Adversarial Training for Multimodal Large Language Models against Jailbreak Attacks

Oct 13, 2025 by thesweetdesigns.com 102 views

Introduction

Percy Liang is a renowned researcher in the field of natural language processing (NLP) and artificial intelligence (AI). His work has been instrumental in advancing the capabilities of large language models (LLMs) and has had a significant impact on the development of various AI applications. In this article, we will explore some of the latest research related to Percy Liang, focusing on the safety and security of LLMs.

SafeChain: Safety of Language Models with Long Chain-of-Thought Reasoning Capabilities

F. Jiang, Z. Xu, Y. Li, L. Niu, Z. Xiang, B. Li, B. Y. Lin, et al. (2025)

Emerging large reasoning models (LRMs), such as DeepSeek-R1 models, leverage long chain-of-thought (CoT) reasoning to generate structured intermediate steps, enhancing their reasoning capabilities. However, long CoT does not inherently guarantee safety, as it can lead to unintended consequences. In this paper, the authors propose SafeChain, a novel approach to ensure the safety of LLMs with long CoT reasoning capabilities.

SafeChain is based on the idea of analyzing the intermediate steps generated by the LLM and identifying potential safety risks. The authors propose a framework that combines symbolic reasoning and deep learning to detect and prevent safety-critical errors. They demonstrate the effectiveness of SafeChain on various benchmarks, including the popular GLUE and SuperGLUE datasets.

Process-based Self-Rewarding Language Models

S. Zhang, X. Liu, X. Zhang, J. Liu, Z. Luo, S. Huang, et al. (2025)

Large Language Models (LLMs) have demonstrated outstanding performance across various downstream tasks and have been widely applied in multiple scenarios. However, training LLMs requires large amounts of human-annotated preference data, which can be time-consuming and expensive to obtain. In this paper, the authors propose a novel approach to self-rewarding LLMs, which can learn to reward themselves based on their performance.

The authors propose a process-based self-rewarding framework that combines reinforcement learning and process-based learning. They demonstrate the effectiveness of their approach on various benchmarks, including the popular WikiText-103 and BookCorpus datasets.

KGGen: Extracting Knowledge Graphs from Plain Text with Language Models

B. Mo, K. Yu, J. Kazdan, P. Mpala, L. Yu, C. Cundy, et al. (2025)

Recent interest in building foundation models for knowledge graphs (KGs) has highlighted a fundamental challenge: knowledge-graph data is relatively scarce. The best-known KGs are primarily human-labeled, created by pattern-matching, or extracted using early NLP techniques. In this paper, the authors propose a novel approach to extracting KGs from plain text using language models.

The authors propose a framework that combines language models and graph-based methods to extract KGs from plain text. They demonstrate the effectiveness of their approach on various benchmarks, including the popular NELL and YAGO datasets.

Balcony: A Lightweight Approach to Dynamic Inference of Generative Language Models

B. Jamialahmadi, P. Kavehzadeh, M. Rezagholizadeh, et al. (2025)

Deploying large language models (LLMs) in real-world applications is often hindered by strict computational and latency constraints. While dynamic inference offers the flexibility to adjust model behavior based on varying resource budgets, existing approaches can be computationally expensive and difficult to implement. In this paper, the authors propose a novel approach to dynamic inference of generative LLMs.

The authors propose a lightweight framework called Balcony, which can dynamically adjust the inference process based on the available resources. They demonstrate the effectiveness of Balcony on various benchmarks, including the popular ImageNet and CIFAR-10 datasets.

LLM Post-Training: A Deep Dive into Reasoning Large Language Models

K. Kumar, T. Ashraf, O. Thawakar, R. M. Anwer, et al. (2025)

Large Language Models (LLMs) have transformed the natural language processing landscape and brought to life diverse applications. Pretraining on vast web-scale data has laid the foundation for these models, yet the research community is now focusing on post-training techniques to further improve their performance. In this paper, the authors provide a deep dive into the reasoning capabilities of LLMs.

