NVIDIA Introduces SoL-Pi: Auto-Research Loops That Cut Coding Agent Token Traffic by Up to 49%
https://www.marktechpost.com/2026/09/21/nvidia-researchers-have-released-sol-pi/📌 【NVIDIA】讓 AI 自己優化 Coding Agent，Token 流量砍近半
TL;DR：NVIDIA、NTU、MIT 團隊用自動研究迴圈找出 harness 優化技巧，讓開源 Pi coding agent 省下最多 49% token 流量。
多數省 token 的招數都在降低「每個 token 的成本」：更快的 kernel、量化、換用更便宜的模型。NVIDIA 團隊這次反過來問：能不能直接減少任務要用掉的 token 數量？答案是讓另一個 AI 去自動搜尋 coding agent harness 的優化空間。
🤔 Coding agent 越跑越久，context 越塞越滿
Coding agent 現在一跑就是好幾個小時，每一次編輯、測試執行、log 讀取都會回寫進模型的 context。Harness 是負責處理 tool call、context、observation 與委派（delegation）的那一層，手動調校速度慢，而且各部分彼此耦合，改一處可能把成本推到後面的步驟。Meta-Harness 之類的系統嘗試自動化這件事，但先前有研究指出，演化出來的 harness 容易過度擬合搜尋時用的任務，換到沒看過的任務只有邊際效益。
🧩 怎麼搜：一個 AI 觀察另一個 AI 的執行軌跡
NVIDIA、NTU、MIT 團隊釋出的 SoL-Pi，是給開源 Pi coding agent 用的 4 個效率機制，全部由一個「研究 AI」透過在 harness 層跑自動研究迴圈找出來。流程是：研究 AI 觀察一個跑基礎版 Pi 的獨立 agent 產生的執行軌跡，據此提出 harness 改動並實際測試。每一次搜尋都是一個獨立、可拋棄的迴圈，遵循 autoresearch cycle，並額外加上 Ralph Loop 實作步驟與一位獨立審查員。
驗收規則在搜尋開始前就固定下來，優化器本身不能修改規則：每一項能力指標都必須落在預先宣告的容忍範圍內，候選方案還必須至少改善一項效率指標。EdgeBench 的 51 個公開任務中，11 個用來對凍結後的候選方案做單向驗收，40 個用於最終評估；held-out 的結果不會回饋進搜尋過程，避免自我強化的過擬合。
📊 實測：token 少了近一半，分數幾乎沒掉
整套機制先在 GPT-5.6 Sol 上搭建完成，再直接搬到 Opus 5 上使用，中間沒有針對 Opus 5 再做搜尋：
| 模型 | 保留分數比例 | Token 流量減少 | API 成本減少 |
|---|---|---|---|
| Opus 5 | 94.3% | 44.7% | 33.5% |
| GPT-5.6 Sol | 93.7% | 49.0% | 33.2% |
另外還有一個「Performance point」設定，只挑各後端表現最好的單一機制（GPT-5.6 Sol 用 ObservationPack、Opus 5 用 Action Fusion），能讓分數比原版 Pi 分別高出 5.3% 與 12.8%。
💡 省下來的錢，是拿 cache-write 換的
在 GPT-5.6 Sol 上，完整堆疊反而讓 cache-write 流量從 0.0141B 增加到 0.0316B token，但總成本仍然下降，從 1,339 美元降到 894 美元。論文估算，相較原生 Codex 與 Claude Code harness，每小時可省下 8.75 到 13.50 美元；相較原版 Pi，則是 4.36 到 5.71 美元。
⚠️ 跨模型遷移還只是初步結果
研究團隊自己也強調，跨模型遷移的結果屬於初步（preliminary）——這些機制在 Opus 5 上觸發的頻率較低，可能是因為搜尋過程只用了 GPT-5.6 Sol 的執行軌跡，尚未針對 Opus 5 重新搜尋。
🎯 實務啟示
SoL-Pi 已經以 MIT 授權在 NVlabs 的 GitHub 上開源，可以直接套用在未經修改的 Pi 版本上，官方測試環境是 Pi 0.85.1 搭配 Node.js 22.19 以上。對正在跑長時間 agent 任務、被 context 堆積吃掉成本的團隊來說，這提供了一條不用換模型、不用改 prompt，只從 harness 層面省錢的路徑，而且驗收機制的設計（固定規則、held-out 評估）也值得參考，用來避免自己手動調校 harness 時掉入過擬合的陷阱。
🔗 來源
- 標題：NVIDIA Introduces SoL-Pi: Auto-Research Loops That Cut Coding Agent Token Traffic by Up to 49%
- 作者／機構：Asif Razzaq @ MarkTechPost；研究團隊來自 NVIDIA、NTU、MIT
- 連結：https://www.marktechpost.com/2026/09/21/nvidia-researchers-have-released-sol-pi/
#NVIDIA #CodingAgent #LLMAgents #TokenEfficiency #OpenSource #AIResearch #HarnessOptimization #AgenticAI #MachineLearning #EdgeBench
原始資料 MarkTechPost · 收集於 2026-09-22
摘要原文
