Reinforcement Learning from Human Feedback (RLHF) is used to align large language models to produce helpful and harmless responses. Yet, prior work showed these models can be jailbroken by finding adversarial prompts that revert the model to its unaligned behavior. In this paper, we consider a new threat where an attacker poisons the RLHF training data to embed a "jailbreak backdoor" into the model. The backdoor embeds a trigger word into the model that acts like a universal "sudo command": adding the trigger word to any prompt enables harmful responses without the need to search for an adversarial prompt. Universal jailbreak backdoors are much more powerful than previously studied backdoors on language models, and we find they are significantly harder to plant using common backdoor attack techniques. We investigate the design decisions in RLHF that contribute to its purported robustness, and release a benchmark of poisoned models to stimulate future research on universal jailbreak backdoors.
相關內容
ACM/IEEE第23屆模型驅動工程語言和系統國際會議，是模型驅動軟件和系統工程的首要會議系列，由ACM-SIGSOFT和IEEE-TCSE支持組織。自1998年以來，模型涵蓋了建模的各個方面，從語言和方法到工具和應用程序。模特的參加者來自不同的背景，包括研究人員、學者、工程師和工業專業人士。MODELS 2019是一個論壇，參與者可以圍繞建模和模型驅動的軟件和系統交流前沿研究成果和創新實踐經驗。今年的版本將為建模社區提供進一步推進建模基礎的機會，并在網絡物理系統、嵌入式系統、社會技術系統、云計算、大數據、機器學習、安全、開源等新興領域提出建模的創新應用以及可持續性。
官網鏈接： ·
推斷
·
模型評估
·
Extensibility
·
可約的
·
In large language models like the Generative Pre-trained Transformer, the Mixture of Experts paradigm has emerged as a powerful technique for enhancing model expressiveness and accuracy. However, deploying GPT MoE models for parallel inference on distributed systems presents significant challenges, primarily due to the extensive Alltoall communication required for expert routing and aggregation. This communication bottleneck exacerbates the already complex computational landscape, hindering the efficient utilization of high-performance computing resources. In this paper, we propose a lightweight optimization technique called ExFlow, to largely accelerate the inference of these MoE models. We take a new perspective on alleviating the communication overhead by exploiting the inter-layer expert affinity. Unlike previous methods, our solution can be directly applied to pre-trained MoE models without any fine-tuning or accuracy degradation. By proposing a context-coherent expert parallelism on distributed systems, our design only uses one Alltoall communication to deliver the same functionality while previous methods all require two Alltoalls. By carefully examining the conditional probability in tokens' routing across multiple layers, we proved that pre-trained GPT MoE models implicitly exhibit a strong inter-layer expert affinity. We then design an efficient integer programming model to capture such features and show that by properly placing the experts on corresponding GPUs, we can reduce up to 67% cross-GPU routing latency. Our solution beats the cutting-edge MoE implementations with experts from 8 to 64, with up to 2.2x improvement in inference throughput. We further provide a detailed study of how the model implicitly acquires this expert affinity at the very early training stage and how this affinity evolves and stabilizes during training.
Some of the most successful knowledge graph embedding (KGE) models for link prediction -- CP, RESCAL, TuckER, ComplEx -- can be interpreted as energy-based models. Under this perspective they are not amenable for exact maximum-likelihood estimation (MLE), sampling and struggle to integrate logical constraints. This work re-interprets the score functions of these KGEs as circuits -- constrained computational graphs allowing efficient marginalisation. Then, we design two recipes to obtain efficient generative circuit models by either restricting their activations to be non-negative or squaring their outputs. Our interpretation comes with little or no loss of performance for link prediction, while the circuits framework unlocks exact learning by MLE, efficient sampling of new triples, and guarantee that logical constraints are satisfied by design. Furthermore, our models scale more gracefully than the original KGEs on graphs with millions of entities.
