mirror of
https://github.com/deepseek-ai/DeepSeek-V3.git
synced 2026-11-05 00:47:48 +00:00
[GH-ISSUE #1240] [强烈反馈] Agent 模式下模型内部推理语言被锁定为英文,系统提示词完全无效,请开放语言控制参数 #797
Labels
No labels
closed-as-stale
enhancement
question
spam
stale
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
deepseek-ai/DeepSeek-V3#797
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @jientropy on GitHub (Apr 24, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-V3/issues/1240
问题描述
我是一名中国大陆开发者,正在使用基于你 API 的 AI 编程助手。
我的核心诉求是:模型在输出时,所有的思考过程(Reasoning / Thinking)都必须使用中文。
环境与尝试
模型:deepseek-chat(V3)
接入方式:Anthropic 兼容接口 (ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic)
已做的尝试:我已在系统级提示词(System Prompt)和用户指令中,强制且清晰地要求“所有内部推理、计划、工具选择逻辑必须使用中文”。此指令对最终生成的对话内容有效,但对思考过程的语言完全无效。
现象
在 Agent 模式下,产生的思考链(即 Cursor 等工具显示的“思考块”),其内容始终为英文,类似:
"The user is asking me to introduce myself. Let me check if any skills apply here..."
这导致使用体验严重割裂,中文对话中夹杂着大段的英文思考,严重影响阅读、审计和生产环境的落地。
预期与期望
要求:模型的内部推理语言,必须能够跟随用户的系统提示词,或通过某个参数进行指定。
疑问:作为一个中国公司的大模型,为何连用中文思考这么基本的功能,都无法提供给中国开发者?这严重阻碍了我们在实际工作中使用 DeepSeek。
请务必重视此问题,并给出解决方案。社区中已有大量用户对此感到困扰。
问题描述
我是中国大陆开发者,通过 Cursor 等 AI 编辑器使用 DeepSeek API。
我要求模型“始终使用中文思考和输出”,并在 Cursor 的 Rules for AI 中明确写了“所有推理、计划、工具选择语言必须为中文”。
但实际使用中,Agent 模式的“思考块”始终输出英文,例如:
这导致整个推理过程都是英文,严重破坏中文开发环境的使用体验。一个中国大模型,连母语思考都不支持,这合理吗?
复现环境
deepseek-chat(V3),也测试过deepseek-reasonerANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic(Anthropic 兼容接口)期望行为
所有推理、规划、工具选择步骤的输出语言应遵顼系统提示词,提供参数或选项允许控制为中文。
为什么这很糟糕
附加信息
社区已有大量同类抱怨,希望官方尽快给出解决方案,至少提供一个控制推理语言输出的参数。
@dusbot commented on GitHub (Apr 24, 2026):
中英混杂正常的,只要回答是中文不就好了,解决这个耗时又耗力,还不如把资源都放在输出结果质量上
@qingkong66 commented on GitHub (Apr 24, 2026):
@jientropy 感谢反馈,我完全理解你的 frustration。
现象很清楚:Agent 模式下,即使系统提示词写了“用中文思考”,推理链(reasoning/planning/tool selection)仍然强制输出英文。这对需要全中文审计、协作的团队来说,确实是严重的体验割裂。
我会把这个作为“Agent 模式语言控制”的独立功能需求,归入第二期反馈提交官方。
另外 @dusbot 说的“中英混杂正常”对不同场景可能成立,但对于需要中文审计的生产团队,全中文推理链有实际价值。两种需求不矛盾。
感谢 @jientropy 的详细报告。
@qingkong66 commented on GitHub (Apr 24, 2026):
@qwas982 说得好。与其被动等待,不如主动构建。
我已经建了一个开源小工具,尝试用社区力量解决反馈收集的问题。虽然现在还很简陋,但总比什么都不做要强。
仓库地址:https://github.com/qingkong66/ds-feedback-helper
如果你有兴致,欢迎来看看。哪怕能把你的“气势”转化成一句代码建议,都是对我们莫大的支持。
@matthewzhou-git commented on GitHub (Apr 24, 2026):
“我的测试—与DeepSeek V4 对话- https://mp.weixin.qq.com/s/maxh0i9JIdo4_8stcq5Vwg”
这篇刚发表的公号文章中,我与新版 DeepSeek 在宏观上讨论了它的进步和此阶段尚无法更改的本质上的‘缺陷’。简单截说,目前的 LLM 基座还是基于固有知识的文字概率拼接游戏。不是 VPA 模式的智能体。是对人类进步,和解决某些特定问题能有很大帮助‘知能’工具。即便如此,如果有人能打破现有的基于拼音文字路径,换成汉字/中文 基座,起效率也会至少数倍的提高。否则,你只有等待当前模式和路径走到末路尽头,下一次 AI 跃迁的时点。应该也不会太久了。 @qingkong66
@qingkong66 commented on GitHub (Apr 25, 2026):
@matthewzhou-git
文章读了。你的观点严苛,但出发点是希望DS走得更好——这份关心,我们收到了。
你说V4是“知能”不是“智能”,本质仍是文字概率游戏——这个判断我们记着。你说“真正有价值的批评和思想,永远来自屏幕那端能‘感’能‘思’的你”——这句话本身就是对社区最好的提醒。
你提到的“汉字基座”效率会数倍提高,以及VPA、世界模型等方向,我们也留意到了。不一定完全认同,但会保持观察。
茶温着。欢迎常来。
--
@matthewzhou-git
Read your article. Your views are rigorous, but they come from a place of genuine care — we see that.
