ai · arxiv/cs.LG · 8 min
Chain-of-Thought Supervision Eliminates Sample Complexity Growth
New theoretical analysis shows intermediate reasoning steps remove dependence on generation length, while end-to-end learning scales unpredictably with sequence depth.
Apr 21, 2026 Read →