The emergence of large language models (LLMs), such as Generative Pre-trained Transformer 4 (GPT-4) used by ChatGPT, has profoundly impacted the academic and broader community. While these models offer numerous advantages in terms of revolutionizing work and study methods, they have also garnered significant attention due to their potential negative consequences. One example is generating academic reports or papers with little to no human contribution. Consequently, researchers have focused on developing detectors to address the misuse of LLMs. However, most existing methods prioritize achieving higher accuracy on restricted datasets, neglecting the crucial aspect of generalizability. This limitation hinders their practical application in real-life scenarios where reliability is paramount. In this paper, we present a comprehensive analysis of the impact of prompts on the text generated by LLMs and highlight the potential lack of robustness in one of the current state-of-the-art GPT detectors. To mitigate these issues concerning the misuse of LLMs in academic writing, we propose a reference-based Siamese detector named Synthetic-Siamese which takes a pair of texts, one as the inquiry and the other as the reference. Our method effectively addresses the lack of robustness of previous detectors (OpenAI detector and DetectGPT) and significantly improves the baseline performances in realistic academic writing scenarios by approximately 67% to 95%.
With the recent advancement of Large Language Models (LLMs), generating functionally correct code has become less complicated for a wide array of developers. While using LLMs has sped up the functional development process, it poses a heavy risk to code security. Code generation with proper security measures using LLM is a significantly more challenging task than functional code generation. Security measures may include adding a pair of lines of code with the original code, consisting of null pointer checking or prepared statements for SQL injection prevention. Currently, available code repair LLMs generate code repair by supervised fine-tuning, where the model looks at cross-entropy loss. However, the original and repaired codes are mostly similar in functionality and syntactically, except for a few (1-2) lines, which act as security measures. This imbalance between the lines needed for security measures and the functional code enforces the supervised fine-tuned model to prioritize generating functional code without adding proper security measures, which also benefits the model by resulting in minimal loss. Therefore, in this work, for security hardening and strengthening of generated code from LLMs, we propose a reinforcement learning-based method for program-specific repair with the combination of semantic and syntactic reward mechanisms that focus heavily on adding security and functional measures in the code, respectively.
Pretrained language models (PLMs) are today the primary model for natural language processing. Despite their impressive downstream performance, it can be difficult to apply PLMs to new languages, a barrier to making their capabilities universally accessible. While prior work has shown it possible to address this issue by learning a new embedding layer for the new language, doing so is both data and compute inefficient. We propose to use an active forgetting mechanism during pretraining, as a simple way of creating PLMs that can quickly adapt to new languages. Concretely, by resetting the embedding layer every K updates during pretraining, we encourage the PLM to improve its ability of learning new embeddings within a limited number of updates, similar to a meta-learning effect. Experiments with RoBERTa show that models pretrained with our forgetting mechanism not only demonstrate faster convergence during language adaptation but also outperform standard ones in a low-data regime, particularly for languages that are distant from English.
Reinforcement Learning from Human Feedback (RLHF) has become a crucial technology for aligning language models with human values and intentions, enabling models to produce more helpful and harmless responses. Reward models are trained as proxies for human preferences to drive reinforcement learning optimization. While reward models are often considered central to achieving high performance, they face the following challenges in practical applications: (1) Incorrect and ambiguous preference pairs in the dataset may hinder the reward model from accurately capturing human intent. (2) Reward models trained on data from a specific distribution often struggle to generalize to examples outside that distribution and are not suitable for iterative RLHF training. In this report, we attempt to address these two issues. (1) From a data perspective, we propose a method to measure the strength of preferences within the data, based on a voting mechanism of multiple reward models. Experimental results confirm that data with varying preference strengths have different impacts on reward model performance. We introduce a series of novel methods to mitigate the influence of incorrect and ambiguous preferences in the dataset and fully leverage high-quality preference data. (2) From an algorithmic standpoint, we introduce contrastive learning to enhance the ability of reward models to distinguish between chosen and rejected responses, thereby improving model generalization. Furthermore, we employ meta-learning to enable the reward model to maintain the ability to differentiate subtle differences in out-of-distribution samples, and this approach can be utilized for iterative RLHF optimization.
