Large language models often produce biased or stereotypical outputs. One way to reduce this possibility is to be more inclusive in our prompts, but doing so may not come naturally to most users. Therefore, we designed a tool that coaches users to write more inclusive prompts—a strategy that leverages design friction to provide a media literacy intervention. Data from a user study (N=344) show that compared to no coaching, inclusive prompt coaching directly increased users’ awareness of algorithmic bias and their perceived prompting efficacy. It also indirectly enhanced their trust in the system and perceived trust calibration through cognitive elaboration. However, inclusive prompt coaching resulted in a less satisfying user experience. These findings have implications for ethical interventions in prompting for better communicating and combating algorithmic bias. We discuss the benefits and limitations of inclusive prompt coaching, as well as ways to balance usability for long-term adoption of generative AI systems.

ChatGPT’s memory feature is designed to provide users with greater control and more helpful responses. Yet, it remains unclear how users perceive this feature in relation to privacy. To address this gap, we conducted interviews with 20 ChatGPT users from diverse backgrounds. Our findings revealed four major characteristics that distinguish ChatGPT’s memory from human memory: perceived unforgetfulness, detailedness, accuracy, and lack of emotions, highlighting the machine-like nature of AI memory. Moreover, both ChatGPT’s memory and human memory were perceived as beneficial for relationship building. Notably, most participants experienced negative expectancy violations after learning what ChatGPT remembered about them. They expressed a strong need for greater visibility, accessibility, transparency, and user control in the design of future memory features. Drawing on users’ suggestions and theoretical frameworks on privacy management, we provide design implications for developing a more transparent, responsible, and user-aligned memory experience that helps them navigate privacy-personalization trade-offs when interacting with LLM-based memories.

Reasoning future unknowable facts on temporal knowledge graphs (TKGs) is a challenging task, holding significant academic and practical values for various fields. Existing studies exploring explainable reasoning concentrate on modeling comprehensible temporal paths relevant to the query. Yet, these path-based methods primarily focus on local temporal paths appearing in recent times, failing to capture the complex temporal paths in TKG and resulting in the loss of longer historical relations related to the query. Motivated by the Dual Process Theory in cognitive science, we propose a Cognitive Temporal Knowledge Extrapolation framework (CognTKE), which introduces a novel temporal cognitive relation directed graph (TCR-Digraph) and performs interpretable global shallow reasoning and local deep reasoning over the TCR-Digraph. Specifically, the proposed TCR-Digraph is constituted by retrieving significant local and global historical temporal relation paths associated with the query. In addition, CognTKE presents the global shallow reasoner and the local deep reasoner to perform global one-hop temporal relation reasoning (System 1) and local complex multi-hop path reasoning (System 2) over the TCR-Digraph, respectively. The experimental results on four benchmark datasets demonstrate that CognTKE achieves significant improvement in accuracy compared to the state-of-the-art baselines and delivers excellent zero-shot reasoning ability.

Reasoning large language models (LLMs) excel in complex tasks, which has drawn significant attention to reinforcement learning (RL) for LLMs. However, existing approaches allocate an equal number of rollouts to all questions during the RL process, which is inefficient. This inefficiency stems from the fact that training on simple questions yields limited gains, whereas more rollouts are needed for challenging questions to sample correct answers. Furthermore, while RL improves response precision, it limits the model’s exploration ability, potentially resulting in a performance cap below that of the base model prior to RL. To address these issues, we propose a mechanism for dynamically allocating rollout budgets based on the difficulty of the problems, enabling more efficient RL training. Additionally, we introduce an adaptive dynamic temperature adjustment strategy to maintain the entropy at a stable level, thereby encouraging sufficient exploration. This enables LLMs to improve response precision while preserving their exploratory ability to uncover potential correct pathways. The code and data is available on: https://anonymous.4open.science/r/E3-RL4LLMs-DB28

