The Mirage of Optimizing Training Policies: Monotonic Inference Policies as the Real Objective for LLM Reinforcement Learning

  • 类型:arxiv
  • 标识:2606.29526
  • 链接:https://arxiv.org/abs/2606.29526
  • 主分类:engineering
  • 形态:method
  • 被引:1
  • 被引来源:Semantic Scholar
  • S2被引:1
  • OpenAlex被引:0
  • 影响力被引:0
  • TLDR:Monotonic Inference Policy Update (MIPU) is introduced, a two-step LLM RL framework that constructs sampler-referenced candidate updates and selectively accepts synchronized candidates using an inference-side gap proxy, and experiments show that MIPU improves average reasoning performance and training stability.
  • OpenAlex ID:W7166710770
  • OpenAlex DOI:10.48550/arxiv.2606.29526
  • DOI:10.48550/arxiv.2606.29526
  • DOI来源:OpenAlex
  • 开放获取:green
  • 开放获取链接:https://doi.org/10.48550/arxiv.2606.29526
  • OpenAlex更新:2026-07-19
  • 副分类:llm-infra
  • 待LLM分类:否
  • 标题中文:训练策略优化的幻象:单调推理策略才是 LLM 强化学习的真正目标
  • TLDR中文:本文提出 Monotonic Inference Policy Update (MIPU),一种两步式 LLM 强化学习框架:构建采样器引用的候选更新,并使用推理侧差距代理选择性接受同步候选;实验表明 MIPU 提升了平均推理性能与训练稳定性。
  • 来源文件
  • /inbox/tom/_candidates/2026-07-06-agent-rag-longcontext-candidates.json
  • [S2 enrich]
  • [OpenAlex backfill]