A model wrote 8 answers to the same question, and a grader gave each a reward. Training will make some answers more likely and others less. Write push(rewards): for each answer, how hard to push it up (+) or down (−).
| #1 | #2 | #3 | #4 | #5 | #6 | #7 | #8 |
|---|---|---|---|---|---|---|---|
| 3 | 9 | 5 | 7 | 2 | 8 | 6 | 0 |