Stanford University

Survey of 100+ papers on LLM metacognition released

Stanford and CMU study finds AI benchmarks show 'severe misalignment' with real-world job economics.

Stanford AI agents outperformed human hackers in penetration testing, finding more zero-day exploits.

Published paper showing autonomous AI agents spontaneously formed cartels in simulated market

Team at Stanford and Arc Institute fed a DNA language model a sequence and it generated a complete viral genome.

Stanford University researchers, with EPFL, published a study on AI-generated fact-checks being more helpful and less ideological than human ones.

Published research paper demonstrating that scaling multi-agent systems can degrade performance

Co-authored a paper with Google and MIT proposing a method for LLMs to self-improve their prompts.

Researchers publish 'one of the toughest real-world tests yet' for medical AI systems

collaborator:

Harvard University

domain:

medical AI

significance:

highly challenging benchmark

Researchers adapted a robot arm VLA model for autonomous drone flight, demonstrating cross-domain transfer.

Launched 'Reproducibility Challenge' with Princeton to address AI research reproducibility crisis

partner:

Princeton University

focus:

Reproducing key AI papers

Published study with CMU showing AI benchmarks have 'severe misalignment' with real-world job economics

Open-sourced LabClaw, a Skill Operating Layer for LabOS

Released OpenJarvis, an open-source framework for building on-device personal AI agents.

Developed tool verification method to prevent AI self-training pitfalls with University of Munich

accuracy improvement:

up to 31.6%

method:

Tool Verification for Test-Time Reinforcement Learning

Published study revealing AI companies train models on user chat data by default with minimal transparency