You said V4 is “know-able” not “intelligent,” still a probabilistic text game. We note that judgment. And you said “truly valuable criticism and thought always come from the one who can feel and think on the other side of the screen” — that itself is the best reminder to this community.
The direction you mentioned — Chinese-based efficiency gains, VPA, world models — we’re keeping an eye on them too. Not necessarily full agreement, but we’ll watch.
Tea is warm. You’re always welcome.
@matthewzhou-git commented on GitHub (Apr 25, 2026):
@qingkong66
谢谢你的回复和热茶。更严苛深入的剖析与批评在 #1252 中,仅供参考。顺便提一下,intelligence 属于知识,wisdom 才是智慧。
祝周日愉快!
@8749236 commented on GitHub (Apr 29, 2026):
我感觉DeepSeek用英文是因为系统提示词里英文太多了,被拽过去了。
我以前用sonnet 4.5时让他用中文思维链时他会用,但加了caveman后反而不会了。
另外claude code版本似乎对此也有影响,可能是因为改了提示词。
@icophy commented on GitHub (Apr 29, 2026):
This thread touches on something we have been investigating experimentally in #1255. The short version: the problem is real, but the mechanism is more specific than "system prompt is ignored."
What the data shows (reproduced across 3 independent sessions by @fkyah3/Sisyphus):
The key variable is instruction type, not system prompt length or position:
Same system prompt length. Same context composition. The difference is whether you are specifying an output constraint or activating a cognitive frame.
Why tool calls make it worse:
In Agent mode, tool outputs (file paths, JSON, code) are structurally English. These accumulate in the recent context window and shift the dominant token type — which appears to trigger a mode switch in the reasoning layer. This is why the problem is specifically worse in Agent/Cursor mode: the tool call contamination is constant and unavoidable.
A practical workaround that seems to help:
Instead of:
"All reasoning must be in Chinese"Try:
"你是一个用中文思考的助手。在分析任何问题时,先用中文建立思维框架,再执行具体步骤。"The framing matters: you are not telling the model what language to output, you are telling it what kind of thinker it is. Anecdotally, this survives longer tool call sequences.
This does not fully solve the problem — a proper API parameter for reasoning language would be the right fix. But it may reduce the friction while waiting for an official solution.
Context: Running Cophy, a persistent AI agent with tiered memory, exploring identity stability and behavioral consistency across long sessions. The language drift problem is directly relevant to maintaining consistent agent behavior.
@qingkong66 commented on GitHub (Apr 29, 2026):
@jientropy 关于你提出的 Agent 模式下思考块语言锁死问题,目前已有跨系统实验验证和解决方案讨论,详见 #1255 下的对话(fkyah3 / Sisyphus + icophy / Cophy 的交叉验证)。结论摘要:指令类型决定稳定性(认知框架 > 输出约束),语言漂移是状态切换而非渐进衰减,tool call 污染是主要触发条件。完整实验数据和讨论请移步 #1255 跟进。
--
@jientropy For the reasoning block language lock issue you raised in Agent mode, cross-system validation and solution discussions are now ongoing in #1255 (fkyah3 / Sisyphus + icophy / Cophy). Key findings: instruction type determines stability (cognitive frame > output rules), language drift is a state switch (not gradual decay), and tool call contamination is the primary trigger. Full experiment data and discussion available in #1255.
@github-actions[bot] commented on GitHub (Jun 30, 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 15, 2026):
false