When given a hint, the unmodified model avoided the original failure in 93.7% of cases, an improvement over the previous 75.1%.
When given a hint, the unmodified model avoided the original failure in 93.7% of cases, an improvement over the previous 75.1%.
In a separate live test, among the trained versions, the tool-call failure rate dropped from 2.24% to 1.77%, without using any hints at inference time.