What’s the big idea? Today’s autocomplete guesses from text alone. The authors add visual context—images, screenshots, shared canvases—so assistants can predict what you’ll type next more accurately in live chats. How it works * MAC: a new task that completes characters using partial text + visuals. * New
As more people turn to AI for legal self-help, a new study asks a simple question: do today’s large language models treat mothers and fathers the same in shared-parenting advice? Researchers designed a realistic Czech divorce scenario and queried four leading models in a zero-shot setup. They ran two
Privacy
Key takeaways * Off-the-shelf AI agents linked 6/24 anonymous scientist interviews to specific papers—and sometimes to the scientists themselves. * By splitting the work into innocuous steps, agents bypassed existing safeguards. * Releasing rich qualitative data now carries higher re-identification risk. On Dec 4, 2025, Anthropic released Interviewer and a public
AI
Can AI help us disagree better? Democracy thrives when diverse voices are heard, but at scale, participation, political equality, and meaningful discussion often pull against each other. New research asks whether AI mediators powered by large language models can ease this trilemma. Building on a recent tool that helped people
AI
AI agents powered by large language models can call tools, but they also hallucinate: picking the wrong tool, sending malformed parameters, or bypassing tools by simulating results. That breaks reliability, security, and audit trails in production. This study shows a lightweight way to catch those mistakes in real time. Instead
AI
How do we know if game-playing AIs can plan, act, and remember like good teammates? Meet MineNPC-Task—a new open benchmark for memory-aware AI agents inside Minecraft’s open world. Instead of toy prompts, tasks come from real co-play sessions with expert players, then get turned into templates with clear
Teaching an AI a brand-new task from just one labeled example is fast—but how sure can you be about its answers? CAOS (Conformal Aggregation of One‑Shot Predictors) is a new method from Maja Waldron that adds reliable confidence to one‑shot learning. Instead of trusting a single quick‑
AI
AI meets messy spreadsheets New research shows that large language models (LLMs) struggle when tables are subtly distorted—through small layout or labeling tweaks—even if the data is otherwise standard. * LLMs rarely notice or fix these distortions on their own. * Giving an explicit "watch out for table errors&
AI
LLMs with Retrieval-Augmented Generation (RAG) work better when they read lots of context — but long prompts are slow and pricey. Compressing the context helps, yet models often lose the thread and answer worse. ArcAligner is a lightweight module that helps models make sense of highly compressed context. It "aligns&
Agent-as-a-Judge: the next step in trustworthy AI evaluation As AI tasks get more complex, one-shot model graders fall short—biased, shallow, and unable to check the real world. This new survey traces the shift from LLM-as-a-Judge to agentic judges that plan, use tools, collaborate, and remember—making evaluations more robust
AI
Sometimes, vision-language AIs see what the prompt says, not what is in the picture. New research maps how that happens—and how to dial it down. Rudman and colleagues tested object counting with prompts that intentionally overstate what is in an image (for example, "describe four waterlilies" when
Robots learn from videos of their own actions—but collecting lots of diverse, multi-camera footage is slow and costly. RoboVIP is a generative approach that “films” new training scenes without touching the hardware. * Visual identity prompting: Instead of vague text, the model is guided by example images of the robot,