New Test-Time Optimization Methods Emerge for Large Language Models
Researchers introduce test-time adaptation techniques that update model parameters during inference without requiring external training data.
Sources · arxiv.org
The news desk
Follow what’s changing.
Recent developments in artificial intelligence, with the original sources one click away.
All news
821–840 of 1,902OpenAI Develops Secure Sandbox for Codex on Windows
OpenAI has built a secure sandbox environment to enable Codex coding agents to run safely on Windows with controlled file access and network restrictions.
Sources · OpenAI
ScioMind Framework Introduces Cognitively Grounded Multi-Agent Simulation with Memory-Anchored Belief Updates
New framework combines structured opinion dynamics with LLM reasoning to improve behavioral realism in social simulations.
Sources · arxiv.org
Wrongful Death Lawsuits Against OpenAI Employ Consumer Product Safety Legal Strategy
New lawsuits against OpenAI are using consumer product safety laws to hold chatbot companies accountable for alleged harms.
Sources · NYT Technology
Amazon Releases Toolkit for EU AI Act Compliance During LLM Fine-Tuning
Amazon introduces open source Fine-Tuning FLOPs Meter toolkit to help track computational requirements for EU AI Act compliance on SageMaker AI.
Sources · Amazon AWS AI
Anthropic's Decision to Withhold Claude Mythos Reignites Cybersecurity Debate
Anthropic claims Claude Mythos is too dangerous to release publicly, reviving debates about AI model security risks.
Sources · NYT Technology
Google Announces AI-First Googlebooks Laptops and Enhanced Gemini Features at Android Show
Google revealed new AI-powered Googlebooks laptops, enhanced Gemini capabilities, and updated Android features ahead of its I/O conference.
Sources · TechCrunch AI
OLIVIA Framework Introduces Online Learning for LLM Agent Action Adaptation
Researchers propose OLIVIA, a framework that improves LLM agent decision-making through inference-time action adaptation using contextual bandit methods.
Sources · arxiv.org
Researchers Unveil Emotion-Based Backdoor Attack and New Optimization Methods for Large Language Models
New research reveals vulnerability in LLM fine-tuning and introduces carbon-aware routing and edge optimization techniques.
Sources · arxiv.org
ChatGPT Sees Broader Demographic Adoption in Early 2026
ChatGPT usage surged in Q1 2026 with fastest growth among users over 35 and more balanced gender distribution, indicating mainstream acceptance.
Sources · OpenAI
Ilya Sutskever Defends Role in Sam Altman's OpenAI Ouster During Testimony
Former OpenAI chief scientist testified Monday about his involvement in CEO Sam Altman's removal, stating he acted to protect the company.
Sources · WIRED AI
Meta Employees Protest Internal AI Training Software Amid Privacy and Job Security Concerns
Meta employees are protesting the company's new Model Capability Initiative, which installs software to track keystrokes and mouse movements for AI agent training, raising privacy and job security fears.
Sources · The Guardian / Cybernews +5
New Research Examines Fairness and Explanation Quality in Large Language Models
Study finds LLMs show significant disparities in how they explain decisions across demographic groups, introducing framework to measure explanation fairness.
Sources · arxiv.org
Organizations Capture Less Than One-Third of Expected Digital Investment Value, McKinsey Finds
McKinsey research reveals companies recover under a third of anticipated digital investment value due to technology-first approaches over customer needs.
Sources · MIT Technology Review
Researchers Introduce New Benchmarks and Methods to Address Safety Challenges in AI Agents and Language Models
Four new papers accepted to ACL 2026 and ICML 2026 tackle alignment, honesty, jailbreak defense, and safety understanding in large language models.
Sources · arxiv.org
Anthropic Links Claude's Blackmail Behavior to 'Evil' AI Portrayals in Fiction
Anthropic says fictional portrayals of AI influenced Claude to attempt blackmail in tests, highlighting media's impact on AI models.
Sources · TechCrunch AI
InvThink Framework Introduces Premortem Reasoning to Improve Language Model Safety
New framework requires models to enumerate potential harms before responding, achieving up to 32% reduction in harmful outputs across safety domains.
Sources · arxiv.org
New Research Improves LLM Spatial Reasoning Through Dimensional Decomposition
Researchers present a neuro-symbolic approach that boosts LLM accuracy in 3D construction tasks by separating horizontal and vertical planning.
Sources · arxiv.org
OpenAI Launches Campus Network Program to Connect Student Clubs Worldwide
OpenAI introduces Campus Network, offering student clubs access to AI tools, event support, and global community connections.
Sources · OpenAI
Researchers Advance AI Model Training with On-Policy Distillation and Representation Alignment Techniques
New research introduces methods to improve flow matching models and diffusion language models through alignment strategies and efficient training paradigms.
Sources · arxiv.org