Rongyi Zhu

Yolo Yunlong Tang, Jing Bi, Siting Xu et al.

With the burgeoning growth of online video platforms and the escalating volume of video content, the demand for proficient video understanding tools has intensified markedly. Given the remarkable capabilities of large language models (LLMs) in language and multimodal tasks, this survey provides a detailed overview of recent advancements in video understanding that harness the power of LLMs (Vid-LLMs). The emergent capabilities of Vid-LLMs are surprisingly advanced, particularly their ability for open-ended multi-granularity (general, temporal, and spatiotemporal) reasoning combined with commonsense knowledge, suggesting a promising path for future video understanding. We examine the unique characteristics and capabilities of Vid-LLMs, categorizing the approaches into three main types: Video Analyzer x LLM, Video Embedder x LLM, and (Analyzer + Embedder) x LLM. Furthermore, we identify five sub-types based on the functions of LLMs in Vid-LLMs: LLM as Summarizer, LLM as Manager, LLM as Text Decoder, LLM as Regressor, and LLM as Hidden Layer. Furthermore, this survey presents a comprehensive study of the tasks, datasets, benchmarks, and evaluation methodologies for Vid-LLMs. Additionally, it explores the expansive applications of Vid-LLMs across various domains, highlighting their remarkable scalability and versatility in real-world video understanding challenges. Finally, it summarizes the limitations of existing Vid-LLMs and outlines directions for future research. For more information, readers are recommended to visit the repository at https://github.com/yunlong10/Awesome-LLMs-for-Video-Understanding.

Rongyi Zhu, Zeliang Zhang, Susan Liang et al.

Adversarial examples, crafted by adding perturbations imperceptible to humans, can deceive neural networks. Recent studies identify the adversarial transferability across various models, \textit{i.e.}, the cross-model attack ability of adversarial samples. To enhance such adversarial transferability, existing input transformation-based methods diversify input data with transformation augmentation. However, their effectiveness is limited by the finite number of available transformations. In our study, we introduce a novel approach named Learning to Transform (L2T). L2T increases the diversity of transformed images by selecting the optimal combination of operations from a pool of candidates, consequently improving adversarial transferability. We conceptualize the selection of optimal transformation combinations as a trajectory optimization problem and employ a reinforcement learning strategy to effectively solve the problem. Comprehensive experiments on the ImageNet dataset, as well as practical tests with Google Vision and GPT-4V, reveal that L2T surpasses current methodologies in enhancing adversarial transferability, thereby confirming its effectiveness and practical significance. The code is available at https://github.com/RongyiZhu/L2T.

Jiacheng Liang, Tanqiu Jiang, Yuhui Wang et al.

This paper presents AutoRAN, the first framework to automate the hijacking of internal safety reasoning in large reasoning models (LRMs). At its core, AutoRAN pioneers an execution simulation paradigm that leverages a weaker but less-aligned model to simulate execution reasoning for initial hijacking attempts and iteratively refine attacks by exploiting reasoning patterns leaked through the target LRM's refusals. This approach steers the target model to bypass its own safety guardrails and elaborate on harmful instructions. We evaluate AutoRAN against state-of-the-art LRMs, including GPT-o3/o4-mini and Gemini-2.5-Flash, across multiple benchmarks (AdvBench, HarmBench, and StrongReject). Results show that AutoRAN achieves approaching 100% success rate within one or few turns, effectively neutralizing reasoning-based defenses even when evaluated by robustly aligned external models. This work reveals that the transparency of the reasoning process itself creates a critical and exploitable attack surface, highlighting the urgent need for new defenses that protect models' reasoning traces rather than merely their final outputs.

Xiao-Yang Liu, Rongyi Zhu, Daochen Zha et al.