Recent studies have combined Mixture of Experts (MoE) and Parameter-Efficient Fine-tuning (PEFT) to fine-tune large language models (LLMs), holding excellent performance in multi-task scenarios while remaining resource-efficient. However, existing MoE approaches still exhibit the following limitations: (1) Current methods fail to consider that different LLM layers capture features at varying levels of granularity, leading to suboptimal performance. (2) Task-level routing methods lack generalizability to unseen tasks. (3) The uncertainty introduced by load imbalance loss undermines the effective specialization of the experts. To address these challenges, we propose HMoRA, a Hierarchical fine-tuning method that combines MoE and LoRA, employing hybrid routing that integrates token-level and task-level routing in a hierarchical manner. This hierarchical hybrid routing allows the model to more efficiently capture both fine-grained token information and broader task contexts. To improve the certainty of expert selection, a novel routing auxiliary loss is introduced. This auxiliary function also enhances the task router's ability to differentiate tasks and its generalization to unseen tasks. Additionally, several optional lightweight designs have been proposed to significantly reduce both the number of trainable parameters and computational costs. Experimental results demonstrate that HMoRA outperforms full fine-tuning across multiple NLP benchmarks, while fine-tuning only 3.9\% of the parameters. The code is available on: https://github.com/LiaoMengqi/HMoRA.

Zero-shot stance detection on social media (ZSSD-SM) aims to distinguish the attitude in tweets towards an unseen target. Previous work capture latent variables between source and target domains to perform this task, but the lack of context knowledge hinders the detection performance. Recent studies have been devoted to obtaining the accurate representation of tweets by bringing additional facts from Knowledge Graph (KG), showing promising performance. However, these knowledge injection methods still suffer from two challenges: (i) The pipeline of knowledge injection causes error accumulation and (ii) irrelevant knowledge makes them fail to understand the semantics. In this paper, we propose a novel knowledge injection method for ZSSD-SM, which adopts two training stages, namely knowledge compression and task guidance, to flexibly inject knowledge into the pre-trained language model (PLM) and adaptively expand tweets context. Specifically, in the knowledge compression stage, the latent representation of KG is reconstructed by the triplet denoising task and compressed into external matrices; while in the task guidance stage, the frozen matrices are employed to guide the PLM to adaptively extract its own context-related knowledge, and then complete the fine-tuning of the ZSSD-SM task. Extensive experiments on multiple datasets show the effectiveness of our proposed method. The code is available at: https://github.com/ShuohaoLin/KPatch.

Recommender systems (RS) have become increasingly vital for guiding health actions. While traditional systems filter content based on either demographics, personal history of activities, or preferences of other users, newer systems use social media information to personalize recommendations, based either on the users’ own activities and/or those of their friends on social media platforms. However, we do not know if these approaches differ in their persuasiveness. To find out, we conducted a user study of a fitness plan recommender system (N = 341), wherein participants were randomly assigned to one of six personalization approaches, with half of them given a choice to switch to a different approach. Data revealed that social media-based personalization threatens users’ identity and increases privacy concerns. Users prefer personalized health recommendations based on their own preferences. Choice enhances trust by providing users with a greater sense of agency and lowering their privacy concerns. These findings provide design implications for RS, especially in the preventive health domain.

Three of the most common approaches used in recommender systems are content-based filtering (matching users’ preferences with products’ characteristics), collaborative filtering (matching users with similar preferences), and demographic filtering (catering to users based on demographic characteristics). Do users’ intuitions lead them to trust one of these approaches over others, independent of the actual operations of these different systems? Does their faith in one type or another depend on the quality of the recommendation, rather than how the recommendation appears to have been derived? We conducted an empirical study with a prototype of a movie recommender system to find out. A 3 (Ostensible Recommender Type: Content vs. Collaborative vs. Demographic Filtering) x 2 (Recommendation Quality: Good vs. Bad) experiment (N=226) investigated how users evaluate systems and attribute responsibility for the recommendations they receive. We found that users trust systems that use collaborative filtering more, regardless of the system's performance. They think that they themselves are responsible for good recommendations but that the system is responsible for bad recommendations (reflecting a self-serving bias). Theoretical insights, design implications and practical solutions for the cold start problem are discussed.

AI systems collect our personal information in order to provide personalized services, raising privacy concerns and making users leery. As a result, systems have begun emphasizing overt over covert collection of information by directly asking users. This poses an important question for ethical interaction design, which is dedicated to improving user experience while promoting informed decision-making: Should the interface tout the benefits of information disclosure and frame itself as a help-provider? Or, should it appear as a help-seeker? We decided to find out by creating a mockup of a news recommendation system called Mindz and conducting an online user study (N=293) with the following four variations: AI system as help seeker vs. help provider vs. both vs. neither. Data showed that even though all participants received the same recommendations, power users tended to trust a help-seeking Mindz more whereas non-power users favored one that is both help-seeker and help-provider.