Salesforce研究：微调Agent模型需保留原有工作流适配
New Salesforce AI research finds, once you tune an agent’s prompts, tools, and w…
Agent 开发者必读，研究用数据证明盲目“蒸馏”强模型会破坏原有工作流，针对性修正才是提效关键。
New Salesforce AI research finds, once you tune an agent’s prompts, tools, and workflow around a weaker model, copying a stronger model can make it worse; targeted fixes to its own failures work better.
新的 Salesforce AI 研究发现，一旦你围绕较弱的模型调整了智能体的提示词、工具和流程，复制一个更强的模型反而会使表现变差；针对其自身失败的定向修复效果更好。
The weaker model did not need to think like Gemini; it needed Gemini to correct the places where its own approach failed.
较弱的模型不需要像 Gemini 那样思考；它需要 Gemini 来纠正其自身方法失败的地方。
Salesforce tested this across 7 enterprise tasks with Qwen3-Coder-30B-A3B.
Salesforce 使用 Qwen3-Coder-30B-A3B 在 7 个企业任务中测试了这一点。
Tuning the agent setup around Qwen raised average success from 29.2% to 78.0%.
围绕 Qwen 调整智能体设置后，平均成功率从 29.2% 提升至 78.0%。
A stronger Gemini model reached 93.6% in that same setup, so teaching Qwen to copy Gemini looked like the obvious next move.
在该相同设置下，更强的 Gemini 模型达到了 93.6% 的成功率，因此让 Qwen 学习复制 Gemini 看起来是显而易见的下一步。
Instead, fine-tuning Qwen with complete Gemini runs dropped success to 63.1%, with performance falling on all 7 tasks.
相反，使用完整的 Gemini 运行记录对 Qwen 进行微调，导致成功率降至 63.1%，且在所有 7 个任务上的性能均下降。
Qwen learned useful knowledge, but it also copied Gemini’s way of planning, which no longer fit the setup built around Qwen’s original behavior.
Qwen 学到了有用的知识，但也复制了 Gemini 的规划方式，这不再适合围绕 Qwen 原始行为构建的设置。
The fix was to keep Qwen’s own failed runs and let Gemini correct only the step where Qwen went wrong.
解决方案是保留 Qwen 自身的失败运行记录，并让 Gemini 仅纠正 Qwen 出错的那一步骤。
That pushed success to 79.7% without losing the earlier gains.
这使得成功率提升至 79.7%，且未丧失之前的成果。
the lesson: tune the model without breaking the agent setup already working around it.
教训：调整模型时不要破坏已围绕其工作正常的智能体设置。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力