Θ: W_Q, W_K, W_V, W₀, η-rule, LM weights
x_t → q_t, k_t, v_t
→
k_t, v_t, W_(t−1)
↓ inner update
W_t
→
z_t = f(q_t; W_t)
↓ remaining LM
next-token logits