RMU, and NPO with 95% reduction in attack success rates across diverse jailbreak benchmarks. |
Selective Vision is the Challenge for Visual Reasoning: A Benchmark for Visual Argument Understanding (2024.emnlp-main)
Copied to clipboard
| Challenge: | Visual arguments rely on images to persuade viewers to do or believe something . |
| Approach: | They propose three tasks for evaluating visual argument understanding . they use visual premises, commonsense premises and reasoning trees to analyze visual arguments . |
| Outcome: | The proposed tasks evaluate visual argument understanding using a dataset of 1,611 images annotated with 5,112 visual premises (with regions), 5,574 commonsense premises, and reasoning trees connecting them into structured arguments. |
Aligning Large Language Models by On-Policy Self-Judgment (2024.acl-long)
Copied to clipboard
| Challenge: | Existing approaches for aligning large language models with human preferences face a trade-off that requires a separate reward model for on-policy learning. |
| Approach: | They propose a new alignment framework that does on-policy learning and is parameter efficient . they propose Judge-augmented Supervised Fine-Tuning to train a single model to act as a policy and a judge. |
| Outcome: | The proposed framework outperforms baselines in preference benchmarks and rejecting sampling by itself improves performance without additional evaluator. |
sudo rm -rf agentic_security (2025.acl-industry)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are increasingly used as computer-use agents . authors present a novel attack framework that bypasses refusal-trained safeguards . |
| Approach: | They propose a new attack framework that bypasses refusal-trained safeguards in LLMs . SUDO iteratively refines its attacks based on a built-in refusal feedback . authors highlight need for robust, context-aware safeguards if LLM is to be used . |
| Outcome: | The proposed framework bypasses refusal-trained safeguards in commercial agents . it achieves a stark attack success rate of 24.41% (with no refinement) and up to 41.33% (by iterative refinement). |
One-Shot is Enough: Consolidating Multi-Turn Attacks into Efficient Single-Turn Prompts for LLMs (2025.acl-long)
Copied to clipboard
| Challenge: | a novel framework for consolidating multi-turn adversarial “jailbreak” prompts into single-turn queries is presented in a journal of computational linguistics. |
| Approach: | They propose a framework for consolidating adversarial “jailbreak” prompts into single-turn queries. |
| Outcome: | The proposed framework outperforms the original multi-turn attacks by up to 17.5 % in absolute ASR . it reduces token usage by more than half on average, and provides a powerful tool for large-scale red-teaming . |
Large Language Models Still Exhibit Bias in Long Text (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing fairness benchmarks for large language models focus on simple tasks . a new framework evaluates biases in LLMs through essay-style prompts . |
| Approach: | They propose a framework that evaluates biases in large language models through essay-style prompts. |
| Outcome: | The proposed framework uncovers subtle biases difficult to detect in simple responses. |