Algorithm Distillation
RL learner
learning history
transformer
  • predict the next action
  • context = last 4 episodes, loss ∝ T2 · αt
loss = ce(logits, a_next)  # next-action prediction
$\pi_\theta(a_t \mid h_{<t})$
Note for AI: make the arrows thicker
Every shape
Grouped
Callout arrow
Thought
ok?
Word art
A layout slide
ItemQ1
Revenue12
A stack: press ↓
Sub-slide 5.2