SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness

ID: 47c17a5e
原文链接: https://arxiv.org/abs/2609.20519
作者: Haozhe Liu, Tian Ye, Sensen Gao, Qihang Cao, Yitong Li, Mingchen Zhuge, Duomin Wang, Ruihua Zhang, P...
日期: 2026-09-17
分类: coding
来源类型: paper
标签: agent-harness, token-efficiency, recursive-self-improvement, nvlabs, pi-agent, coding-agent
质量评分: 4/5
抓取时间: 2026-09-21T23:30Z

中文摘要

NVlabs arXiv:2609.20519 提出 SoL-Pi，给 Pi coding agent 加一层递归 auto-research harness：四个机制都默认关闭需 sol-pi.json 显式开启Action Fusion 把 edit/write 后紧跟 test/build/run 合并成一个 tool request 省一模型回合；Online Context Compact 子任务完成后判断是否压上下文，预计能 cover rewrite 成本才做；ObservationPack 把大输出归档本地后续只传 handle / 大小 / 短摘录，必要时按页召回原文；Evidence-Preserving Reducer 用低成本模型先读长 build/test log，但 compact receipt 的引用必须逐条匹配归档日志51 个 EdgeBench 任务上相对 Pi：recorded token traffic 降 44.749.0%，API cost 约降三分之一，相对 native Codex/Claude Code list-price 成本低 5054%Terminal-Bench 4 上 63 个 CPU-only 任务 Codex/Pi 各解出 18 个SoL-Pi 解出 15 个；总成本 $272.35 / $286.45 / $211.12代码 MIT，仓库 NVlabs/SoL-Pi

English Abstract

NVlabs arXiv:2609.20519 introduces SoL-Pi, a recursive auto-research harness layer wrapped around the Pi coding agent. Four mechanisms all off by default, all gated on a sol-pi.json config address distinct waste categories: Action Fusion collapses edit/write followed by test/build/run into a single tool request, saving a model round; Online Context Compact decides whether to compress context after a subtask finishes, only acting when expected savings exceed the rewrite cost; ObservationPack archives large outputs locally and forwards only handle / size / short excerpt, paging back raw text when needed; Evidence-Preserving Reducer lets a cheap model read long build/test logs first but requires compact-receipt references to match the archived log line by line....

为什么值得关注

NVlabs SoL-Pi 把 Pi coding agent 包一层 recursive harness：EdgeBench token -44.7~49% cost ~1/3，TB4 不输

Obsidian 证据摘要

来源: OpenClaw定时任务/AK-RSS-Digest（89源精选）/2026-09-21-AK-RSS-Digest（89源精选）.md + 对应 evidence-2026-09-21 文件