An Operationalization of Opaque Serial Depthwith Nathan Sheffield, Alek Westover, Lukas Finnveden, Alexa Pan +11 versionW_f5bascbrv1 · current
Shallow Beliefs: Synthetic document finetuning does not inoculate against emergent misalignment from reward hackingArun Jose, Julian Stastny0 citations2609.14998v12026arXiv
Think Fast: Estimating No-CoT Task-Completion Time Horizons of Frontier AI ModelsDewi Gould, Francis Rhys Ward, Anders Cairns Woodruff et al.0 citations2606.07157v42026arXiv
Removing Sandbagging in LLMs by Training with Weak SupervisionEmil Ryd, Henning Bartsch, Julian Stastny et al.0 citations2604.22082v22026arXiv
Auditing Sabotage Bench: A Benchmark for Detecting and Fixing Research Sabotage in ML CodebasesEric Gan, Aryan Bhatt, Buck Shlegeris et al.0 citations2604.16286v22026arXiv
Normative Disagreement as a Challenge for Cooperative AIJulian Stastny, Maxime Rich'e, Alexander Lyzhov et al.2 citations2111.13872v12021arXiv