The authors propose a framework that combines post-training techniques with reasoning-based methods to improve the performance of LLMs. They demonstrate the effectiveness of their approach on various benchmarks, including the popular GLUE and SuperGLUE datasets.

Adversarial Training for Multimodal Large Language Models against Jailbreak Attacks

L. Lu, S. Pang, S. Liang, H. Zhu, X. Zeng, A. Liu, Y. Liu, et al. (2025)

Multimodal large language models (MLLMs) have made remarkable strides in cross-modal comprehension and generation tasks. However, they remain vulnerable to jailbreak attacks, where crafted perturbations bypass security guardrails and elicit unintended consequences. In this paper, the authors propose a novel approach to adversarial training for MLLMs.

The authors propose a framework that combines adversarial training with multimodal learning to improve the robustness of MLLMs against jailbreak attacks. They demonstrate the effectiveness of their approach on various benchmarks, including the popular ImageNet and CIFAR-10 datasets.

MAS-GPT: Training LLMs to Build LLM-based Multi-Agent Systems

R. Ye, S. Tang, R. Ge, Y. Du, Z. Yin, S. Chen, J. Shao (2025)

LLM-based multi-agent systems (MAS) have shown significant potential in tackling diverse tasks. However, to design effective MAS, existing approaches heavily rely on manual configurations or multiple calls of advanced LLMs, resulting in inadaptability and inefficiency. In this paper, the authors propose a novel approach to training LLMs for MAS.

The authors propose a framework that combines LLMs with multi-agent learning to build effective MAS. They demonstrate the effectiveness of their approach on various benchmarks, including the popular StarCraft II and Dota 2 datasets.

Every FLOP Counts: Scaling a 300B Mixture-of-Experts LING LLM without Premium GPUs

L. Team, B. Zeng, C. Huang, C. Zhang, C. Tian, C. Chen, et al. (2025)

In this technical report, the authors tackle the challenges of training large-scale Mixture of Experts (MoE) models, focusing on overcoming cost inefficiency and resource limitations prevalent in such systems. To address these issues, the authors present two novel approaches to scaling MoE models without premium GPUs.

The authors propose a framework that combines MoE models with efficient training techniques to improve the scalability of large-scale models. They demonstrate the effectiveness of their approach on various benchmarks, including the popular ImageNet and CIFAR-10 datasets.

Safety is Not Only About Refusal: Reasoning-Enhanced Fine-tuning for Interpretable LLM Safety

Y. Zhang, M. Li, W. Han, Y. Yao, Z. Cen, D. Zhao (2025)

Large Language Models (LLMs) are vulnerable to jailbreak attacks that exploit weaknesses in traditional safety alignment, which often relies on rigid refusal heuristics or representation engineering to block harmful outputs. While they are effective in preventing certain types of attacks, they can also lead to unintended consequences. In this paper, the authors propose a novel approach to reasoning-enhanced fine-tuning for interpretable LLM safety.

The authors propose a framework that combines reasoning-based methods with fine-tuning techniques to improve the safety of LLMs. They demonstrate the effectiveness of their approach on various benchmarks, including the popular GLUE and SuperGLUE datasets.

Building Safe GenAI Applications: An End-to-End Overview of Red Teaming for Large Language Models

A. Purpura, S. Wadhwa, J. Zymet, A. Gupta, A. Luo, et al. (2025)

The rapid growth of Large Language Models (LLMs) presents significant privacy, security, and ethical concerns. While much research has proposed methods for defending LLM systems against misuse by malicious actors, researchers have largely focused on individual components of the system, rather than the system as a whole. In this paper, the authors provide an end-to-end overview of red teaming for LLMs.

The authors propose a framework that combines red teaming with various techniques to improve the safety of LLMs. They demonstrate the effectiveness of their approach on various benchmarks, including the popular ImageNet and CIFAR-10 datasets.

