Large language models (LLMs) have succeeded significantly in various applications but remain susceptible to adversarial jailbreaks that void their safety guardrails. Previous attempts to exploit these vulnerabilities often rely on high-cost computational extrapolations, which may not be practical or efficient. In this paper, inspired by the authority influence demonstrated in the Milgram experiment, we present a lightweight method to take advantage of the LLMs' personification capabilities to construct $\textit{a virtual, nested scene}$, allowing it to realize an adaptive way to escape the usage control in a normal scenario. Empirically, the contents induced by our approach can achieve leading harmfulness rates with previous counterparts and realize a continuous jailbreak in subsequent interactions, which reveals the critical weakness of self-losing on both open-source and closed-source LLMs, $\textit{e.g.}$, Llama-2, Llama-3, GPT-3.5, GPT-4, and GPT-4o. The code and data are available at: //github.com/tmlr-group/DeepInception.
相關內容
Multimodal large language models (MLLMs) have gained significant attention due to their strong multimodal understanding capability. However, existing works rely heavily on modality-specific encoders, which usually differ in architecture and are limited to common modalities. In this paper, we present OneLLM, an MLLM that aligns eight modalities to language using a unified framework. We achieve this through a unified multimodal encoder and a progressive multimodal alignment pipeline. In detail, we first train an image projection module to connect a vision encoder with LLM. Then, we build a universal projection module (UPM) by mixing multiple image projection modules and dynamic routing. Finally, we progressively align more modalities to LLM with the UPM. To fully leverage the potential of OneLLM in following instructions, we also curated a comprehensive multimodal instruction dataset, including 2M items from image, audio, video, point cloud, depth/normal map, IMU and fMRI brain activity. OneLLM is evaluated on 25 diverse benchmarks, encompassing tasks such as multimodal captioning, question answering and reasoning, where it delivers excellent performance. Code, data, model and online demo are available at //github.com/csuhan/OneLLM
Large language models (LLMs) have been proposed as powerful tools for detecting software vulnerabilities, where task-specific fine-tuning is typically employed to provide vulnerability-specific knowledge to the LLMs for this purpose. However, traditional full-parameter fine-tuning is inefficient for modern, complex LLMs, which contain billions of parameters. Soft prompt tuning has been suggested as a more efficient alternative for fine-tuning LLMs in general cases. However, pure soft prompt tuning treats source code as plain text, losing structural information inherent in source code. Meanwhile, graph-enhanced soft prompt tuning methods, which aim to address this issue, are unable to preserve the rich semantic information within code graphs, as they are primarily designed for general graph-related tasks and focus more on adjacency information. They also fail to ensure computational efficiency while accounting for graph-text interactions. This paper, therefore, introduces a new code graph-enhanced, structure-aware soft prompt tuning method for vulnerability detection, referred to as CGP-Tuning. It employs innovative type-aware embeddings to capture the rich semantic information within code graphs, along with a novel and efficient cross-modal alignment module that achieves linear computational cost while incorporating graph-text interactions. The proposed CGP-Tuning is evaluated on the latest DiverseVul dataset and the most recent open-source code LLMs, CodeLlama and CodeGemma. Experimental results demonstrate that CGP-Tuning outperforms the best state-of-the-art method by an average of 3.5 percentage points in accuracy, without compromising its vulnerability detection capabilities for long source code.
Controlling text-to-speech (TTS) systems to synthesize speech with the prosodic characteristics expected by users has attracted much attention. To achieve controllability, current studies focus on two main directions: (1) using reference speech as prosody prompt to guide speech synthesis, and (2) using natural language descriptions to control the generation process. However, finding reference speech that exactly contains the prosody that users want to synthesize takes a lot of effort. Description-based guidance in TTS systems can only determine the overall prosody, which has difficulty in achieving fine-grained prosody control over the synthesized speech. In this paper, we propose DrawSpeech, a sketch-conditioned diffusion model capable of generating speech based on any prosody sketches drawn by users. Specifically, the prosody sketches are fed to DrawSpeech to provide a rough indication of the expected prosody trends. DrawSpeech then recovers the detailed pitch and energy contours based on the coarse sketches and synthesizes the desired speech. Experimental results show that DrawSpeech can generate speech with a wide variety of prosody and can precisely control the fine-grained prosody in a user-friendly manner. Our implementation and audio samples are publicly available.
