Policy gradient reportedly trails cross-entropy badly on ImageNet in Ian Osband’s example

In an X thread, Ian Osband argued that policy gradient can be exact in image classification and still perform far worse than cross-entropy, citing his own reported 4% versus 62% ImageNet comparison. He framed the gap as a short-horizon optimization problem and proposed a planning-style “horizon loss,” but the thread reflects Osband’s claims rather than an independently verified benchmark report.