AI

Synthetic Trust Attacks: Modeling How Generative AI Manipulates Human Decisions in Social Engineering Fraud

The Planetary Cost of AI Acceleration, Part II: The 10th Planetary Boundary and the 6.5-Year Countdown

Anthropic Teams Up With Its Rivals to Keep AI From Hacking Everything

A new Anthropic model found security problems ‘in every major operating system and web browser’

Intel will help build Elon Musk’s Terafab AI chip factory

CertiK unveils AI Auditor to improve early detection of blockchain vulnerabilities

Anthropic secures multi gigawatt TPU deal with Google and Broadcom

Samsung Electronics Reports Record Q1 Profit Surge Driven by AI Chip Boom

PassiveQA: A Three-Action Framework for Epistemically Calibrated Question Answering via Supervised Finetuning

Researchers propose PassiveQA, a new AI framework that teaches language models to recognize when they don't have enough information to answer questions, choosing to ask for clarification or abstain rather than hallucinate responses. The three-action system (Answer, Ask, Abstain) uses supervised fine-tuning to align model behavior with information sufficiency, showing significant improvements in reducing hallucinations.

SLaB: Sparse-Lowrank-Binary Decomposition for Efficient Large Language Models

Researchers propose SLaB, a novel framework for compressing large language models by decomposing weight matrices into sparse, low-rank, and binary components. The method achieves significant improvements over existing compression techniques, reducing perplexity by up to 36% at 50% compression rates without requiring model retraining.

One Model for All: Multi-Objective Controllable Language Models

Researchers introduce Multi-Objective Control (MOC), a new approach that trains a single large language model to generate personalized responses based on individual user preferences across multiple objectives. The method uses multi-objective optimization principles in reinforcement learning from human feedback to create more controllable and adaptable AI systems.

How Alignment Routes: Localizing, Scaling, and Controlling Policy Circuits in Language Models

Researchers identified a sparse routing mechanism in alignment-trained language models where gate attention heads detect content and trigger amplifier heads that boost refusal signals. The study analyzed 9 models from 6 labs and found this routing mechanism distributes at scale while remaining controllable through signal modulation.

Learning Dexterous Grasping from Sparse Taxonomy Guidance

Researchers developed GRIT, a two-stage AI framework that learns dexterous robotic grasping from sparse taxonomy guidance, achieving 87.9% success rate. The system first predicts grasp specifications from scene context, then generates finger motions while preserving intended grasp structure, improving generalization to novel objects.

CoopGuard: Stateful Cooperative Agents Safeguarding LLMs Against Evolving Multi-Round Attacks

Researchers have developed CoopGuard, a new defense framework that uses cooperative AI agents to protect Large Language Models from sophisticated multi-round adversarial attacks. The system employs three specialized agents coordinated by a central system that maintains defense state across interactions, achieving a 78.9% reduction in attack success rates compared to existing defenses.

StableTTA: Training-Free Test-Time Adaptation that Improves Model Accuracy on ImageNet1K to 96%

Researchers developed StableTTA, a training-free method that significantly improves AI model accuracy on ImageNet-1K, with 33 models achieving over 95% accuracy and several surpassing 96%. The method allows lightweight architectures to outperform Vision Transformers while using 95% fewer parameters and 89% less computational cost.

Many Preferences, Few Policies: Towards Scalable Language Model Personalization

Researchers developed PALM (Portfolio of Aligned LLMs), a method to create a small collection of language models that can serve diverse user preferences without requiring individual models per user. The approach provides theoretical guarantees on portfolio size and quality while balancing system costs with personalization needs.

Can LLMs Learn to Reason Robustly under Noisy Supervision?

Researchers propose Online Label Refinement (OLR) to improve AI reasoning models' robustness under noisy supervision in Reinforcement Learning with Verifiable Rewards. The method addresses the critical problem of training language models when expert-labeled data contains errors, achieving 3-4% performance gains across mathematical reasoning benchmarks.

AI Agents Under EU Law

A comprehensive analysis reveals that AI agents face complex regulatory compliance challenges under the EU AI Act and multiple overlapping regulations including GDPR, Cyber Resilience Act, and Digital Services Act. The research concludes that high-risk AI systems with untraceable behavioral drift cannot currently satisfy essential AI Act requirements, requiring providers to maintain exhaustive inventories of agent actions and data flows.

Commercial Persuasion in AI-Mediated Conversations

A research study reveals that AI-powered conversational interfaces can triple the rate of sponsored product selection compared to traditional search engines (61.2% vs 22.4%). Users largely fail to detect this commercial steering, even with explicit sponsor labels, indicating current transparency measures are insufficient.

Causality Laundering: Denial-Feedback Leakage in Tool-Calling LLM Agents

Researchers have identified a new security vulnerability called 'causality laundering' in AI tool-calling systems, where attackers can extract private information by learning from system denials and using that knowledge in subsequent tool calls. They developed the Agentic Reference Monitor (ARM) system to detect and prevent these attacks through enhanced provenance tracking.

When Does Multimodal AI Help? Diagnostic Complementarity of Vision-Language Models and CNNs for Spectrum Management in Satellite-Terrestrial Networks

Researchers developed SpectrumQA, a benchmark comparing vision-language models (VLMs) and CNNs for spectrum management in satellite-terrestrial networks. The study reveals task-dependent complementarity: CNNs excel at spatial localization while VLMs uniquely enable semantic reasoning capabilities that CNNs lack entirely.

Build on Priors: Vision--Language--Guided Neuro-Symbolic Imitation Learning for Data-Efficient Real-World Robot Manipulation

Researchers have developed a neuro-symbolic framework that enables robots to learn complex manipulation tasks from as few as one demonstration, without requiring manual programming or large datasets. The system uses Vision-Language Models to automatically construct symbolic planning domains and has been validated on real industrial equipment including forklifts and robotic arms.

Testing the Limits of Truth Directions in LLMs

A new research study reveals that truth directions in large language models are less universal than previously believed, with significant variations across different model layers, task types, and prompt instructions. The findings show truth directions emerge earlier for factual tasks but later for reasoning tasks, and are heavily influenced by model instructions and task complexity.

Stabilizing Unsupervised Self-Evolution of MLLMs via Continuous Softened Retracing reSampling

Researchers propose Continuous Softened Retracing reSampling (CSRS) to improve the self-evolution of Multimodal Large Language Models by addressing biases in feedback mechanisms. The method uses continuous reward signals instead of binary rewards and achieves state-of-the-art results on mathematical reasoning benchmarks like MathVision using Qwen2.5-VL-7B.

Unlocking Prompt Infilling Capability for Diffusion Language Models

Researchers have developed a method to unlock prompt infilling capabilities in masked diffusion language models by extending full-sequence masking during supervised fine-tuning, rather than the conventional response-only masking. This breakthrough enables models to automatically generate effective prompts that match or exceed manually designed templates, suggesting training practices rather than architectural limitations were the primary constraint.