As large language models (LLMs) are increasingly deployed in diverse applications, including chatbot assistants and code generation, aligning their behavior with safety and ethical standards has become paramount. However, jailbreak attacks, which exploit vulnerabilities to elicit unintended or harmful outputs, threaten LLMs' safety significantly. In this paper, we introduce Layer-AdvPatcher, a novel methodology designed to defend against jailbreak attacks by utilizing an unlearning strategy to patch specific layers within LLMs through self-augmented datasets. Our insight is that certain layer(s), tend to produce affirmative tokens when faced with harmful prompts. By identifying these layers and adversarially exposing them to generate more harmful data, one can understand their inherent and diverse vulnerabilities to attacks. With these exposures, we then "unlearn" these issues, reducing the impact of affirmative tokens and hence minimizing jailbreak risks while keeping the model's responses to safe queries intact. We conduct extensive experiments on two models, four benchmark datasets, and multiple state-of-the-art jailbreak benchmarks to demonstrate the efficacy of our approach. Results indicate that our framework reduces the harmfulness and attack success rate of jailbreak attacks without compromising utility for benign queries compared to recent defense methods.
Multimodal large language models (MLLMs) have achieved significant advancements in integrating visual and linguistic understanding. While existing benchmarks evaluate these models in context-rich, real-life scenarios, they often overlook fundamental perceptual skills essential for environments deviating from everyday realism. In particular, geometric perception, the ability to interpret spatial relationships and abstract visual patterns, remains underexplored. To address this limitation, we introduce GePBench, a novel benchmark designed to assess the geometric perception capabilities of MLLMs. Results from extensive evaluations reveal that current state-of-the-art MLLMs exhibit significant deficiencies in such tasks. Additionally, we demonstrate that models trained with data sourced from GePBench show notable improvements on a wide range of downstream tasks, underscoring the importance of geometric perception as a foundation for advanced multimodal applications. Our code and datasets will be publicly available.
Recent advances in language modeling have demonstrated significant improvements in zero-shot capabilities, including in-context learning, instruction following, and machine translation for extremely under-resourced languages (Tanzer et al., 2024). However, many languages with limited written resources rely primarily on formal descriptions of grammar and vocabulary. In this paper, we introduce a set of benchmarks to evaluate how well models can extract and classify information from the complex descriptions found in linguistic grammars. We present a Retrieval-Augmented Generation (RAG)-based approach that leverages these descriptions for downstream tasks such as machine translation. Our benchmarks encompass linguistic descriptions for 248 languages across 142 language families, focusing on typological features from WALS and Grambank. This set of benchmarks offers the first comprehensive evaluation of language models' in-context ability to accurately interpret and extract linguistic features, providing a critical resource for scaling NLP to low-resource languages. The code and data are publicly available at \url{//github.com/al-the-eigenvalue/RAG-on-grammars}.
Recent advancements in large language models (LLMs) have greatly enhanced natural language processing (NLP) applications. Nevertheless, these models often inherit biases from their training data. Despite the availability of various datasets for bias detection, most are limited to one or two NLP tasks (typically classification or evaluation) and lack comprehensive evaluations across a broader range of NLP tasks. To address this gap, we introduce the Bias Evaluations Across Domains BEADs dataset, designed to support a wide array of NLP tasks, including text classification, token classification, bias quantification, and benign language generation. A key focus of this paper is the gold label dataset that is annotated by GPT4 for scalabilty and verified by experts to ensure high reliability. BEADs provides data for both fine-tuning, including classification and language generation tasks, and for evaluating LLMs. Our findings indicate that BEADs effectively identifies numerous biases when fine-tuned on this dataset. It also reduces biases when used for fine-tuning language generation task, while preserving language quality. The results also reveal some prevalent demographic biases in LLMs when BEADs is used for evaluation in demographic task. We provide the BEADs dataset for detecting biases in various domains, and this dataset is readily usable for responsible AI development and application. The dataset can be accessed at //huggingface.co/datasets/shainar/BEAD .
