Google DeepMind announced Genie 4, a real-time world model that generates interactive simulations with native audio and can be steered with text.
Why it matters: For agent builders, Genie 4 offers a way to test multi-step behavior in steerable simulated environments before bringing systems into the real world, as Google describes. Its stated limits—especially drift and inconsistent long-term behavior—are important when assessing whether a simulation is reliable enough for evaluation or training.
The HoS abstract reports that learned per-head attention gates reduced RL-Hammer attack success against GPT-OSS-20B from 97% to 10%.
Why it matters: For agent builders, the abstract suggests attention-level instruction-data separation may improve robustness without the reported utility losses seen in some alternatives under few-shot prompting. Its results cover two models and specified attack and benchmark settings, so they do not establish performance beyond those evaluations.
The abstract introduces a benchmark that measures harness reliability separately from agent task performance.
Why it matters: For agent builders, the reported gap between task success and harness-criterion compliance suggests that outcome-only evaluation can miss reliability failures. The benchmark offers a way to examine harness behavior and attribute failures separately from agent performance.