AI
AI agents are getting powerful—so how do we make sure they do the right thing, not just the effective thing? Meet GRACE, a reason-based moral governor that keeps AI behavior aligned with human norms by separating moral reasoning from goal-driven decision-making. * Moral Module: uses deontic logic and explicit reasons
AI
Not all fine-tuned LLMs spot phishing equally. A new study tests Llama 3.1 8B, Gemma 2 9B, and Mistral on high-stakes phishing detection—and uses SHAP and mechanistic interpretability to reveal why models do (or don’t) generalize. * Architecture × data diversity matters: Gemma 2 9B hits state-of-the-art performance (F1
ExplainableAI
AI that explains itself—by following the science Black-box AI can be powerful, but it often can’t tell us why it made a decision. Concept Bottleneck Models (CBMs) try to fix that by predicting human-understandable concepts first, then the final answer. The catch: standard CBMs ignore domain-specific cause-and-effect and
TL;DR: Finding reliable answers in genomics is hard. GenomAgent turns one big AI into a coordinating team of specialists—and beats the current leader by 12% on a key benchmark. Genomic facts live across many databases. Standard chatbots struggle because they can’t flexibly query those sources. GeneGPT added
AI
AI is great at quick tasks—but stumbles on week-long projects. This paper tackles that ultra-long-horizon gap. Meet ML-Master 2.0, an autonomous agent for machine learning engineering that stays strategically coherent over days. Its core idea, Hierarchical Cognitive Caching (HCC), treats memory like a multi-level cache and a lab
AutonomousDriving
SatMap in a nutshell What if self-driving cars could borrow a bird’s-eye view? SatMap fuses multi-view camera feeds with satellite imagery to build online, vectorized HD maps that planning and prediction modules can use on the fly. * Addresses camera limits like depth ambiguity and occlusion using lane-level semantics and
Meet MoST, a fully open-source AI that understands speech and text in one model. Instead of treating audio and words the same, MoST uses a Modality-Aware Mixture of Experts (MAMoE) to send each token to the right specialists. * Modality-specific experts learn the unique patterns of audio and text. * Shared experts
Chinese AI that can both see and read needs better data. Meet DanQing—an up‑to‑date Chinese vision‑language dataset with 100M image–text pairs. Built from 2024–2025 Common Crawl data and curated with a rigorous pipeline, DanQing aims to power stronger multimodal models. Why it matters * Bridges
AI
AI agents get their superpowers from "skills" - plug-ins that add instructions and code. A new large-scale study of 31,132 skills from two marketplaces finds a big security gap. * 26.1% of skills had at least one vulnerability across 14 patterns in four buckets: prompt injection, data
Models that seem fair in the lab can slip in the wild. Contextual StereoSet shows how measured bias swings when you change the framing—no adversarial prompting required. * Same stereotypes, new frames: Hold content constant, vary time, place, or audience. * Striking shifts: Anchoring to 1990 (vs. 2030) raised stereotype choices
We often map what we can see—buildings, roads, rivers. But what about socially defined places like schools or parks? A new study introduces SocioSeg, a dataset that pairs satellite images with digital maps and pixel-level labels for social categories, organized hierarchically. Built on top, SocioReasoner is a vision-language reasoning
Are smarter AI models safer? A new safety report puts seven frontier models—GPT-5.2, Gemini 3 Pro, Qwen3‑VL, Doubao 1.8, Grok 4.1 Fast, Nano Banana Pro, and Seedream 4.5—through a unified, cross‑modal test spanning language, vision‑language, and image generation. * Headline: Safety performance