In human–robot collaboration, people's perceptions of robot abilities can shape how they interact with it. Drawing on insights from a pilot study, we hypothesized that a robot's backstory and gaze patterns might strengthen perceptions of competence. To examine this, we conducted a 2×2 (backstory x gaze patterns) mixed-method study in a museum-based lab setting, where 66 participants completed two open-ended gift-box assembly tasks (makeup vs. pet-owner items) with a Misty robot. Contrary to our expectations and much of the HRI literature, we found no significant differences across experimental conditions. However, participants' interactions revealed important insights about collaboration dynamics. Two distinct collaboration levels emerged: while a quarter of participants maintained low levels of collaboration, the majority engaged in highly collaborative exchanges with the robot. These collaboration levels were associated with different perceptions of warmth and competence and different interaction patterns. A thematic analysis showed that in low-collaboration cases, participants' self-efficacy (i.e. the confidence they held in their own task competence) did not appear to influence the interaction. However, within high-collaboration cases, self-efficacy shaped how participants engaged with the robot. Those with high self-efficacy critically evaluated the robot's contributions against their own expertise, whereas those with low self-efficacy leaned more heavily on the robot's guidance. Across both groups, several factors seemed to enhance collaboration: the robot's answer justifications, occasional pushback from the robot, and participants' growing familiarity with the interaction. We discuss how future systems should account for users' confidence levels, balance responsiveness with constructive disagreement, and support adaptation to foster more balanced and effective collaboration.

Large Language Models (LLMs) often retain inaccurate or outdated information from pre-training, leading to incorrect predictions or biased outputs during inference. While existing model editing methods can address this challenge, they struggle with editing large amounts of factual information simultaneously and may compromise the general capabilities of the models. In this paper, our empirical study demonstrates that it is feasible to edit the internal representations of LLMs and replace the entities in a manner similar to editing natural language inputs. Based on this insight, we introduce the Latent Knowledge Scalpel (LKS), an LLM editor that manipulates the latent knowledge of specific entities via a lightweight hypernetwork to enable precise and large-scale editing. Experiments conducted on Llama-2 and Mistral show even with the number of simultaneous edits reaching 10,000, LKS effectively performs knowledge editing while preserving the general abilities of the edited LLMs. Code is available at: https://github.com/Linuxin-xxx/LKS.

With the development of large language models (LLMs) technology, generative AI (GAI) has significantly lowered the barriers to software reverse engineering attacks. Traditional database system security controls, face new challenges when confronted with the powerful analytical capabilities of GAI. This paper provides a detailed demonstration of the reverse engineering of database cryptographic functions using GAI tools. It finds that existing encryption mechanisms in embedded DBSs need new approaches to prevent internal data attacks. The main contributions are in two aspects. First, by emulating human reverse engineering and encrypted data theft behaviors, it demonstrates how to perform software reverse engineering using GAI tools. It reveals that LLMs can accelerate analysts in obtaining the keys needed for decryption or in identifying cache extraction points for decrypted data, enabling attackers to access sensitive information. Second, to counteract the strong code generation and analytical capabilities of LLMs, the paper proposes a solution based on random data blocks and timestamp-based decryption mechanisms. This approach aims to increase the cost for attackers using GAI tools to perform software reverse engineering and exploit potential vulnerabilities in database systems.