Research Papers 论文研究 7d ago Updated 7d ago 更新于 7天前 49

Procedural Memory Distillation: Online Reflection for Self-Improving Language Models 程序记忆蒸馏:用于自我改进语言模型的在线反思

Introduces Procedural Memory Distillation (PMD), a method that captures cross-episode procedural signals often lost in standard RLVR and self-distillation approaches. Organizes memory into three abstraction levels: raw trajectories, self-reflected strategies, and recurring behavioral patterns, all extracted online. Implements a co-evolutionary loop where the policy updates the memory and the memory-conditioned self-teacher supervises the policy. Achieves significant performance gains over SDPO, 提出程序记忆蒸馏(PMD)方法,旨在解决现有强化学习验证奖励(RLVR)中跨回合过程信息丢失的问题。 构建三级抽象程序记忆(原始轨迹、自我反思策略、高阶行为模式),通过在线自监督机制将经验内化为模型权重。 采用“共同进化”设计原则,策略生成更新记忆,记忆指导策略训练,实现无需外部记忆的推理效率。 在Qwen3-8B和OLMo3-Instruct-7B上显著超越SDPO基线,SCIKNOWEVAL提升3.8-5.5%,LIVECODEBENCH提升7.9-13.6%。 消融实验证明共同进化的必要性,冻结记忆或策略任一环节均导致性能下降超过10%。

65
Hot 热度
75
Quality 质量
70
Impact 影响力

Analysis 深度分析

TL;DR

  • Introduces Procedural Memory Distillation (PMD), a method that captures cross-episode procedural signals often lost in standard RLVR and self-distillation approaches.
  • Organizes memory into three abstraction levels: raw trajectories, self-reflected strategies, and recurring behavioral patterns, all extracted online.
  • Implements a co-evolutionary loop where the policy updates the memory and the memory-conditioned self-teacher supervises the policy.
  • Achieves significant performance gains over SDPO, improving by 3.8-5.5% on SCIKNOWEVAL and 7.9-13.6% on LIVECODEBENCH.
  • Results in a memory-free model at inference time, as procedural knowledge is fully distilled into the policy weights.

Why It Matters

This approach addresses a critical limitation in current reinforcement learning frameworks for LLMs: the inability to retain and reuse rich procedural information across multiple training episodes. By converting transient rollout data into persistent, distilled knowledge, PMD enables models to learn from their own historical failures and successes more effectively, leading to substantial improvements in reasoning and coding benchmarks without increasing inference overhead.

Technical Details

  • Methodology: PMD converts cross-episode signals into reusable procedural memory, which acts as a training scaffold. This memory is distilled into the policy's weights, allowing for a memory-free inference state.
  • Memory Abstraction Levels: The memory structure operates at three levels: (1) raw trajectories, (2) self-reflected strategies and lessons, and (3) higher-level behavioral patterns that recur across different problems.
  • Co-Evolution Mechanism: The core design involves a feedback loop where the policy generates rollouts that update the memory, while a memory-conditioned self-teacher uses this accumulated experience to supervise the student model during its own rollouts.
  • Empirical Validation: Tested on Qwen3-8B and OLMo3-Instruct-7B, showing superior performance compared to SDPO. Ablation studies indicate that freezing either the memory or the policy component reduces performance by more than 10%, highlighting the necessity of the co-evolutionary process.

Industry Insight

  • Efficiency in Training: The ability to distill complex procedural knowledge into static weights suggests a pathway to more efficient training loops that do not require maintaining large external memory buffers during inference, reducing deployment complexity.
  • Focus on Process over Outcome: Moving beyond simple reward verification to capturing "how" solutions were derived (procedural memory) could become a standard practice for improving model robustness in complex reasoning tasks like coding and scientific evaluation.
  • Importance of Co-Evolution: The significant drop in performance when decoupling memory and policy updates underscores the need for integrated training architectures where knowledge retention and policy optimization are mutually reinforcing rather than sequential steps.

TL;DR

  • 提出程序记忆蒸馏(PMD)方法,旨在解决现有强化学习验证奖励(RLVR)中跨回合过程信息丢失的问题。
  • 构建三级抽象程序记忆(原始轨迹、自我反思策略、高阶行为模式),通过在线自监督机制将经验内化为模型权重。
  • 采用“共同进化”设计原则,策略生成更新记忆,记忆指导策略训练,实现无需外部记忆的推理效率。
  • 在Qwen3-8B和OLMo3-Instruct-7B上显著超越SDPO基线,SCIKNOWEVAL提升3.8-5.5%,LIVECODEBENCH提升7.9-13.6%。
  • 消融实验证明共同进化的必要性,冻结记忆或策略任一环节均导致性能下降超过10%。

为什么值得看

本文针对大模型强化学习中“遗忘”过程性知识的痛点,提出了将动态经验转化为静态权重的有效路径。其“共同进化”框架为提升模型在复杂推理任务中的自我改进能力提供了新的理论视角和技术范式。

技术解析

  • 核心机制:PMD不依赖episode-level的单一信号,而是从模型自身的rollout中提取跨回合的过程性信息,将其转化为可复用的程序记忆,并在训练期间蒸馏到策略网络的参数中。
  • 记忆结构:程序记忆分为三个抽象层级:1) 原始轨迹数据;2) 经过自我反思生成的策略与教训;3) 跨问题重复出现的高阶行为模式。所有信息均在线提取。
  • 训练流程:引入一个“记忆条件的自教师”(memory-conditioned self-teacher),利用累积的经验对当前策略生成的rollout进行监督,促使学生模型逐步内化程序知识。
  • 共同进化原理:策略生成用于更新记忆的rollout,而记忆又塑造了用于更新策略的监督信号,两者相互促进。实验显示,若固定记忆或策略,性能均大幅落后于联合优化版本。

行业启示

  • 从即时反馈到长期记忆:RLHF/RLVR领域需关注如何利用历史交互数据中的过程性知识,避免每次训练都从零开始探索,通过内化经验提升样本效率。
  • 无记忆推理的重要性:PMD证明了复杂的在线学习机制可以在推理阶段简化为标准的权重前向传播,平衡了训练时的丰富性与部署时的轻量级需求。
  • 自监督反思的价值:利用模型自身的反思结果作为监督信号,结合多层级的抽象提取,是提升大模型在代码和科学推理等复杂任务中鲁棒性的关键方向。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

LLM 大模型 Training 训练 Research 科学研究