Teacher depth across training
Qwen3.8-27B · Enterprise Chat
Depth distribution, τ = 0.1
Steps 0–9
Steps 10–19
Steps 20–29
Steps 30–39
Steps 40–49
Steps 50–62
Depth distribution, τ = 0.1
0%
5%
10%
15%
20%
25%
30%
Share of policy tokens
0
1
2
3
4
5–8
9–16
17–32
33–64
65–128
129+
Depth Dₜ (teacher tokens aligned from position t)
Share with Dₜ ≥ 9
τ = 0.01
τ = 0.03
τ = 0.1
τ = 0.3
Share with Dₜ ≥ 9
0%
20%
40%
60%
80%
100%
Share with Dₜ ≥ 9
0–9
10–19
20–29
30–39
40–49
50–62
Training steps