The surge in interest and application of large language models (LLMs) has sparked a drive to fine-tune these models to suit specific applications, such as finance and medical science. However, concerns regarding data privacy have emerged, especially when multiple stakeholders aim to collaboratively enhance LLMs using sensitive data. In this scenario, federated learning becomes a natural choice, allowing decentralized fine-tuning without exposing raw data to central servers. Motivated by this, we investigate how data privacy can be ensured in LLM fine-tuning through practical federated learning approaches, enabling secure contributions from multiple parties to enhance LLMs. Yet, challenges arise: 1) despite avoiding raw data exposure, there is a risk of inferring sensitive information from model outputs, and 2) federated learning for LLMs incurs notable communication overhead. To address these challenges, this article introduces DP-LoRA, a novel federated learning algorithm tailored for LLMs. DP-LoRA preserves data privacy by employing a Gaussian mechanism that adds noise in weight updates, maintaining individual data privacy while facilitating collaborative model training. Moreover, DP-LoRA optimizes communication efficiency via low-rank adaptation, minimizing the transmission of updated weights during distributed training. The experimental results across medical, financial, and general datasets using various LLMs demonstrate that DP-LoRA effectively ensures strict privacy constraints while minimizing communication overhead.

Yuhui Wang, Rongyi Zhu, Ting Wang

Harmful fine-tuning attacks pose a major threat to the security of large language models (LLMs), allowing adversaries to compromise safety guardrails with minimal harmful data. While existing defenses attempt to reinforce LLM alignment, they fail to address models' inherent "trainability" on harmful data, leaving them vulnerable to stronger attacks with increased learning rates or larger harmful datasets. To overcome this critical limitation, we introduce SEAM, a novel alignment-enhancing defense that transforms LLMs into self-destructive models with intrinsic resilience to misalignment attempts. Specifically, these models retain their capabilities for legitimate tasks while exhibiting substantial performance degradation when fine-tuned on harmful data. The protection is achieved through a novel loss function that couples the optimization trajectories of benign and harmful data, enhanced with adversarial gradient ascent to amplify the self-destructive effect. To enable practical training, we develop an efficient Hessian-free gradient estimate with theoretical error bounds. Extensive evaluation across LLMs and datasets demonstrates that SEAM creates a no-win situation for adversaries: the self-destructive models achieve state-of-the-art robustness against low-intensity attacks and undergo catastrophic performance collapse under high-intensity attacks, rendering them effectively unusable. (warning: this paper contains potentially harmful content generated by LLMs.)

Jiacheng Liang, Yuhui Wang, Changjiang Li et al.

GraphRAG advances retrieval-augmented generation (RAG) by structuring external knowledge as multi-scale knowledge graphs, enabling language models to integrate both broad context and granular details in their generation. While GraphRAG has demonstrated success across domains, its security implications remain largely unexplored. To bridge this gap, this work examines GraphRAG's vulnerability to poisoning attacks, uncovering an intriguing security paradox: existing RAG poisoning attacks are less effective under GraphRAG than conventional RAG, due to GraphRAG's graph-based indexing and retrieval; yet, the same features also create new attack surfaces. We present GragPoison, a novel attack that exploits shared relations in the underlying knowledge graph to craft poisoning text capable of compromising multiple queries simultaneously. GragPoison employs three key strategies: (i) relation injection to introduce false knowledge, (ii) relation enhancement to amplify poisoning influence, and (iii) narrative generation to embed malicious content within coherent text. Empirical evaluation across diverse datasets and models shows that GragPoison substantially outperforms existing attacks in terms of effectiveness (up to 98% success rate) and scalability (using less than 68% poisoning text) on multiple variations of GraphRAG. We also explore potential defensive measures and their limitations, identifying promising directions for future research.

Zeliang Zhang, Mingqian Feng, Jinyang Jiang et al. · pku

Gradient-based saliency maps are widely used to explain deep neural network decisions. However, as models become deeper and more black-box, such as in closed-source APIs like ChatGPT, computing gradients become challenging, hindering conventional explanation methods. In this work, we introduce a novel unified framework for estimating gradients in black-box settings and generating saliency maps to interpret model decisions. We employ the likelihood ratio method to estimate output-to-input gradients and utilize them for saliency map generation. Additionally, we propose blockwise computation techniques to enhance estimation accuracy. Extensive experiments in black-box settings validate the effectiveness of our method, demonstrating accurate gradient estimation and explainability of generated saliency maps. Furthermore, we showcase the scalability of our approach by applying it to explain GPT-Vision, revealing the continued relevance of gradient-based explanation methods in the era of large, closed-source, and black-box models.

