Open Models Almanac
Methodspost-trainingreinforcement learning algorithm

single mention · post-training

Zero-reward correction for externally supported trajectories

Resets a trajectory's effective reward to zero when evidence shows it relied on external or leaked answers, then recomputes group statistics.

The sources on this page only mention it in passing. The line above is the extraction's one-line gloss. See reinforcement learning algorithm for the methods filed beside it.

Filed under post-training › reinforcement learning algorithm