Distilling Feedback into Memory-as-a-Tool
- ID: 65ce651a
- 原文链接: https://arxiv.org/abs/2601.05960
- PDF: https://arxiv.org/pdf/2601.05960
- 作者: Víctor Gallego
- 日期 / 版本: Submitted on 9 Jan 2026 (v1), last revised 17 Mar 2026 (this version, v2)
- 分类: agents
- 来源类型: paper
- arXiv 分类: Computation and Language (cs.CL)
- 标签: llm-agent, memory, feedback, tool-use, inference-cost, 2601-05960
- 质量评分: 4/5
- 抓取时间: 2026-07-27T04:25:26.452521+00:00
中文导读
论文提出把推理时产生的 critique / feedback 转换为可检索 guideline 的 Memory-as-a-Tool 框架:Agent 通过文件式记忆和工具调用,在后续任务中复用这些反馈摘要称其在 Rubric Feedback Bench 上能快速接近 test-time refinement pipeline 的表现,同时显著降低推理成本
为什么值得关注
把一次性反馈蒸馏成可检索的 memory-as-a-tool,用外部文件记忆降低反复推理成本
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:论文提出把推理时产生的 critique / feedback 转换为可检索 guideline 的 Memory-as-a-Tool 框架:Agent 通过文件式记忆和工具调用,在后续任务中复用这些反馈摘要称其在 Rubric Feedback Bench 上能快速接近 test-time refinement pipeline 的表现,同时显著降低推理成本。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:Distilling Feedback into Memory-as-a-Tool
- 作者:Víctor Gallego
- arXiv:https://arxiv.org/abs/2601.05960
- 发布时间 / 修订:Submitted on 9 Jan 2026 (v1), last revised 17 Mar 2026 (this version, v2)
- arXiv 分类:Computation and Language (cs.CL)
- arXiv 备注:Code and data links listed on arXiv; journal reference: ICLR 2026 Workshop on Memory for LLM-Based Agentic Systems
- 关联标签:llm-agent, memory, feedback, tool-use, inference-cost, 2601-05960
English Abstract
We propose a framework that amortizes the cost of inference-time reasoning by converting transient critiques into retrievable guidelines, through a file-based memory system and agent-controlled tool calls. We evaluate this method on the Rubric Feedback Bench, a novel dataset for rubric-based learning. Experiments demonstrate that our augmented LLMs rapidly match the performance of test-time refinement pipelines while drastically reducing inference cost.
English Summary
We propose a framework that amortizes the cost of inference-time reasoning by converting transient critiques into retrievable guidelines, through a file-based memory system and agent-controlled tool calls. We evaluate this method on the Rubric Feedback Bench, a novel dataset for rubric-based learning. Experiments demonstrate that our augmented LLMs rapidly match the performance of test-time refinement pipelines while drastically reducing inference cost.
Obsidian Notes
- 内容获取路径:优先尝试
opencli arxiv paper 2601.05960 -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/。