Can a 7B Model Beat Frontier LLMs at Auditing Move? Inside SuiAudit Pure17B 模型能贏過前沿 LLM 嗎?SuiAudit Pure1 的做法

Can a 7B Model Beat Frontier LLMs at Auditing Move? Inside SuiAudit Pure17B 模型能贏過前沿 LLM 嗎?SuiAudit Pure1 的做法

7B 模型能贏過前沿 LLM 嗎?SuiAudit Pure1 的做法
In short摘要

How SuiAudit Pure1, a 7B model, reached 31.4% accuracy auditing Sui Move — ahead of Claude Opus 4.5 and GPT-5.2 — by curating hard negatives.SuiAudit Pure1 這個 7B 模型如何靠困難負樣本微調,在 Sui Move 稽核基準上取得 31.4% 準確率,領先 Claude Opus 4.5 與 GPT-5.2。

The Question Behind the Paper

Our team at SuiAudit Lab has published a research paper, Can Small Models Beat Frontier LLMs? Introducing SuiAudit Pure1: Specialized Auditing via Extreme Low-Resource Fine-Tuning, on TechRxiv (DOI: 10.36227/techrxiv.177204934.48017152/v1). I was a key contributor alongside Chia-Lin Chen and Fang-Yu Lien; the resulting product, SuiAudit, is live at suiaudit.wal.app.

The question in the title is the honest version of what we were testing. If a general-purpose frontier model already writes and reads code well, what is left for a 7B-parameter model specialized on one language? Our answer is that the gap frontier models leave on Sui Move is not a gap in capability. It is a gap in what they were trained to expect.

Solidity Bias Is a Data Problem, Not a Scale Problem

General-purpose LLMs auditing Sui Move are plagued by what we call “Solidity Bias”: they mistakenly apply Ethereum vulnerabilities to Sui’s object-centric architecture.

My reading is that this is a wrong prior rather than a shortfall of capability, and that framing decides the whole approach: the fix is data that contradicts the prior.

So the dataset is the center of the work. It blends positive vulnerability samples, categorized A-F, with adversarial “hard negative” examples: safe code patterns engineered to trip up broader LLMs. The hard negatives are the part that matters. Training only on real vulnerabilities teaches a model to find something; training against code that looks vulnerable under Solidity assumptions but is safe under Sui’s teaches it when to say nothing.

The Pipeline, and How the Numbers Were Obtained

Pure1 is a 7B-parameter model built through a streamlined fine-tuning pipeline: Supervised Fine-Tuning (SFT) for domain mastery, followed by Direct Preference Optimization (DPO) enhanced with Chain-of-Thought reasoning.

Evaluation is on our own Sui-specific benchmark, where Pure1 reached 31.4% accuracy, ahead of Claude Opus 4.5 (26.6%) and GPT-5.2 (17.2%) — a margin of +4.7% over the strongest generalist model.

Pure1 is built on Mistral-7B-Instruct-v0.2 and quantized to 4-bit NF4, which is what lets it fit on a single NVIDIA T4 (16GB) or RTX 3060 (12GB). It runs fully offline, preserving developer privacy while delivering real-time insights via VS Code and Chrome extensions. We have open-sourced the core infrastructure on GitHub and shared the weights on Kaggle.

Where It Has Been Used

Pure1 is deployed in production.

The work also took 2nd Place at the 2025 Sui Hackathon Mandarin and 3rd at the Sui Taipei Dev Hackathon.

What I Take Away

The result I care about is narrow and, I think, generalizable: on a task where the dominant failure is a transferred prior rather than missing capability, curating the negatives beat scaling the model. That is a cheap lever, and it is available to anyone working in a language the frontier labs have little data for.

If you’re building on Sui or interested in AI-driven security, the full paper and the live tool are both linked above.

論文背後的問題

我們 SuiAudit Lab 團隊在 TechRxiv 上發表了一篇研究論文,Can Small Models Beat Frontier LLMs? Introducing SuiAudit Pure1: Specialized Auditing via Extreme Low-Resource Fine-Tuning(DOI: 10.36227/techrxiv.177204934.48017152/v1)。我與 Chia-Lin Chen、Fang-Yu Lien 同為主要貢獻者;由此而生的產品 SuiAudit 已上線於 suiaudit.wal.app

標題裡的問題就是我們真正在測的東西的誠實版本。如果一個通用的前沿模型本來就很會讀、很會寫程式碼,那一個只專精單一語言、7B 參數的模型還剩下什麼空間?我們的答案是:前沿模型在 Sui Move 上留下的落差,不是能力的落差,而是它們被訓練成「預期什麼」的落差。

Solidity 偏誤是資料問題,不是規模問題

通用 LLM 在稽核 Sui Move 時,深受我們稱為「Solidity 偏誤」的問題所苦:它們會把以太坊的漏洞樣態錯誤地套到 Sui 以物件為中心的架構上。

我的解讀是,這是先驗錯了,而不是能力不足;而這個框架決定了整套做法:解法是用與該先驗相牴觸的資料。

所以資料集才是這份工作的核心。它把分類為 A 到 F 的正向漏洞樣本,和對抗性的「困難負樣本」混在一起:那些刻意設計來絆倒較廣泛 LLM 的安全程式碼樣態。困難負樣本才是關鍵。只拿真實漏洞訓練,教會模型的是把東西找出來;拿那些在 Solidity 假設下看起來有漏洞、在 Sui 的規則下卻安全的程式碼來訓練,教會它的是什麼時候該閉嘴。

訓練管線,以及數字怎麼來的

Pure1 是一個 7B 參數模型,透過精簡的微調管線建成:先以監督式微調(SFT)取得領域掌握度,再用結合思維鏈推理的直接偏好優化(DPO)加強。

評測在我們自己的 Sui 專屬基準上進行,Pure1 取得 31.4% 的準確率,領先 Claude Opus 4.5 的 26.6% 與 GPT-5.2 的 17.2%——比最強的通用模型高出 4.7 個百分點。

Pure1 以 Mistral-7B-Instruct-v0.2 為基礎模型,量化為 4-bit NF4,因此塞得進單張 NVIDIA T4(16GB)或 RTX 3060(12GB)。它可完全離線執行,在保護開發者隱私的同時,透過 VS Code 與 Chrome 擴充功能提供即時洞察。我們已把核心基礎設施開源在 GitHub 上,並在 Kaggle 分享權重。

它被用在哪裡

Pure1 已部署於正式環境。

這份工作也拿下 2025 Sui Hackathon Mandarin 第二名,以及 Sui Taipei Dev Hackathon 第三名。

我的收穫

我在意的結論很窄,但我認為可以推廣:在一個主要失敗來自「先驗被錯誤遷移」而非「能力不足」的任務上,整理負樣本勝過放大模型。那是便宜的槓桿,而且任何在前沿實驗室資料稀少的語言上工作的人都拿得到。

如果你正在 Sui 上開發,或對 AI 驅動的安全有興趣,完整論文與線上工具的連結都在上面。