arxivcs.LGcs.DC2026-07-16
LongStraw: Long-Context RL Beyond 2M Tokens under a Fixed GPU Budget
Changhai Zhou, Kieran Liu, Yuhua Zhou, Qian Qiao, Jun Gao, Harry Zhang, et al.
A widening gap separates million-token inference from RL post-training, which remains at 256K tokens or below. The gap matters for AI agents, whose observations, tool outputs, documents, and decisions accumulate over long trajectories. Unlike inference, GRPO scores and backpropag…