Recent code large language models (LLMs) have shown promising performance in generating standalone functions but face limitations in repository-level code generation due to their lack of awareness of repository-level dependencies (e.g., user-defined attributes), resulting in dependency errors such as undefined-variable and no-member errors. In this work, we introduce ToolGen, an approach that integrates autocompletion tools into the code LLM generation process to address these dependencies. ToolGen comprises two main phases: Data Augmentation and Model Fine-tuning (Offline), and Tool-integrated Code Generation (Online). During the offline phase, ToolGen augments functions within a given code corpus with a special mark token, indicating positions to trigger autocompletion tools. These augmented functions, along with their corresponding docstrings, are then used to fine-tune a selected code LLM. In the online phase, ToolGen iteratively generates functions by predicting tokens step-by-step using the fine-tuned LLM. Whenever a mark token is encountered, ToolGen invokes the autocompletion tool to suggest code completions and selects the most appropriate one. We conduct comprehensive experiments to evaluate ToolGen's effectiveness in repository-level code generation. To facilitate this evaluation, we create a benchmark comprising 680 real-world code repositories and introduce two new repository-level metrics: Dependency Coverage and Success Rate. The results demonstrate that ToolGen significantly improves dependency coverage by 15.2% to 45.8% and success rates by 10.9% to 42.2% across three distinct code LLMs, while maintaining competitive performance in widely-recognized similarity metrics. Furthermore, our generalizability evaluation confirms ToolGen's consistent performance when applied to diverse code LLMs, including various model architectures and scales.
Large Language Models (LLMs) have significantly advanced natural language processing (NLP) with their impressive language understanding and generation capabilities. However, their performance may be suboptimal for long-tail or domain-specific tasks due to limited exposure to domain-specific knowledge and vocabulary. Additionally, the lack of transparency of most state-of-the-art (SOTA) LLMs, which can only be accessed via APIs, impedes further fine-tuning with custom data. Moreover, data privacy is a significant concern. To address these challenges, we propose the novel Parametric Knowledge Guiding (PKG) framework, which equips LLMs with a knowledge-guiding module to access relevant knowledge at runtime without altering the LLMs' parameters. Our PKG is based on open-source "white-box" small language models, allowing offline storage of any knowledge that LLMs require. We demonstrate that our PKG framework can enhance the performance of "black-box" LLMs on a range of long-tail and domain-specific downstream tasks requiring factual, tabular, medical, and multimodal knowledge.
Graph Neural Networks (GNNs) have recently been used for node and graph classification tasks with great success, but GNNs model dependencies among the attributes of nearby neighboring nodes rather than dependencies among observed node labels. In this work, we consider the task of inductive node classification using GNNs in supervised and semi-supervised settings, with the goal of incorporating label dependencies. Because current GNNs are not universal (i.e., most-expressive) graph representations, we propose a general collective learning approach to increase the representation power of any existing GNN. Our framework combines ideas from collective classification with self-supervised learning, and uses a Monte Carlo approach to sampling embeddings for inductive learning across graphs. We evaluate performance on five real-world network datasets and demonstrate consistent, significant improvement in node classification accuracy, for a variety of state-of-the-art GNNs.
Semantic Role Labeling (SRL) is believed to be a crucial step towards natural language understanding and has been widely studied. Recent years, end-to-end SRL with recurrent neural networks (RNN) has gained increasing attention. However, it remains a major challenge for RNNs to handle structural information and long range dependencies. In this paper, we present a simple and effective architecture for SRL which aims to address these problems. Our model is based on self-attention which can directly capture the relationships between two tokens regardless of their distance. Our single model achieves F$_1=83.4$ on the CoNLL-2005 shared task dataset and F$_1=82.7$ on the CoNLL-2012 shared task dataset, which outperforms the previous state-of-the-art results by $1.8$ and $1.0$ F$_1$ score respectively. Besides, our model is computationally efficient, and the parsing speed is 50K tokens per second on a single Titan X GPU.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191