LongStraw: Long-Context RL Beyond 2M Tokens under a Fixed GPU Budget

2026-07-16

AI Summary

Current reinforcement learning (RL) post-training methods suffer from context lengths far shorter than those supported by modern inference systems, hindering agents’ ability to accumulate information over long trajectories. This work proposes LongStraw, an architecture-aware execution stack tailored for million-token-scale RL post-training. By integrating autograd-free shared prompt evaluation, selective retention of critical model states, and segmented replay with response branching, LongStraw substantially reduces training graph memory overhead under fixed GPU budgets. The approach enables, for the first time, grouped scoring and backpropagation over 2.1M tokens for the Qwen3.6-27B model on just eight H20 GPUs—with peak memory increasing by only 0.21 GB—and achieves end-to-end training of the GLM-5.2 model on 32 H20 GPUs, scaling to 4.46M tokens in stress tests while supporting both hybrid attention and mixture-of-experts (MoE) architectures.

Abstract

A growing gap separates inference context lengths from RL post-training: inference systems are approaching million-token contexts, while post-training workloads often remain at 256K tokens or below and rely on length generalization at deployment. The gap is especially important for AI agents, whose observations, tool outputs, documents, and prior decisions accumulate over long trajectories. LongStraw is an architecture-aware execution stack for million-token RL post-training under a fixed GPU budget, instantiated with Group Relative Policy Optimization (GRPO). It evaluates the shared prompt without autograd, retains only model-specific state needed by later tokens, and replays short response branches one at a time, reducing the live training graph at the cost of additional replay time. We implement it for the hybrid recurrent and full-attention Qwen3.6-27B and the compressed-attention mixture-of-experts GLM-5.2. On eight H20 GPUs, LongStraw completes grouped Qwen scoring and response backward at 2.1M positions for groups of 2 and 8; increasing the group size adds only 0.21 GB of peak allocated memory, while a separate stress test reaches 4.46M positions. On 32 H20 GPUs, we validate the end-to-end LongStraw execution path for a 2.1M-token prompt across all 78 layers of GLM-5.2. These experiments establish execution capacity rather than complete training correctness because the captured prompt state is detached and some distributed forward and gradient composition paths remain incomplete.

Problem

• long-context reinforcement learning
• RL post-training
• context length gap
• AI agents
• fixed GPU budget

Innovation

• Long-context RL
• memory-efficient training
• architecture-aware execution
• Group Relative Policy Optimization
• attention compression