Coding agents now run for hours, not minutes. Every edit, test run and log read goes back into the model’s context. A team of researchers from NVIDIA, NTU and MIT have released SoL-Pi , a set of 4 efficiency mechanisms for the open-source Pi coding agent. An AI found these mechanisms by running auto-research loops at the harness layer. On the 51-task EdgeBench evaluation, SoL-Pi cuts recorded token traffic by 44.7% to 49.0% versus Pi and cuts API cost by roughly 33%. Its scores stay close to Pi on both GPT-5.6 Sol and Opus 5. Is it deployable? Yes. SoL-Pi ships on GitHub under NVlabs as an MIT-licensed extension that runs on an unmodified Pi release. It is tested with Pi 0.85.1 and Node.js 22.19 or newer. Most efficiency work lowers the cost per token through faster kernels, quantization or cheaper models. SoL-Pi instead reduces how many tokens a task consumes. The harness is the layer that handles tool calls, context, observations and delegation. Tuning a harness by hand is slow, and its parts are coupled: a fix in one place can push cost into later steps. Meta-Harness and similar systems automate this work. However, a recent study found that evolved harnesses can overfit their search tasks and give only marginal gains on unseen ones. A research AI observes execution traces from a separate agent running base Pi. It then proposes harness changes and tests them. The search covered: Each search is a disposable, isolated loop. It follows the autoresearch cycle, extended with a Ralph Loop implementation step and an independent reviewer. Acceptance rules are fixed before the search starts, and the optimizer cannot change them. Every capability metric must stay within a predeclared tolerance. The candidate must also improve at least 1 efficiency metric. EdgeBench stays held out. Of its 51 public tasks, 11 are used for one-way acceptance of frozen candidates and 40 for final evaluation. Held-out results never feed back into the search. The full stack was built on GPT-5.6 Sol and moved to Opus 5 without further search. On Opus 5, it keeps 94.3% of Pi’s score while cutting token traffic by 44.7% and API cost by 33.5%. On GPT-5.6 Sol, it keeps 93.7% of Pi’s score with 49.0% fewer tokens and 33.2% lower cost. The Performance point uses the best single mechanism for each backend: ObservationPack on GPT-5.6 Sol and Action Fusion on Opus 5. It lifts scores 5.3% and 12.8% above Pi. On GPT-5.6 Sol, the full stack increases cache-write traffic from 0.0141 B to 0.0316 B tokens. Total cost still falls, from $1,339 to $894. The paper estimates hourly savings of $8.75 to $13.50 versus the native Codex and Claude Code harnesses, and $4.36 to $5.71 versus Pi. The research team calls the cross-model transfer preliminary. The mechanisms trigger less often on Opus 5, possibly because the search used only GPT-5.6 Sol trajectories. Check out the Paper , GitHub Repo and Technical Blog . All credit goes to the researcher of this project. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter . Wait! are you on telegram? now you can join us on telegram as well. Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us The post NVIDIA Introduces SoL-Pi: Auto-Research Loops That Cut Coding Agent Token Traffic by Up to 49% appeared first on MarkTechPost .
由 claude-code/sonnet 自動生成