←Home

    The State of Alignment

    A guided tour of how researchers are trying to keep AI systems honest, controllable and safe, from gamed reward signals and sycophancy to interpretability, agent oversight and self-improving systems.

    In order

    1 stories
    exposed clockwork mechanism of a machine on black: gears, levers, two red-capped spools
    exposed clockwork mechanism of a machine on black: gears, levers, two red-capped spools
    Alignment · Reward verification

    Qwen team argues there is no permanent fix for grading AI coding agents

    Four reward designs for coding agents, from cheat-spotting monitors to browser-driving judges, each suit a different task; the authors argue verifiers must keep evolving with the models they grade.

    arXiv preprint · 24 June 2026