LongStraw: Long-Context RL Beyond 2M Tokens under a Fixed GPU Budget
LongStraw trains multi-million-token RL prompts by keeping only the needed prompt state live, then replaying policy responses one at a time.
The paper says the bottleneck in long-context GRPO is live state and gradients, not just attention cost. LongStraw captures the shared prompt without autograd, scores old-policy and reference branches without a graph, and bounds the active training graph to the response suffix. The authors instantiate it for Qwen3.6-27B and GLM-5.2, with separate retained-state and replay paths under one transaction contract. HF Daily Papers' note
The paper says the bottleneck in long-context GRPO is live state and gradients, not just attention cost. LongStraw captures the shared prompt without autograd, scores old-policy and reference branches without a graph, and bounds the active training graph to the response suffix. The authors instantiate it for Qwen3.6-27B and GLM-5.2, with separate retained-state and replay paths under one transaction contract. HF Daily Papers' note
score 6