Abstract:An Artificial Intelligence (AI) agent is a software entity that autonomously performs tasks or makes decisions based on pre-defined objectives and data inputs. AI agents, capable of perceiving user inputs, reasoning and planning tasks, and executing actions, have seen remarkable advancements in algorithm development and task performance. However, the security challenges they pose remain under-explored and unresolved. This survey delves into the emerging security threats faced by AI agents, categorizing them into four critical knowledge gaps: unpredictability of multi-step user inputs, complexity in internal executions, variability of operational environments, and interactions with untrusted external entities. By systematically reviewing these threats, this paper highlights both the progress made and the existing limitations in safeguarding AI agents. The insights provided aim to inspire further research into addressing the security threats associated with AI agents, thereby fostering the development of more robust and secure AI agent applications.
Abstract:Preference datasets are essential for incorporating human preferences into pre-trained language models, playing a key role in the success of Reinforcement Learning from Human Feedback. However, these datasets often demonstrate conflicting alignment objectives, leading to increased vulnerability to jailbreak attacks and challenges in adapting downstream tasks to prioritize specific alignment objectives without negatively impacting others. In this work, we introduce a novel statistical metric, Alignment Dimension Conflict, to quantify the degree of conflict within preference datasets. We then present \texttt{Hummer} and its fine-grained variant, \texttt{Hummer-F}, as innovative pairwise preference datasets with reduced-conflict alignment objectives. \texttt{Hummer} is built based on UltraFeedback and is enhanced by AI feedback from GPT-4, marking as the first preference dataset aimed at reducing the competition between alignment objectives. Furthermore, we develop reward models, HummerRM and HummerRM-F, which employ a hybrid sampling approach to balance diverse alignment objectives effectively. This sampling method positions HummerRM as an ideal model for domain-specific further fine-tuning and reducing vulnerabilities to attacks.
Abstract:Large language models (LLMs) have strong capabilities in solving diverse natural language processing tasks. However, the safety and security issues of LLM systems have become the major obstacle to their widespread application. Many studies have extensively investigated risks in LLM systems and developed the corresponding mitigation strategies. Leading-edge enterprises such as OpenAI, Google, Meta, and Anthropic have also made lots of efforts on responsible LLMs. Therefore, there is a growing need to organize the existing studies and establish comprehensive taxonomies for the community. In this paper, we delve into four essential modules of an LLM system, including an input module for receiving prompts, a language model trained on extensive corpora, a toolchain module for development and deployment, and an output module for exporting LLM-generated content. Based on this, we propose a comprehensive taxonomy, which systematically analyzes potential risks associated with each module of an LLM system and discusses the corresponding mitigation strategies. Furthermore, we review prevalent benchmarks, aiming to facilitate the risk assessment of LLM systems. We hope that this paper can help LLM participants embrace a systematic perspective to build their responsible LLM systems.
Abstract:Learning the representation of data with hierarchical structures in the hyperbolic space attracts increasing attention in recent years. Due to the constant negative curvature, the hyperbolic space resembles tree metrics and captures the tree-like properties naturally, which enables the hyperbolic embeddings to improve over traditional Euclidean models. However, many real-world hierarchically structured data such as taxonomies and multitree networks have varying local structures and they are not trees, thus they do not ubiquitously match the constant curvature property of the hyperbolic space. To address this limitation of hyperbolic embeddings, we explore the complex hyperbolic space, which has the variable negative curvature, for representation learning. Specifically, we propose to learn the embeddings of hierarchically structured data in the unit ball model of the complex hyperbolic space. The unit ball model based embeddings have a more powerful representation capacity to capture a variety of hierarchical structures. Through experiments on synthetic and real-world data, we show that our approach improves over the hyperbolic embedding models significantly.
Abstract:Model-based reinforcement learning (MBRL) has shown its advantages in sample-efficiency over model-free reinforcement learning (MFRL). Despite the impressive results it achieves, it still faces a trade-off between the ease of data generation and model bias. In this paper, we propose a simple and elegant model-embedding model-based reinforcement learning (MEMB) algorithm in the framework of the probabilistic reinforcement learning. To balance the sample-efficiency and model bias, we exploit both real and imaginary data in the training. In particular, we embed the model in the policy update and learn $Q$ and $V$ functions from the real data set. We provide the theoretical analysis of MEMB with the Lipschitz continuity assumption on the model and policy. At last, we evaluate MEMB on several benchmarks and demonstrate our algorithm can achieve state-of-the-art performance.