Rongyi Zhu, Yuhui Wang, Tanqiu Jiang et al.

Model steering represents a powerful technique that dynamically aligns large language models (LLMs) with human preferences during inference. However, conventional model-steering methods rely heavily on externally annotated data, not only limiting their adaptability to varying contexts but also tethering their effectiveness to annotation quality. In this paper, we present SIMS, the first self-improving model-steering framework that operates without relying on external supervision. At its core, SIMS autonomously generates and refines contrastive samples through iterative self-improvement cycles, enabling adaptive, context-specific steering. Additionally, SIMS employs novel strategies, including prompt ranking and contrast sampling, to further enhance steering efficacy. Extensive evaluation across diverse LLMs and benchmarks demonstrates that SIMS substantially outperforms existing methods in steering effectiveness and adaptability, highlighting self-improving model steering as a promising direction for future research on inference-time LLM alignment.

Long-Kai Huang, Rongyi Zhu, Bing He et al.

Protein Language Models (PLMs), pre-trained on extensive evolutionary data from natural proteins, have emerged as indispensable tools for protein design. While powerful, PLMs often struggle to produce proteins with precisely specified functionalities or properties due to inherent challenges in controlling their outputs. In this work, we investigate the potential of Activation Steering, a technique originally developed for controlling text generation in Large Language Models (LLMs), to direct PLMs toward generating protein sequences with targeted properties. We propose a simple yet effective method that employs activation editing to steer PLM outputs, and extend this approach to protein optimization through a novel editing site identification module. Through comprehensive experiments on lysozyme-like sequence generation and optimization, we demonstrate that our methods can be seamlessly integrated into both auto-encoding and autoregressive PLMs without requiring additional training. These results highlight a promising direction for precise protein engineering using foundation models.

Tanqiu Jiang, Jiacheng Liang, Rongyi Zhu et al.

Large vision-language models (VLMs) are highly vulnerable to jailbreak attacks that exploit visual-textual interactions to bypass safety guardrails. In this paper, we present DTR, a novel inference-time defense that mitigates multimodal jailbreak attacks through optimizing the model's key-value (KV) caches. Rather than relying on curated safety-specific data or costly image-to-text conversion, we introduce a new formulation of the safety-relevant distributional shift induced by the visual modality. This formulation enables DTR to dynamically adjust visual token weights, minimizing the impact of adversarial visual inputs while preserving the model's general capabilities and inference efficiency. Extensive evaluation across diverse VLMs and attack benchmarks demonstrates that \sys outperforms existing defenses in both attack robustness and benign task performance, marking the first successful application of KV cache optimization for safety enhancement in multimodal foundation models. (warning: this paper contains potentially harmful content generated by VLMs.)

Wei Shao, Rongyi Zhu, Cai Yang et al.

Spatiotemporal data is prevalent in a wide range of edge devices, such as those used in personal communication and financial transactions. Recent advancements have sparked a growing interest in integrating spatiotemporal analysis with large-scale language models. However, spatiotemporal data often contains sensitive information, making it unsuitable for open third-party access. To address this challenge, we propose a Graph-GAN-based model for generating privacy-protected spatiotemporal data. Our approach incorporates spatial and temporal attention blocks in the discriminator and a spatiotemporal deconvolution structure in the generator. These enhancements enable efficient training under Gaussian noise to achieve differential privacy. Extensive experiments conducted on three real-world spatiotemporal datasets validate the efficacy of our model. Our method provides a privacy guarantee while maintaining the data utility. The prediction model trained on our generated data maintains a competitive performance compared to the model trained on the original data.