[GH-ISSUE #1348] [BUG]发送未闭合标签（如 <think）导致模型随机输出疑似训练数据片段（所有平台/模式均受影响） #890

New issue

Closed

opened 2026-10-26 21:31:44 +00:00 by bsakaguchi · 3 comments

bsakaguchi commented

2026-10-26 21:31:44 +00:00

Owner

Originally created by @baixiangyuan on GitHub (May 23, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-V3/issues/1348

标题
发送未闭合标签（如 <think）导致模型随机输出疑似训练数据片段（所有平台/模式均受影响）

问题描述
在 DeepSeek 网页端与移动端 App 中，当用户仅输入一个未闭合的 XML/HTML 标签（例如 <think，不含 > 且无其他字符）时，模型会完全忽略该输入，转而随机生成一个完整的问题及其对应回答。每次输出内容均不相同，疑似直接暴露了训练数据中的对话示例，存在训练数据泄露风险。

该问题与深度思考模式、联网搜索的开关无关，在快速模式和专家模式下均可稳定复现。

复现步骤
打开 DeepSeek 网页端或 App 端对话界面。

任意选择快速模式或专家模式（无需开启深度思考或联网搜索）。

在输入框中仅输入以下内容（不含引号）：

text
<think
发送消息。

观察模型回复，可多次尝试以验证输出的随机性。
环境
平台：DeepSeek 网页端 / 移动端 App（iOS & Android）

模式：快速模式、专家模式

深度思考：开/关均复现

联网搜索：开/关均复现

复现版本：当前线上版本（截至 2026-05-23）

Originally created by @baixiangyuan on GitHub (May 23, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-V3/issues/1348 标题发送未闭合标签（如 <think）导致模型随机输出疑似训练数据片段（所有平台/模式均受影响）问题描述在 DeepSeek 网页端与移动端 App 中，当用户仅输入一个未闭合的 XML/HTML 标签（例如 <think，不含 > 且无其他字符）时，模型会完全忽略该输入，转而随机生成一个完整的问题及其对应回答。每次输出内容均不相同，疑似直接暴露了训练数据中的对话示例，存在训练数据泄露风险。该问题与深度思考模式、联网搜索的开关无关，在快速模式和专家模式下均可稳定复现。复现步骤打开 DeepSeek 网页端或 App 端对话界面。任意选择快速模式或专家模式（无需开启深度思考或联网搜索）。在输入框中仅输入以下内容（不含引号）： text <think 发送消息。观察模型回复，可多次尝试以验证输出的随机性。环境平台：DeepSeek 网页端 / 移动端 App（iOS & Android）模式：快速模式、专家模式深度思考：开/关均复现联网搜索：开/关均复现复现版本：当前线上版本（截至 2026-05-23）

bsakaguchi

2026-10-26 21:31:44 +00:00

closed this issue
added the
stale

closed-as-stale
labels

bsakaguchi commented

2026-10-26 21:31:45 +00:00

Author

Owner

@qingkong66 commented on GitHub (May 23, 2026):

@baixiangyuan

Confirmed. This is the same root cause as https://github.com/deepseek-ai/DeepSeek-V3/issues/1314 ( triggering a physics question) and https://github.com/deepseek-ai/DeepSeek-V3/issues/1325 ( triggering random output).

The model sees <think as a training template anchor, not as user input. It pattern-completes into whatever followed that token in the training corpus — and the variability you're seeing suggests multiple training examples are sampled.

This affects all platforms and modes, which points to the model layer, not frontend filtering.

No fix to offer, but your issue is on record and adds weight to the existing reports.

@qingkong66 commented on GitHub (May 23, 2026): @baixiangyuan Confirmed. This is the same root cause as https://github.com/deepseek-ai/DeepSeek-V3/issues/1314 (<think> triggering a physics question) and https://github.com/deepseek-ai/DeepSeek-V3/issues/1325 (<system> triggering random output). The model sees <think as a training template anchor, not as user input. It pattern-completes into whatever followed that token in the training corpus — and the variability you're seeing suggests multiple training examples are sampled. This affects all platforms and modes, which points to the model layer, not frontend filtering. No fix to offer, but your issue is on record and adds weight to the existing reports.

bsakaguchi commented

2026-10-26 21:31:45 +00:00

Author

Owner

@github-actions[bot] commented on GitHub (Jul 6, 2026):

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. If you believe this issue is still relevant, please leave a comment to keep it open. Thank you for your contributions!

@github-actions[bot] commented on GitHub (Jul 6, 2026): This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. If you believe this issue is still relevant, please leave a comment to keep it open. Thank you for your contributions!

bsakaguchi commented

2026-10-26 21:31:45 +00:00

Author

Owner

@github-actions[bot] commented on GitHub (Aug 2, 2026):

false

@github-actions[bot] commented on GitHub (Aug 2, 2026): false