Learning Hierarchical Procedural Memory for LLM Agents through Bayesian Selection and Contrastive Refinement
- ID: 0a89a060
- 原文链接: https://arxiv.org/abs/2512.18950
- PDF: https://arxiv.org/pdf/2512.18950
- 作者: Saman Forouzandeh, Wei Peng, Parham Moradi, Xinghuo Yu, Mahdi Jalili
- 日期 / 版本: Submitted on 22 Dec 2025
- 分类: agents
- 来源类型: paper
- arXiv 分类: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
- 标签: llm-agent, procedural-memory, planning, bayesian-selection, benchmarks, 2512-18950
- 质量评分: 4/5
- 抓取时间: 2026-07-27T04:25:26.452521+00:00
中文导读
论文提出 MACLA：冻结底层 LLM，把学习和适配放在外部层级 procedural memory 中摘要描述其从轨迹抽取可复用 procedure，用 Bayesian posterior 跟踪可靠性，通过 expected-utility 选择动作，并用成功/失败对比细化过程；在 ALFWorldWebShopTravelPlannerInterCodeSQL 四个基准上报告 78.1% 平均表现
为什么值得关注
MACLA 将适配能力放进外部层级过程记忆，用贝叶斯可靠性和成败对比来改进 Agent
这篇论文值得放进 AAIF 的原因，是它把 Agent 系统里的一个具体工程问题讲清楚：论文提出 MACLA：冻结底层 LLM，把学习和适配放在外部层级 procedural memory 中摘要描述其从轨迹抽取可复用 procedure，用 Bayesian posterior 跟踪可靠性，通过 expected-utility 选择动作，并用成功/失败对比细化过程；在 ALFWorldWebShopTravelPlannerInterCodeSQL 四个基准上报告 78.1% 平均表现。以下内容基于 arXiv 页面元数据、摘要与条目已有摘要整理，未补充论文摘要之外的实验细节。
关键信息
- 论文标题：Learning Hierarchical Procedural Memory for LLM Agents through Bayesian Selection and Contrastive Refinement
- 作者：Saman Forouzandeh, Wei Peng, Parham Moradi, Xinghuo Yu, Mahdi Jalili
- arXiv：https://arxiv.org/abs/2512.18950
- 发布时间 / 修订：Submitted on 22 Dec 2025
- arXiv 分类：Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
- arXiv 备注：Accepted at The 25th International Conference on Autonomous Agents and Multi-Agent Systems (AAMAS 2026). 21 pages including references, with 7 figures and 8 tables. Code is publicly available at the authors GitHub repository.
- 关联标签：llm-agent, procedural-memory, planning, bayesian-selection, benchmarks, 2512-18950
English Abstract
We present MACLA, a framework that decouples reasoning from learning by maintaining a frozen large language model while performing all adaptation in an external hierarchical procedural memory. MACLA extracts reusable procedures from trajectories, tracks reliability via Bayesian posteriors, selects actions through expected-utility scoring, and refines procedures by contrasting successes and failures. Across four benchmarks (ALFWorld, WebShop, TravelPlanner, InterCodeSQL), MACLA achieves 78.1 percent average performance, outperforming all baselines. On ALFWorld unseen tasks, MACLA reaches 90.3 percent with 3.1 percent positive generalization. The system constructs memory in 56 seconds, 2800 times faster than the state-of-the-art LLM parameter-training baseline, compressing 2851 trajectories into 187 procedures. Experimental results demonstrate that structured external memory with Bayesian selection and contrastive refinement enables sample-efficient, interpretable, and continually improving agents without LLM parameter updates.
English Summary
We present MACLA, a framework that decouples reasoning from learning by maintaining a frozen large language model while performing all adaptation in an external hierarchical procedural memory. MACLA extracts reusable procedures from trajectories, tracks reliability via Bayesian posteriors, selects actions through expected-utility scoring, and refines procedures by contrasting successes and failures. Across four benchmarks (ALFWorld, WebShop, TravelPlanner, InterCodeSQL), MACLA achieves 78.1 percent average performance, outperforming all baselines. On ALFWorld unseen tasks, MACLA reaches 90.3 percent with 3.1 percent positive generalization. The system constructs memory in 56 seconds, 2800 times faster than the state-of-the-art LLM parameter-training baseline, compressing 2851 trajectories into 187 procedures. Experimental results demonstrate that structured external m
Obsidian Notes
- 内容获取路径：优先尝试
opencli arxiv paper 2512.18950 -f json，本轮遭遇 arXiv API HTTP 429 后，改由 arXiv 页面元数据与摘要回填。 - 中文导读与价值判断均锚定在条目已有摘要、论文摘要、作者、日期与分类信息上。
- 现代站点生成器按
content/{entry.id}.md查找内容页；本文件写入 canonical content 目录，而不是openclaw/content/。