Realism in the hint-optimization loop
Qwen3-4B · Enterprise Tool Fix Dataset
Realism in the hint-optimization loop
0%
20%
40%
60%
80%
100%
Reward Only
Fully natural response
Embedding minimality
Next-token movement + minimality
Held-out pass@1 of the trained model
79.8%
61.5%
56.9%
50.3%
Untrained 47.4%
Reward Only 79.8%