InfiniteWeb: Endless Practice Websites for Smarter GUI Agents What if AI helpers could rehearse tasks on thousands of realistic websites before helping you? InfiniteWeb makes that possible by automatically building full, functional web environments for training GUI agents. * Problem: real training sites are scarce, and single-page demos don't
LLMs often stick to what they “remember,” even when you show them better, up‑to‑date evidence. ContextFocus offers a fix. It’s a lightweight “activation steering” technique that nudges models to follow the provided context over their internal guesses—no finetuning required and minimal slowdown at inference. * Improves contextual
Why it matters Machine learning on health records often stalls because hospitals store events in many formats. That hurts reproducibility and sharing. What’s new * MEDS-OWL: a small, clear vocabulary (ontology) that gives common meanings to MEDS event data, so computers agree on who did what, when. * meds2rdf: a Python
Tiny robot, big brain What if low-cost robots could make their own decisions—no constant cloud connection needed? In this study, researchers embedded an autonomous software agent (written in AgentSpeak) into a small two-wheeled robot. Using only its onboard sensors and reasoning, the bot explored and solved a maze—on
AI
Can creators of generative AI face criminal charges if users make child sexual abuse material (CSAM) with their models? A new multidisciplinary study of German law says: sometimes, yes. The authors examine realistic scenarios and find that liability may extend beyond the user to independent developers, researchers, and company representatives—
Do big AI models really "remember" your phone number? A new study says: often, no. Past claims of PII leakage may be inflated by "lexical cues"—obvious hints in the prompt (like giving part of a name or address) that let models complete patterns rather than
What matters for AI safety alignment? We ran a large-scale study of 32 popular language and reasoning models to see what truly improves—or harms—safety. * Scope: six intrinsic model traits, three attack styles, five safety datasets, 56 jailbreaks plus four chain‑of‑thought (CoT) attacks; 4.6M API calls.
Jailbreak attacks on chatbots keep evolving. HoneyTrap is a new defense that doesn’t just refuse—it deceives attackers. * How it works: Four cooperating agents—Threat Interceptor, Misdirection Controller, Forensic Tracker, and System Harmonizer—detect risky prompts, steer attackers into safe decoys, trace tactics, and keep normal users unaffected. * Put
AI
Can video AIs guide real robots? Meet the WoW-World-Eval Embodied Turing Test Robots need 'world models' that predict what happens next. A new benchmark, WoW-World-Eval (Wow, wo, val), tests video foundation models on 609 robot manipulation samples across five skills: perception, planning, prediction, generalization, and execution. * 22-metric score
AI
Multi-agent AI systems can solve complex tasks, but over long runs they may slowly veer off course — a phenomenon this paper calls agent drift. As interactions pile up, behavior degrades, decisions slip, and agents stop aligning with each other. * Semantic drift: answers stray from the original intent. * Coordination drift: consensus
Deep neural networks are powerful but often opaque. This paper introduces xDNN(ASP), a method that turns a trained network into human-readable logic rules, so you can see not just what it predicts, but why. * It extracts a logic program (via Answer Set Programming) that aims to mirror the network&
NLP
Arctic-ABSA brings aspect-based sentiment analysis closer to real life. Instead of only “positive/negative/neutral,” it adds nuance, multilingual reach, and reasoning for better business decisions. * Five classes: positive, negative, neutral, mixed, and unknown. * Finds aspect-level opinions and the overall tone in one pass. * Multilingual: one model holds 87–91%