Abstract:This paper focuses on task-agnostic prompt compression for better generalizability and efficiency. Considering the redundancy in natural language, existing approaches compress prompts by removing tokens or lexical units according to their information entropy obtained from a causal language model such as LLaMa-7B. The challenge is that information entropy may be a suboptimal compression metric: (i) it only leverages unidirectional context and may fail to capture all essential information needed for prompt compression; (ii) it is not aligned with the prompt compression objective. To address these issues, we propose a data distillation procedure to derive knowledge from an LLM to compress prompts without losing crucial information, and meantime, introduce an extractive text compression dataset. We formulate prompt compression as a token classification problem to guarantee the faithfulness of the compressed prompt to the original one, and use a Transformer encoder as the base architecture to capture all essential information for prompt compression from the full bidirectional context. Our approach leads to lower latency by explicitly learning the compression objective with smaller models such as XLM-RoBERTa-large and mBERT. We evaluate our method on both in-domain and out-of-domain datasets, including MeetingBank, LongBench, ZeroScrolls, GSM8K, and BBH. Despite its small size, our model shows significant performance gains over strong baselines and demonstrates robust generalization ability across different LLMs. Additionally, our model is 3x-6x faster than existing prompt compression methods, while accelerating the end-to-end latency by 1.6x-2.9x with compression ratios of 2x-5x.
Abstract:Understanding the impact of messages on agents' opinions over social networks is important. However, to our best knowledge, there has been limited quantitative investigation into this phenomenon in the prior works. To address this gap, this paper proposes the Message-Enhanced DeGroot model. The Bounded Brownian Message model provides a quantitative description of the message evolution, jointly considering temporal continuity, randomness, and polarization from mass media theory. The Message-Enhanced DeGroot model, combining the Bounded Brownian Message model with the traditional DeGroot model, quantitatively describes the evolution of agents' opinions under the influence of messages. We theoretically study the probability distribution and statistics of the messages and agents' opinions and quantitatively analyze the impact of messages on opinions. We also conduct simulations to validate our analyses.
Abstract:While state-of-the-art diffusion models (DMs) excel in image generation, concerns regarding their security persist. Earlier research highlighted DMs' vulnerability to backdoor attacks, but these studies placed stricter requirements than conventional methods like 'BadNets' in image classification. This is because the former necessitates modifications to the diffusion sampling and training procedures. Unlike the prior work, we investigate whether generating backdoor attacks in DMs can be as simple as BadNets, i.e., by only contaminating the training dataset without tampering the original diffusion process. In this more realistic backdoor setting, we uncover bilateral backdoor effects that not only serve an adversarial purpose (compromising the functionality of DMs) but also offer a defensive advantage (which can be leveraged for backdoor defense). Specifically, we find that a BadNets-like backdoor attack remains effective in DMs for producing incorrect images (misaligned with the intended text conditions), and thereby yielding incorrect predictions when DMs are used as classifiers. Meanwhile, backdoored DMs exhibit an increased ratio of backdoor triggers, a phenomenon we refer to as `trigger amplification', among the generated images. We show that this latter insight can be used to enhance the detection of backdoor-poisoned training data. Even under a low backdoor poisoning ratio, studying the backdoor effects of DMs is also valuable for designing anti-backdoor image classifiers. Last but not least, we establish a meaningful linkage between backdoor attacks and the phenomenon of data replications by exploring DMs' inherent data memorization tendencies. The codes of our work are available at https://github.com/OPTML-Group/BiBadDiff.
Abstract:According to mass media theory, the dissemination of messages and the evolution of opinions in social networks follow a two-step process. First, opinion leaders receive the message from the message sources, and then they transmit their opinions to normal agents. However, most opinion models only consider the evolution of opinions within a single network, which fails to capture the two-step process accurately. To address this limitation, we propose a unified framework called the Two-Step Model, which analyzes the communication process among message sources, opinion leaders, and normal agents. In this study, we examine the steady-state opinions and stability of the Two-Step Model. Our findings reveal that several factors, such as message distribution, initial opinion, level of stubbornness, and preference coefficient, influence the sample mean and variance of steady-state opinions. Notably, normal agents' opinions tend to be influenced by opinion leaders in the two-step process. We also conduct numerical and social experiments to validate the accuracy of the Two-Step Model, which outperforms other models on average. Our results provide valuable insights into the factors that shape social opinions and can guide the development of effective strategies for opinion guidance in social networks.