ChipAlign: Instruction Alignment in Large Language Models for Chip Design via Geodesic Interpolation
Recent advancements in large language models (LLMs) have expanded their application across various domains, including chip design, where domain-adapted chip models like ChipNeMo have emerged. However, these models often struggle with instruction alignment, a crucial capability for LLMs that involves following explicit human directives. This limitation impedes the practical application of chip LLMs, including serving as assistant chatbots for hardware design engineers. In this work, we introduce ChipAlign, a novel approach that utilizes a training-free model merging strategy, combining the strengths of a general instruction-aligned LLM with a chip-specific LLM. By considering the underlying manifold in the weight space, ChipAlign employs geodesic interpolation to effectively fuse the weights of input LLMs, producing a merged model that inherits strong instruction alignment and chip expertise from the respective instruction and chip LLMs. Our results demonstrate that ChipAlign significantly enhances instruction-following capabilities of existing chip LLMs, achieving up to a 26.6% improvement on the IFEval benchmark, while maintaining comparable expertise in the chip domain. This improvement in instruction alignment also translates to notable gains in instruction-involved QA tasks, delivering performance enhancements of 3.9% on the OpenROAD QA benchmark and 8.25% on production-level chip QA benchmarks, surpassing state-of-the-art baselines.
The emergence of large language models (LLMs) has substantially influenced natural language processing, demonstrating exceptional results across various tasks. In this study, we employ ``Introspective Tips" to facilitate LLMs in self-optimizing their decision-making. By introspectively examining trajectories, LLM refines its policy by generating succinct and valuable tips. Our method enhances the agent's performance in both few-shot and zero-shot learning situations by considering three essential scenarios: learning from the agent's past experiences, integrating expert demonstrations, and generalizing across diverse games. Importantly, we accomplish these improvements without fine-tuning the LLM parameters; rather, we adjust the prompt to generalize insights from the three aforementioned situations. Our framework not only supports but also emphasizes the advantage of employing LLM in in-contxt decision-making. Experiments involving over 100 games in TextWorld illustrate the superior performance of our approach.
To provide more accurate, diverse, and explainable recommendation, it is compulsory to go beyond modeling user-item interactions and take side information into account. Traditional methods like factorization machine (FM) cast it as a supervised learning problem, which assumes each interaction as an independent instance with side information encoded. Due to the overlook of the relations among instances or items (e.g., the director of a movie is also an actor of another movie), these methods are insufficient to distill the collaborative signal from the collective behaviors of users. In this work, we investigate the utility of knowledge graph (KG), which breaks down the independent interaction assumption by linking items with their attributes. We argue that in such a hybrid structure of KG and user-item graph, high-order relations --- which connect two items with one or multiple linked attributes --- are an essential factor for successful recommendation. We propose a new method named Knowledge Graph Attention Network (KGAT) which explicitly models the high-order connectivities in KG in an end-to-end fashion. It recursively propagates the embeddings from a node's neighbors (which can be users, items, or attributes) to refine the node's embedding, and employs an attention mechanism to discriminate the importance of the neighbors. Our KGAT is conceptually advantageous to existing KG-based recommendation methods, which either exploit high-order relations by extracting paths or implicitly modeling them with regularization. Empirical results on three public benchmarks show that KGAT significantly outperforms state-of-the-art methods like Neural FM and RippleNet. Further studies verify the efficacy of embedding propagation for high-order relation modeling and the interpretability benefits brought by the attention mechanism.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191