1 paper · 1 filter
Navid Rezazadeh, Arash Gholami Davoodi
Reasoning models often generate long traces in which useful self-correction and unproductive revision are hard to distinguish. We study this distinction through backtracking dynami…