AI Briefing — May 18, 2026
🔬 Research
MMSkills: Towards Multimodal Skills for General Visual Agents
92 upvotes · Kangning Zhang et al. · Paper
MMSkills introduces a framework for packaging multimodal procedural knowledge that visual agents can reuse across tasks. Instead of relying solely on text prompts or code, each skill combines textual procedures with visual state cards and keyframes, enabling agents to recognize relevant contexts and interpret visual progress indicators. The framework includes an automated trajectory-to-skill generator and a branch-loaded agent architecture that temporarily inspects visual references before distilling guidance for the main agent, showing consistent improvements across GUI and game benchmarks.
PhysBrain 1.0 Technical Report
83 upvotes · Shijie Lian et al. · Paper
PhysBrain tackles the limited physical understanding in vision-language-action models by mining structured supervision from large-scale human egocentric video. Their data engine extracts scene elements, spatial dynamics, and depth-aware relationships to create question-answer pairs for training, then transfers this physical commonsense to robot policies through capability-preserving adaptation. The approach achieves SOTA results across multiple benchmarks including ERQA, PhysBench, and RoboCasa, with particularly strong out-of-domain performance suggesting that human video provides an effective bridge from multimodal understanding to robotic action.
CiteVQA: Benchmarking Evidence Attribution for Trustworthy Document Intelligence
78 upvotes · Dongsheng Ma et al. · Paper
CiteVQA exposes a critical reliability gap in document AI: models frequently produce correct answers while citing wrong evidence regions—a phenomenon they term "Attribution Hallucination." The benchmark requires models to provide element-level bounding-box citations alongside answers across 1,897 questions in seven domains, evaluating both jointly through Strict Attributed Accuracy. Even the strongest system (Gemini-3.1-Pro-Preview) achieves only 76.0% SAA, while the best open-source model reaches just 22.5%, highlighting a major trustworthiness issue for high-stakes applications in law, finance, and medicine.
🛠 Tools
tinyhumansai/openhuman — Personal AI Super Intelligence Platform
1,690 stars/day · Repo
OpenHuman positions itself as a private, simple yet powerful personal AI assistant built in Rust. With 15,320 total stars and active daily development, it emphasizes local processing and privacy while providing what the creators call "super intelligence" capabilities. The project is in early beta with multi-language documentation support and an active community across Discord and Reddit, suggesting strong momentum in the personal AI assistant space.
Imbad0202/academic-research-skills — Claude Code Academic Pipeline
1,302 stars/day · Repo
This comprehensive Claude Code plugin suite automates the academic research pipeline from initial planning to final publication. With 10,156 stars, it handles reference hunting, citation formatting, data verification, and logical consistency checking while maintaining human-in-the-loop oversight for critical thinking tasks. The tool emphasizes quality over automation, featuring style calibration that learns from past work and quality checks to avoid machine-generated prose patterns—positioning itself as a copilot rather than autopilot for academic writing.
ruvnet/RuView — WiFi-Based Spatial Intelligence
963 stars/day · Repo
RuView transforms ordinary WiFi signals into a comprehensive spatial sensing system using Channel State Information from ESP32 sensors. The platform detects presence through walls, monitors vital signs contactlessly, recognizes activities, and maps environments—all without cameras or wearables. With 59,488 total stars, it's positioned as beta software with known limitations on single-core ESP32 variants, but demonstrates the potential for turning existing WiFi infrastructure into pervasive sensing networks through pure signal processing.
📡 Analysis
GDS weighs in on the NHS's decision to retreat from Open Source
May 17, 2026 · Post · tags: open-source, security, ai-ethics
The Government Digital Service has publicly pushed back against the NHS's decision to close their open source repositories following Project Glasswing vulnerability reports. Their key recommendation—"Keep open by default"—represents what Terence Eden interprets as a rare public escalation within UK civil service culture. This signals growing tension between security concerns and transparency principles in government technology, particularly as AI security research increasingly targets public sector codebases.
Quoting Julia Evans on CSS Mastery
May 16, 2026 · Post · tags: css, julia-evans
Julia Evans advocates for treating CSS as a serious technology worthy of deep understanding rather than dismissing it as "hard." She argues that many developer frustrations stem from not recognizing that CSS addresses genuinely complex problems like responsive layout and cross-browser compatibility. Her perspective shift from fighting CSS to respecting its design decisions transformed her development experience, suggesting that technology adoption often requires mindset changes as much as skill acquisition.
Sources: HuggingFace Papers API, GitHub Trending, simonwillison.net