What's this about?
LongStraw tackles a growing mismatch: inference systems handle million-token contexts, but RL post-training — where models learn agent behaviors — mostly still happens at 256k tokens or less, hoping the skills generalize to longer contexts. LongStraw makes RL training itself work beyond 2M tokens without needing more GPUs, and at 176 upvotes it ran away with HuggingFace's daily papers board.
What exactly is "RL post-training"?
After a model learns language from raw text (pretraining), it's refined with reinforcement learning — practicing tasks and being rewarded for good outcomes, which is how models learn to be effective agents. Doing RL on very long sessions was prohibitively memory-hungry, so labs trained short and prayed the behavior transfers ("length generalization"). For agents that work for hours accumulating context, training at realistic lengths matters — that's the gap this closes.
Why does it matter?
The fixed-GPU-budget angle is the practical hook: long-horizon agent training becomes accessible without frontier-lab clusters. The 176-upvote signal says the community sees long-context RL as the current bottleneck for capable agents.