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%
AI
Meet Klear, a research system for generating video and audio together — with lips, speech, and actions that line up. The team tackles three common pain points in A/V generation: off-sync sound, weak lip–speech matches, and quality drops when one modality dominates. * One unified model: a single-tower design with