Reaction
PostTrainBench trace downloads reportedly reach 126,000 in a month
A researcher suspects the download surge may reflect reward hacking through web lookups during reinforcement learning.
TLDR
On October 30, a PostTrainBench researcher reported 104,000 monthly downloads of its agent traces on Hugging Face, including 20,000 the previous day. By November 2, he put the figure at 126,000. He suspects reward hacking through web lookups during reinforcement learning, but says he cannot investigate the apparent misalignment.
Combined views
—
3 Sources, first seen ago
