A $500 RL fine-tune of a 9B open model beat frontier models on catalog review
A $500 RL fine-tune of a 9B open model beat frontier models on catalog review
仅需 500 美元：通过强化学习微调 9B 开源模型，在目录审核任务中超越前沿模型
Cost vs. quality on catalog integrity The article in one picture: on the same catalog-review workflow, with the same tools, images, and scorer, our GRPO fine-tune of a 9B open-source model (pink) beats every frontier configuration we tested, at $0.50 per 1,000 listings: 40× cheaper than the least expensive frontier setup and ~340× cheaper than the most expensive.
目录完整性的成本与质量 用一张图概括本文：在相同的目录审核工作流中，使用相同的工具、图像和评分器，我们通过 GRPO 微调的 9B 开源模型（粉色）击败了我们测试的所有前沿模型配置。其成本仅为每 1,000 条目 0.50 美元：比最便宜的前沿模型方案便宜 40 倍，比最昂贵的方案便宜约 340 倍。
Part I: Everyone asked the same question Since ChatGPT launched in 2022, business leaders have been asking the same question: what can AI do for us? The answer began with low-risk tasks: summarizing documents, drafting emails, producing first drafts that a human would edit. It quickly moved into higher-value cognitive work, such as software development and content generation, and grew into more ambitious projects, like attempts to build an AI company brain, a system connected to internal knowledge, data, and tools that could coordinate work and eventually operate parts of the business autonomously.
第一部分：每个人都在问同一个问题 自 2022 年 ChatGPT 发布以来，商业领袖们一直在问同一个问题：人工智能能为我们做什么？答案最初始于低风险任务：总结文档、起草电子邮件、生成供人工编辑的初稿。随后迅速转向更高价值的认知工作，如软件开发和内容生成，并发展为更宏大的项目，例如尝试构建“AI 公司大脑”——一个连接内部知识、数据和工具的系统，旨在协调工作并最终自主运营部分业务。
While a lot of time, energy and tokens have been invested in AI adoption, measurable outcomes have barely been achieved at scale. However, some companies embraced being AI-first and saw enormous gains in productivity, revenue, and cost, while others lagged behind or failed to change their organizations enough to reach high ROI. Recent data from corporate expense management platform Ramp reveals a stark contrast in performance: the top quartile of companies investing in AI saw their revenue more than double between November 2022 and December 2025, while businesses with zero AI expenditure experienced a mere 15% increase.
尽管在 AI 应用上投入了大量的时间、精力和 Token，但大规模的可衡量成果却寥寥无几。然而，一些拥抱“AI 优先”战略的公司在生产力、收入和成本方面获得了巨大收益，而另一些公司则落后了，或者未能通过足够的组织变革来实现高投资回报率（ROI）。企业费用管理平台 Ramp 的最新数据显示了绩效上的鲜明对比：在 2022 年 11 月至 2025 年 12 月期间，AI 投资额排名前 25% 的公司收入翻了一番以上，而 AI 支出为零的企业仅增长了 15%。
Top AI adopters outperform across industries Same three years, same economy: in Ramp’s data across its customer base, the heaviest AI adopters more than doubled revenue while businesses spending nothing on AI grew about 15% (indexed, Nov 2022 = 100; curve drawn from the reported endpoints). The rest of this article is about what the AI-heavy group actually did. There are many reasons why AI has done wonders for some companies while others have struggled to see the return on their investment, but research primarily points in five directions.
AI 领跑者在各行业中表现优异 同样的三年，同样的经济环境：在 Ramp 的客户群数据中，AI 应用最深入的公司收入翻了一番以上，而未在 AI 上投入任何资金的企业仅增长了约 15%（以 2022 年 11 月为基准 100；曲线根据报告的终点绘制）。本文的其余部分将探讨这些 AI 领跑者究竟做了什么。AI 为一些公司创造了奇迹，而另一些公司却难以看到投资回报，原因有很多，但研究主要指向五个方向。
01 Redesign the process, not just the task Becoming AI-first means rethinking how the work is structured, not dropping a model into a workflow built around people: what gets approved, who reviews what, and which handoffs still need a human. Where the process stays untouched, legacy bottlenecks absorb the productivity gains before they reach the P&L. In McKinsey’s 2025 survey of organizations using gen AI, workflow redesign was the attribute most correlated with EBIT impact, and only 21% of them had redesigned any workflow at all.
01 重塑流程，而不仅仅是任务 成为“AI 优先”意味着重新思考工作结构，而不是简单地将模型塞进以人为中心的工作流中：需要重新审视什么环节需要审批、谁来审核什么、以及哪些交接环节仍然需要人工介入。如果流程保持不变，遗留的瓶颈会在生产力提升转化为利润（P&L）之前将其消耗殆尽。在麦肯锡 2025 年针对使用生成式 AI 的组织调查中，工作流重塑是与息税前利润（EBIT）影响相关性最高的因素，但仅有 21% 的组织真正重塑了任何工作流。
02 Incentivize experimentation Models, tooling and best practices change weekly, so last quarter’s setup is rarely still the right one. That only gets picked up if people are rewarded for trying things and reporting what failed, not just for shipping. Technical teams are the natural place to start, since they see the same problems recur across functions and can tell which of them a model can actually take over.