Abstract:A hallmark of an AI agent is to mimic human beings to understand and interact with others. In this paper, we propose a collaborative multi-agent reinforcement learning algorithm to learn a \emph{joint} policy through the interactions over agents. To make a joint decision over the group, each agent makes an initial decision and tells its policy to its neighbors. Then each agent modifies its own policy properly based on received messages and spreads out its plan. As this intention propagation procedure goes on, we prove that it converges to a mean-field approximation of the joint policy with the framework of neural embedded probabilistic inference. We evaluate our algorithm on several large scale challenging tasks and demonstrate that it outperforms previous state-of-the-arts.
Abstract:We address a practical problem ubiquitous in modern industry, in which a mediator tries to learn a policy for allocating strategic financial incentives for customers in a marketing campaign and observes only bandit feedback. In contrast to traditional policy optimization frameworks, we rely on a specific assumption for the reward structure and we incorporate budget constraints. We develop a new two-step method for solving this constrained counterfactual policy optimization problem. First, we cast the reward estimation problem as a domain adaptation problem with supplementary structure. Subsequently, the estimators are used for optimizing the policy with constraints. We establish theoretical error bounds for our estimation procedure and we empirically show that the approach leads to significant improvement on both synthetic and real datasets.
Abstract:We consider the networked multi-agent reinforcement learning (MARL) problem in a fully decentralized setting, where agents learn to coordinate to achieve the joint success. This problem is widely encountered in many areas including traffic control, distributed control, and smart grids. We assume that the reward function for each agent can be different and observed only locally by the agent itself. Furthermore, each agent is located at a node of a communication network and can exchanges information only with its neighbors. Using softmax temporal consistency and a decentralized optimization method, we obtain a principled and data-efficient iterative algorithm. In the first step of each iteration, an agent computes its local policy and value gradients and then updates only policy parameters. In the second step, the agent propagates to its neighbors the messages based on its value function and then updates its own value function. Hence we name the algorithm value propagation. We prove a non-asymptotic convergence rate 1/T with the nonlinear function approximation. To the best of our knowledge, it is the first MARL algorithm with convergence guarantee in the control, off-policy and non-linear function approximation setting. We empirically demonstrate the effectiveness of our approach in experiments.
Abstract:Uplift modeling aims to directly model the incremental impact of a treatment on an individual response. It has been widely and successfully used in healthcare analytics and business operations, where one tries to measure the net effect of a new medicine on patients or to understand the impact of a marketing campaign on company revenue. In this work, we address the problem from a new angle and reformulate it as a Markov Decision Process (MDP). This new formulation allows us to handle the lack of explicit labels, to deal with any number of actions (in comparison to the normal two action uplift modeling), and to apply it to applications with responses of general types, which is a challenging task for previous methods. Furthermore, we also design an unbiased metric for more accurate offline evaluation of uplift effects, set up a better reward function for the policy gradient method to solve the problem and adopt some action-based baselines to reduce variance. We conducted extensive experiments on both a synthetic dataset and real-world scenarios, and showed that our method can achieve significant improvement over previous methods.
Abstract:Internet market makers are always facing intense competitive environment, where personalized price reductions or discounted coupons are provided for attracting more customers. Participants in such a price war scenario have to invest a lot to catch up with other competitors. However, such a huge cost of money may not always lead to an improvement of market share. This is mainly due to a lack of information about others' strategies or customers' willingness when participants develop their strategies. In order to obtain this hidden information through observable data, we study the relationship between companies and customers in the Internet price war. Theoretically, we provide a formalization of the problem as a stochastic game with imperfect and incomplete information. Then we develop a variant of Latent Dirichlet Allocation (LDA) to infer latent variables under the current market environment, which represents the preferences of customers and strategies of competitors. To our best knowledge, it is the first time that LDA is applied to game scenario. We conduct simulated experiments where our LDA model exhibits a significant improvement on finding strategies in the Internet price war by including all available market information of the market maker's competitors. And the model is applied to an open dataset for real business. Through comparisons on the likelihood of prediction for users' behavior and distribution distance between inferred opponent's strategy and the real one, our model is shown to be able to provide a better understanding for the market environment. Our work marks a successful learning method to infer latent information in the environment of price war by the LDA modeling, and sets an example for related competitive applications to follow.