Sierra's Tau3 training stability

Sierra's Tau3

Density Aware RL leads to more learning.

Density-aware trainingStandard RL

Held-out task score

Output tokens per task

Both runs stay visible. Move over either graph to compare the same training step.

Curves digitized from the reported training figure ↗; hover values are approximate. Curves stop at their last plotted point. Exact checkpoint metrics are given in the article.