# Salesforce AI 研究：弱模型智能体模仿 Gemini 反而退化，on-policy 修正更有效

- 来源：Rohan Paul (@rohanpaul_ai)
- 发布时间：2026-09-22 04:02
- AIHOT 分数：55
- AIHOT 链接：https://aihot.news/items/cmubom7qi04k9ro4v1qbrpvml
- 原文链接：https://x.com/rohanpaul_ai/status/2102126261279810001

## AI 摘要

Salesforce AI 用 Qwen3-Coder-30B-A3B 在 7 项企业任务上测试发现，把 Qwen 用 Gemini 完整轨迹微调后，成功率从 78.0% 降到 63.1%，7 项任务全部下滑，原因是 Qwen 照搬了 Gemini 的规划方式，与围绕它原有行为搭建的智能体设置不再匹配。

## 正文

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.

The weaker model did not need to think like Gemini; it needed Gemini to correct the places where its own approach failed.

Salesforce tested this across 7 enterprise tasks with Qwen3-Coder-30B-A3B.

Tuning the agent setup around Qwen raised average success from 29.2% to 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.

Instead, fine-tuning Qwen with complete Gemini runs dropped success to 63.1%, with performance falling on all 7 tasks.

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.

The fix was to keep Qwen’s own failed runs and let Gemini correct only the step where Qwen went wrong.

That pushed success to 79.7% without losing the earlier gains.

the lesson: tune the model without breaking the agent setup already working around it.