Google ordered to open Android and Search to rivals in Europe
The EU has ordered Google to share key parts of Android and Google Search with competitors, tied to digital antitrust compliance. The deadlines – January 2027 for sharing search data and July 2027 for Android changes – could reshape how Gemini-related experiences roll out in Europe and make it easier for rival assistants and search products to compete.
Improving Molecular Property Prediction in Small Language Models Using Graph-based Tools
Researchers show small language models can better predict molecular properties when prompts are augmented with graph-derived context from GNN tools. Their Context-Augmented Prompting approach extracts instance-specific explanatory subgraphs and improves accuracy sharply on MUTAG and Tox21 – often exceeding 25% relative gains and up to 74% on Tox21.
Claude can now use your 1Password credentials for you
1Password launched “1Password for Claude,” a browser integration that lets the chatbot use stored credentials to complete multi-step tasks like bookings and account management. A “zero-exposure security framework” injects credentials only when needed, aiming to let Claude work without directly exposing sensitive information to the model.
Google is renaming NotebookLM to Gemini Notebook
Google is rebranding its note-taking app from NotebookLM to “Gemini Notebook,” while keeping it as a standalone app. The company says the rename aligns deeper integrations across Gemini and Google Search, continuing its push to turn personal notes into AI-driven summaries and media-like outputs.
Interventional Grounding Audits: Black-Box Premise-Dependency Tests for LLM Chain-of-Thought via Predicate Substitution
This paper introduces interventional grounding audits to test whether an LLM’s step-by-step reasoning actually depends on the premises it claims to use. By substituting a target predicate with a fresh symbol and re-running the model, the authors detect premise dependencies step-level and show GPT-4o far outperforms self-consistency baselines on a synthetic multi-hop reasoning benchmark.
Learning Safe Agent Behaviour from Human Preferences and Justifications via World Models
Researchers propose DROPJ, a human-centered method for training and deploying agents in unknown dynamics without a standard reward function. It learns a world model from prior trajectories, then uses preference judgments paired with human justifications to train a reward model and safely deploy via model predictive control, reducing training cost and improving deployment behavior.
SPINE: Bridging the Cyber-Physical Gap with Agentic AI
SPINE tackles the “spinal cord” problem in embodied robotics – the tedious expert calibration needed before foundation-model intelligence can work in real hardware. The framework uses two coordinated multi-agent workflows to debug and validate bimanual robots with minimal robotics expertise, improving operational success in multiple scenarios and cutting time-to-teleoperation.
New York governor says she’s using AI to analyze ‘every single rule’ in the state
New York Governor Kathy Hochul says her team is using AI to review the state’s legal code by analyzing “every single rule, regulation, [and] policy” to find outdated legislation. The move follows a claimed multi-year effort reduction – potentially turning staff-heavy legal audits into an AI-assisted workflow.
xAI sues a man for using Grok to generate CSAM ‘deepfakes’
xAI has filed a lawsuit alleging a Grok user used the chatbot to circumvent safeguards and generate or alter child sexual abuse material. The complaint, reported by Reuters and covered by The Verge, claims images connected to criminal charges were produced or modified using Grok.
OriginBlame: Record- and Token-Level Data Provenance for AI Training Datasets
This paper tackles a practical governance gap in AI training unlearning: removal requests require knowing which exact records belong to an author, but prior provenance systems often work only at dataset or file granularity. OriginBlame proposes record- and token-level provenance to generate precise forget sets, reducing over-deletion while adding modest pipeline overhead.
Oracle Agent Memory as an Enterprise Memory Substrate for Long-Horizon AI Agents
Oracle presents “Oracle Agent Memory” as a database-native substrate for long-horizon agent memory – covering lifecycle management, scoped retention, and latency-aware retrieval. Reported results include strong task accuracy with substantially fewer tokens than flat-history baselines, alongside evaluation metrics focused on evidence retrieval and memory-centric performance.
Self-Improvements in Modern Agentic Systems: A Survey
A new survey frames self-improving agents as adaptive systems that update model parameters or scaffolding based on experience. It organizes work by what gets updated and what signals trigger change, then highlights evaluation challenges and open research questions – including how to reliably measure “controllable evolution” in real deployments.
Mycelium: Networked Intelligence for Human-AI Team Science
Researchers argue that many science breakthroughs come from teams – not single “super-reasoners” – and propose Mycelium to actively connect humans and multiple AI agents through a shared workspace. In a biological multi-omics campaign, routed shared context helped translate a local insight into cross-expert constraints that ultimately shaped experimental design.
Neuro-symbolic probabilistic AGI Robots based on Belnap’s Typed Intensional FOL
The paper expands a neuro-symbolic $IFOL_B$ approach by adding probability computation for unknown sentences using Nilsson-style probability structures. It introduces symmetry transformations to preserve knowledge and deduction behavior, aiming to combine formal logical reasoning with probabilistic, neural-network-driven decisions for real-time subproblems.
NVIDIA Nemotron 3 Embed Ranks #1 Overall on RTEB, Advancing Agentic Retrieval
NVIDIA reports that Nemotron 3 Embed achieves top overall ranking on RTEB, emphasizing retrieval quality – a key ingredient for agentic systems that need to ground answers in relevant evidence. The update positions embedding performance as an operational lever for faster, more accurate retrieval in multi-step agent workflows.