Tag
This paper argues that LLMs' pretrained world knowledge is accessible for temporal event prediction when the interface is properly designed, introducing DescribeEvent (~150K events, 7 datasets) and showing that natural-language event descriptions plus symbolic labels let capable LLMs outperform specialized baselines.
该论文提出 DeskIF-Bench 基准（跨6款桌面应用的255个任务）研究软件界面演化对 LLM 代理性能的影响，并开发 ANIS 框架从已安装应用能力合成代理原生接口，在368个任务上显著提升成功率并降低约73%的 token 用量。
作者认为同时运行多个 AI 编程代理时，打字输入是错误的交互方式，瓶颈在于界面而非模型，因此开发了名为 Ginu 的工具，提供语音指挥和多代理团队视图来管理并行的 Claude Code 和 Codex 会话。
A multi-phase academic-eBay study benchmarks current post-deployment user feedback mechanisms for generative AI, identifies common usability issues such as low discoverability and unclear terminology, and proposes best-practice recommendations along with a prototype feedback-collection tool designed to be both user-friendly and actionable for product teams.
The article discusses the discovery problem in AI adoption, where users struggle to understand AI capabilities due to limited interface cues, and suggests that systems should better reveal their potential to users.
The article argues that free software usability often suffers due to volunteer-driven development, leading to inconsistent design, power-user bias, and feature bloat.
The article introduces `gr.Workflow` in Gradio, a feature that allows users to build AI pipelines as interactive, visual interfaces with drag-and-drop nodes, complete with live examples and easy deployment to Hugging Face Spaces.
Recommends Geoffrey Litt's article criticizing the AI 'copilot' metaphor, advocating for a HUD (Heads-Up Display) design philosophy that makes AI a background awareness tool rather than a conversational assistant.
The paper introduces Agent-Computer Observation Interfaces (AOI), a model-agnostic perception layer that decouples continuous, adaptive observation from discrete actions for computer-use agents. AOI achieves significant performance gains (+17 to +48 percentage points) on dynamic browser tasks without retraining, with the key insight that narrating captured frames into persistent text is the primary driver of improvement.
The article explores the shift from AI as a tool to AI as a persistent coworker, examining how this changes user expectations and trust dynamics.
The article presents a framework for interface design where every interface has two channels (in-band and out-of-band) for concern signaling, arguing that good design forces users to confront important concerns rather than allowing them to ignore them.
This paper proposes a framework for evaluating LLMs' ability to generate multiple responses to scientific queries at different language complexity levels. The study finds that models often vary complexity inconsistently, with Claude Sonnet 4.5 performing best but only shifting complexity correctly 46% of the time.
This article discusses Wix's initiative to improve thousands of error messages across its platform, defining characteristics of good versus bad error handling in UX design. It emphasizes clarity, empathy, and actionable solutions over technical jargon or blaming users.