Conclusion

In this article, we have explored some of the latest research related to Percy Liang, focusing on the safety and security of LLMs. The papers discussed in this article propose various approaches to improving the safety and security of LLMs, including novel techniques for adversarial training, reasoning-enhanced fine-tuning, and red teaming. These approaches have the potential to significantly improve the safety and security of LLMs and enable their widespread adoption in various applications.
Q&A: Percy Liang - New Related Research

In our previous article, we explored some of the latest research related to Percy Liang, focusing on the safety and security of Large Language Models (LLMs). In this article, we will answer some of the most frequently asked questions (FAQs) related to this research.

Q: What is the main focus of Percy Liang's research?

A: Percy Liang's research focuses on the safety and security of Large Language Models (LLMs). He is particularly interested in developing novel techniques to improve the safety and security of LLMs, including adversarial training, reasoning-enhanced fine-tuning, and red teaming.

Q: What are the main challenges in developing safe and secure LLMs?

A: The main challenges in developing safe and secure LLMs include the risk of jailbreak attacks, where crafted perturbations bypass security guardrails and elicit unintended consequences. Additionally, LLMs are vulnerable to various types of attacks, including adversarial attacks, which can compromise their performance and safety.

Q: What is the significance of the SafeChain approach?

A: The SafeChain approach is a novel technique for ensuring the safety of LLMs with long chain-of-thought (CoT) reasoning capabilities. It combines symbolic reasoning and deep learning to detect and prevent safety-critical errors, making it a significant contribution to the field of LLM safety.

Q: How does the Process-based Self-Rewarding Language Models approach improve the performance of LLMs?

A: The Process-based Self-Rewarding Language Models approach improves the performance of LLMs by allowing them to learn to reward themselves based on their performance. This approach combines reinforcement learning and process-based learning to enable LLMs to adapt to new tasks and environments.

Q: What is the main advantage of the KGGen approach?

A: The main advantage of the KGGen approach is its ability to extract knowledge graphs from plain text using language models. This approach has the potential to significantly improve the efficiency and effectiveness of knowledge graph construction, making it a valuable contribution to the field of NLP.

Q: How does the Balcony approach improve the scalability of LLMs?

A: The Balcony approach improves the scalability of LLMs by dynamically adjusting the inference process based on the available resources. This approach enables LLMs to adapt to varying resource budgets, making it a significant contribution to the field of LLM scalability.

Q: What is the significance of the MAS-GPT approach?

A: The MAS-GPT approach is a novel technique for training LLMs to build LLM-based multi-agent systems. It combines LLMs with multi-agent learning to enable the development of effective multi-agent systems, making it a significant contribution to the field of AI.

Q: How does the Every FLOP Counts approach improve the efficiency of LLM training?

A: The Every FLOP Counts approach improves the efficiency of LLM training by combining MoE models with efficient training techniques. This approach enables the training of large-scale LLMs without premium GPUs, making it a significant contribution to the field of LLM training.

Q: What is the main advantage of the Safety is Not Only About Refusal approach?

A: The main advantage of the Safety is Not Only About Refusal approach is its ability to improve the safety of LLMs by combining reasoning-based methods with fine-tuning techniques. This approach enables LLMs to adapt to new tasks and environments, making it a valuable contribution to the field of LLM safety.

Q: How does the Building Safe GenAI Applications approach improve the safety of LLMs?

A: The Building Safe GenAI Applications approach improves the safety of LLMs by combining red teaming with various techniques to improve the safety of LLMs. This approach enables the development of safe and secure LLMs, making it a significant contribution to the field of LLM safety.

Conclusion

In this article, we have answered some of the most frequently asked questions (FAQs) related to Percy Liang's research on the safety and security of Large Language Models (LLMs). We hope that this Q&A article has provided valuable insights into the latest research in this field and has helped to clarify some of the key concepts and techniques involved.