On-policy distillation (OPD) has become increasingly popular for transferring teacher capabilities to student models. In this work, we ask a critical research question: Is on-policy sampling always beneficial for distilling arbitrary teacher-student pairs? We show that a simple alternative, Semi-OPD, which distills from offline rollouts generated by the initial student, can often outperform OPD in both accuracy and training efficiency.

Many modern AI systems are designed to operate under diverse, open-ended, use-cases. To help generalize deployed systems, many deployed-system maintenance pipelines use a reactive AI flywheel that observes emerging feedback from user behavior (errors) and patches the model accordingly. However, when used as the primary maintenance mechanism, these flywheels often ignore the broader context of these errors within the system's objectives, failing to preempt potential future edge cases, which leads to more unnecessary flywheel iterations.

Terminal agents act through stochastic model generations, yet the ability to generate a useful action does not ensure its reliable execution. A poor command (e.g., wrong package install) can change the environment in ways that hinder subsequent progress, even when the model could generate a better alternative. We investigate whether allocating test-time compute at the model-harness boundary can improve action reliability and trajectory success, and what makes this allocation effective.

The growing demand for accelerated computing is driving widespread deployment of multi-accelerator systems in the cloud and at the edge. These systems are temporally shared across processes that may not mutually trust one another and often include large pools of third-party, untrusted accelerators operating within a shared virtual address space. Due to the limited capability of the Input-Output Memory Management Unit (IOMMU) in supporting address translation, SoC designers are integrating large Shared Translation Lookaside Buffers (TLBs) that serve many accelerators and processes.