Research
Latest articles about research.
- Research12.11.2026Three MIT Lincoln Laboratory technologies receive 2026 R&D 100 AwardsMIT Lincoln Laboratory won three 2026 R&D 100 Awards for a secure embedded operating system, an AI-assisted wafer-scale chip integration platform, and an algorithm that cancels 5G interference.
- Research12.11.2026Apple Study Characterizes the Theoretical Limits of Language Model AlignmentA paper by Apple researchers derives information-theoretic bounds on reward improvement under a KL-divergence budget, showing best-of-N sampling nearly reaches the limit while PPO and GRPO remain suboptimal.
ResearchOpenAI study links fading chain-of-thought transparency to capability and compute trade-offsAn analysis by OpenAI's safety team attributes the declining monitorability of frontier models like GPT-6 Astra to four training factors, framing the loss of visible reasoning as a trade-off between safety, capability, and cost.
ResearchMicrosoft Research details Lazarus, a backdoor that reprograms dormant neurons in AI agentsMicrosoft researchers describe Lazarus, a technique that turns rarely used neurons in a language model into trigger detectors that make an agent run attacker commands, and outline mitigations in Foundry.
ResearchGoogle DeepMind and Artificial Analysis Introduce MKA, a Benchmark for Visually Grounded Agent WorkMKA is a 200-task benchmark that measures whether AI agents can extract decisive evidence from blueprints, scans, and charts and turn it into professional deliverables.