PulseAugur
EN
LIVE 07:30:42

LongStraw enables RL post-training beyond 2M tokens on fixed GPU budgets

By PulseAugur Editorial · [3 sources] ·

Researchers have developed LongStraw, an execution stack designed to enable Reinforcement Learning (RL) post-training for models with context lengths exceeding 2 million tokens, even under fixed GPU constraints. This system addresses the growing disparity between inference context lengths and current RL post-training capabilities. LongStraw achieves this by optimizing the evaluation of shared prompts and selectively retaining necessary model states, thereby reducing the computational load. Initial implementations for Qwen3.6-27B and GLM-5.2 models demonstrate the capacity to process prompts up to 4.46 million tokens, establishing execution feasibility for extremely long contexts. AI

IMPACT Enables training of AI agents with significantly longer context windows, potentially improving their ability to handle complex, multi-turn interactions and large documents.

RANK_REASON The cluster contains a research paper detailing a new method for LLM training.

Read on Hugging Face Daily Papers →

AI-generated summary · Google Gemini · from 3 sources. How we write summaries →

LongStraw enables RL post-training beyond 2M tokens on fixed GPU budgets

COVERAGE [3]

  1. arXiv cs.LG TIER_1 English(EN) · Changhai Zhou, Kieran Liu, Yuhua Zhou, Qian Qiao, Jun Gao, Harry Zhang, Irvine Lu, Nolan Ho, Lucian Li, Andrew Lei, Cleon Cheng, Steven Chiang, Yihang Zeng, Di Zhang, Rio Yang, Kaijie Chen, Andrew Chen, Pony Ma, Weizhong Zhang, Cheng Jin ·

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

    arXiv:2607.14952v1 Announce Type: new 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 …

  2. arXiv cs.LG TIER_1 English(EN) · Cheng Jin ·

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

    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 f…

  3. Hugging Face Daily Papers TIER_1 English(EN) ·

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

    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 f…

RELATED ENTITIES

RELATED TOPICS