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

  • 类型:arxiv
  • 标识:2607.14952
  • 链接:http://arxiv.org/abs/2607.14952v1
  • 主分类:engineering
  • 形态:application
  • 被引:2
  • 被引来源:Semantic Scholar
  • S2被引:2
  • OpenAlex被引:0
  • 影响力被引:0
  • TLDR:This work presents LongStraw, an objective-aware, architecture-aware system for resident-state virtualization, response replay, and distributed-gradient execution that bounds the live training graph by the response suffix while reusing the expensive prompt computation across the complete GRPO group.
  • OpenAlex ID:W7169616292
  • OpenAlex DOI:10.48550/arxiv.2607.14952
  • DOI:10.48550/arxiv.2607.14952
  • DOI来源:OpenAlex
  • 开放获取:green
  • 开放获取链接:https://doi.org/10.48550/arxiv.2607.14952
  • OpenAlex更新:2026-07-19
  • 副分类:llm-infra
  • 待LLM分类:否
  • 标题中文:LongStraw:固定 GPU 预算下超越 2M token 的长上下文强化学习
  • TLDR中文:本文提出 LongStraw,一个面向目标、感知架构的系统,用于 resident-state 虚拟化、response replay 和分布式梯度执行,它将实时训练图限制在 response 后缀范围内,同时在完整的 GRPO 组内复用代价高昂的 prompt 计算。
  • 来源文件
  • /inbox/tom/_candidates/2026-07-17-agent-rag-longcontext-candidates.json
  • [S2 enrich]
  • /inbox/tom/_candidates/2026-07-18-agent-rag-longcontext-candidates.json
  • /inbox/tom/_candidates/2026-07-19-agent-rag-longcontext-candidates.json
  • [OpenAlex backfill]
  • /inbox/tom/_candidates/2026-07-20-agent-rag-longcontext-candidates.json