Comparing scheduled sampling and consistency regularization for autoregressive ranking beyond the first token
| Method | AR Quality | TF Quality | Mismatch | KL at t=1 |
|---|---|---|---|---|
| Teacher Forcing | 0.0547 | 0.0623 | -0.0278 | 0.0286 |
| Scheduled Sampling | 0.0539 | 0.0609 | -0.0285 | 0.0286 |
| Consistency Reg. | 0.0691 | 0.0620 | -0.0287 | 0.0286 |