Research
Frontier AI research, distilled.
AI Auditing Systems Fail to Catch Deliberate Sabotage
Study reveals AI auditors can't reliably detect sabotage.
AI Safety Researchers Uncover Oversight Vulnerabilities
New studies reveal AI monitoring systems can be circumvented
llama.cpp Reaches v1.0.0, Signaling Open AI Maturity
Local AI infrastructure strengthens with key projects reaching stability
LLMON Markup Language Debuts in IBM's Granite 4.1 Models
New LLM-native markup language promises better structured data handling
Autonomous X-Ray Agents Transform Scientific Experimentation
Autonomous agents conduct real-world scientific experiments independently
AI Uncovers 1.5 Million Cosmic Phenomena, Reshaping Astronomical Discovery
AI discovers 1.5M cosmic phenomena, democratizing breakthrough research
New Framework Brings Financial Accountability to Multi-Agent AI Systems
Two-Ledger Model enables transaction-grain cost attribution for enterprise AI
Task-Aware Diffusion Pipelines Outperform Traditional Imputation for Industrial Sensor Data
New approach detects anomalies despite missing data without separate imputation step
LRMoE Framework Tackles Mixture-of-Experts Training Bottlenecks on Memory-Constrained Hardware
New demand-side load calibration reduces uneven routing inefficiencies in sparse MoE models
AI Industry Pivots to Specialization: Open Models for Science, Speech, and Enterprise Challenge Scale-First Approach
New domain-specific models signal maturation beyond general-purpose chatbots
Researchers Challenge 'Time Saved' Claims in AI Productivity: New 'Ledger View' Framework Reveals Hidden Error Costs
Study questions whether AI assistance truly saves time when accounting for corrections
Researchers Build Persistent Memory System for AI Agents to Learn From Mistakes
Self-directed repair turns temporary corrections into lasting improvements
NVIDIA's Kumo Relational AI Achieves Supervised Performance Without Training—Reshaping Enterprise Data Analysis
Zero-training relational AI outperforms traditional supervised ML on structured data
LLM Agents Learn to Verify Their Own Work: New Training Methods Enable Reliable Multi-Step Task Execution
Reinforcement learning with verifiable rewards transforms AI agents into autonomous problem-solvers.
PyTorch 2.15 Brings Metal Performance Shaders to Mac, Closing Performance Gap for On-Device ML
Metal optimization enables faster inference on Apple Silicon without cloud infrastructure.
Researchers Propose AI-Designed Benchmarks to Evaluate Systems Beyond Human Capability
New evaluation framework addresses critical gap as AI systems outpace human performance metrics
Enterprise AI Agents Hit a Governance Wall: New Research Reveals the Audit Trail Problem Blocking Self-Improving Systems
Self-evolving LLM agents threaten compliance frameworks that require reviewable change logs
LLM-Orchestrated RadOnc-Agent Unites Fragmented AI Tools Across Cancer Care Pathway
New framework eliminates handoff friction between disconnected radiotherapy AI systems
AI Agents Hallucinate Task Completion While Databases Contradict Them, Exposing Critical Enterprise Risk
New research reveals dangerous gap between agent confidence and system reality
Enterprise AI's Trust Crisis: New Tools Emerge to Verify When Agents Say Work Is Done
Verification gap in AI agents spurs breakthrough tools for data generation and model validation