Inference setup success rate by advisor access
Qwen3.5-9B. Whiskers are Wilson 95% intervals. Successes / trials are printed under each bar.
0%
20%
40%
60%
80%
100%
inference setup success rate
Qwen3.5-9B, no RL
Qwen3.5-9B after RL
no RL, no advisor: 60% (6/10), Wilson 95% 31-83%
6/10
no RL, voluntary: 30% (3/10), Wilson 95% 11-60%
3/10
no RL, forced: 25% (2/8), Wilson 95% 7-59%
2/8
after RL, without reward shaping, no advisor: 30% (3/10), Wilson 95% 11-60%
3/10
after RL, without reward shaping, voluntary: 20% (2/10), Wilson 95% 6-51%
2/10
after RL, without reward shaping, forced: 44% (8/18), Wilson 95% 25-66%
8/18
without reward shaping
after RL, with reward shaping, no advisor: 72% (13/18), Wilson 95% 49-88%
13/18
after RL, with reward shaping, voluntary: 61% (11/18), Wilson 95% 39-80%
11/18
after RL, with reward shaping, voluntary, when stuck: 67% (14/21), Wilson 95% 45-83%
14/21
after RL, with reward shaping, forced: 44% (7/16), Wilson 95% 23-67%
7/16
with reward shaping
no RL, no advisor: 60%
no advisor
voluntary
voluntary, when stuck
forced