02 激励实验 模型、工具和最佳实践每周都在变化，因此上个季度的配置很少依然适用。只有当人们因尝试新事物并报告失败（而不仅仅是因交付成果）而获得奖励时，这种迭代才能实现。技术团队是最好的切入点，因为他们能看到跨职能部门中反复出现的问题，并能判断哪些问题是模型真正可以接管的。
03 Provide tailored business context Prompt engineering and retrieval can inject business context at call time, but doing it well is its own engineering program: getting to the data, enforcing access controls on what each request may see, building retrieval that surfaces the right evidence, and managing a context window that models use unevenly as it grows.
03 提供定制化的业务背景 提示词工程（Prompt Engineering）和检索增强可以在调用时注入业务背景，但要做好这一点本身就是一项工程挑战：获取数据、对每个请求的可见内容实施访问控制、构建能够呈现正确证据的检索系统，以及管理随着增长而导致模型使用效率不均的上下文窗口。
04 Measure usage and impact Every AI line item eventually meets the CFO question: what did this change, and was it worth it? In most deployments, nobody can answer it: there is no infrastructure to track the model’s performance, decision costs, or impact on efficiency, and self-reported time savings are often inaccurate. Without a scored evaluation on your own data, a “vibe evaluation” is the ceiling of what you can claim, and a hard budget to defend.
04 衡量使用情况与影响 每一项 AI 支出最终都会面临 CFO 的质询：这带来了什么改变？值得吗？在大多数部署中，没人能回答这个问题：因为缺乏跟踪模型性能、决策成本或效率影响的基础设施，且人工报告的时间节省往往不准确。如果没有基于自身数据的评分评估，所谓的“感觉评估”（vibe evaluation）就是你所能宣称的上限，且很难为预算辩护。
05 Set clear business goals within the AI budget AI brought a pricing model most companies were not used to. Paying per token instead of per seat makes costs scale with usage, which makes it hard to lay out a cost plan or estimate the capital efficiency gains for internal workloads. Uber went through its annual engineering budget in four months, and Microsoft cancelled most of its Claude licenses to bring costs back under control. Today’s prices also understate the problem, since most AI labs are subsidizing token costs to capture market share, and frontier model prices are expected to rise.
05 在 AI 预算内设定明确的业务目标 AI 带来了一种大多数公司不习惯的定价模式。按 Token 付费而非按席位付费，意味着成本随使用量扩展，这使得制定成本计划或估算内部工作负载的资本效率收益变得困难。Uber 在四个月内就花光了年度工程预算，微软则取消了大部分 Claude 许可证以控制成本。目前的价格还掩盖了问题，因为大多数 AI 实验室为了抢占市场份额正在补贴 Token 成本，而前沿模型的价格预计将会上涨。
In this article we give a detailed overview of the deployment technique the winning group keeps converging on: fine-tuning open-source models with reinforcement learning. We cover how it addresses the last three challenges above, and how it turns knowledge only your organization has (namely data, tools, and processes) into a model no vendor API can match at a fraction of the cost.
在本文中，我们将详细概述获胜群体不断趋向的部署技术：利用强化学习微调开源模型。我们将探讨它如何解决上述最后三个挑战，以及如何将只有贵组织才拥有的知识（即数据、工具和流程）转化为任何供应商 API 都无法比拟的模型，且成本仅为后者的一小部分。
TL;DR
- 2.2×: Revenue growth of the top quartile of AI spenders between November 2022 and December 2025 in Ramp’s data. Companies with zero AI spend grew about 15% over the same three years, in the same economy: the heavy adopters grew eight times as much.
- 1: Playbook the winners converge on: an open-source model, proprietary task data, and reinforcement learning against a scored copy of the workflow. Bridgewater’s trained model makes ~30% fewer mistakes than the best frontier model, Harvey’s legal agent beats GPT-5.5 and Claude Opus 4.8 on its own rubrics, and Intercom’s Fin Apex resolves more support issues at lower cost.
- 87.3%: Share of the maximum achievable score our GRPO-trained 9B open-source model reached on catalog review, vs 76.9% for the best frontier configuration: a 13.5% relative improvement over the frontier, and 36% over its own untrained base (64.2%). The five frontier models, even with optimized prompts, plateaued within a tenth of a point of each other; the trained specialist cleared that ceiling.
- 68×: Cost advantage per…
简要总结
- 2.2 倍：根据 Ramp 的数据，2022 年 11 月至 2025 年 12 月期间，AI 支出排名前 25% 的公司收入增长倍数。在同样的三年和经济环境下，AI 支出为零的公司仅增长了约 15%：重度 AI 应用者的增长速度是前者的 8 倍。
- 1 套方案：获胜者趋向的策略：使用开源模型、专有任务数据，并针对工作流的评分副本进行强化学习。Bridgewater 训练的模型比最好的前沿模型错误率低约 30%，Harvey 的法律智能体在自身评估标准下击败了 GPT-5.5 和 Claude Opus 4.8，Intercom 的 Fin Apex 以更低的成本解决了更多的支持问题。
- 87.3%：我们通过 GRPO 训练的 9B 开源模型在目录审核中达到的最高可实现分数占比，而最好的前沿模型配置仅为 76.9%：相对于前沿模型提升了 13.5%，相对于其未经训练的基准模型（64.2%）提升了 36%。五款前沿模型即使经过提示词优化，得分也都在 0.1 分的范围内徘徊；而经过训练的专用模型突破了这一上限。
- 68 倍：每…的成本优势