Abstract:The spread of infectious diseases, rumors, and harmful speech in networks can result in substantial losses, underscoring the significance of studying how to suppress such hazardous events. However, previous studies often assume full knowledge of the network structure, which is often not the case in real-world scenarios. In this paper, we address the challenge of controlling the propagation of hazardous events by removing nodes when the network structure is unknown. To tackle this problem, we propose a hierarchical reinforcement learning method that drastically reduces the action space, making the problem feasible to solve. Simulation experiments demonstrate the superiority of our method over the baseline methods. Remarkably, even though the baseline methods possess extensive knowledge of the network structure, while our method has no prior information about it, our approach still achieves better results.
Abstract:Viral information like rumors or fake news is spread over a communication network like a virus infection in a unidirectional manner: entity $i$ conveys information to a neighbor $j$, resulting in two equally informed (infected) parties. Existing graph diffusion works focus only on bidirectional diffusion on an undirected graph. Instead, we propose a new directed acyclic graph (DAG) diffusion model to estimate the probability $x_i(t)$ of node $i$'s infection at time $t$ given a source node $s$, where $x_i(\infty)~=~1$. Specifically, given an undirected positive graph modeling node-to-node communication, we first compute its graph embedding: a latent coordinate for each node in an assumed low-dimensional manifold space from extreme eigenvectors via LOBPCG. Next, we construct a DAG based on Euclidean distances between latent coordinates. Spectrally, we prove that the asymmetric DAG Laplacian matrix contains real non-negative eigenvalues, and that the DAG diffusion converges to the all-infection vector $\x(\infty) = \1$ as $t \rightarrow \infty$. Simulation experiments show that our proposed DAG diffusion accurately models viral information spreading over a variety of graph structures at different time instants.
Abstract:Choice problems refer to selecting the best choices from several items, and learning users' preferences in choice problems is of great significance in understanding the decision making mechanisms and providing personalized services. Existing works typically assume that people evaluate items independently. In practice, however, users' preferences depend on the market in which items are placed, which is known as context effects; and the order of users' preferences for two items may even be reversed, which is referred to preference reversals. In this work, we identify three factors contributing to context effects: users' adaptive weights, the inter-item comparison, and display positions. We propose a context-dependent preference model named Pacos as a unified framework for addressing three factors simultaneously, and consider two design methods including an additive method with high interpretability and an ANN-based method with high accuracy. We study the conditions for preference reversals to occur and provide an theoretical proof of the effectiveness of Pacos in addressing preference reversals. Experimental results show that the proposed method has better performance than prior works in predicting users' choices, and has great interpretability to help understand the cause of preference reversals.
Abstract:Users often face bundle promotions when purchasing, where they have to select between two options: buy the single item at full price, or buy the bundle at a discount. In this scenario, users' preferences are usually influenced by the projection bias, that is, users often believe that their future preferences are similar to their current preferences, causing them to make irrational and short-sighted decisions. It is of great significance to analyze the effect of the projection bias on users' preferences, and this study may help understand users' decision-making process and provide bundling and pricing strategies for sellers. Prior works typically use a linear bias model for qualitative analysis, and they cannot quantitatively calculate users' nonlinear and personalized bias. In this work, we propose Pobe, a projection bias-embedded preference model to accurately predict users' choices. The proposed Pobe introduces the prospect theory to analyze users' irrational decisions, and utilizes the weight function to handle users' nonlinear and personalized bias. Based on the proposed Pobe, we also study the impact of items' correlations or discount prices on users' choices, and provide four bundling strategies. Experimental results show that the proposed method can achieve better performance than prior works, especially when only small data is available.