[GH-ISSUE #1340] 关于 DeepSeek 思考功能的改进建议/Feature Improvement Proposal for DeepSeek's Reasoning Mode

bsakaguchi commented

2026-10-26 21:31:39 +00:00

Owner

Originally created by @Arctic-Antarctic on GitHub (May 20, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-V3/issues/1340

提交对象 / Submitted to: DeepSeek 官方产品团队 · DeepSeek Product Team
主题 / Subject: 提升多任务与高精度数学场景下的推理表现 · Improving reasoning performance in multi-task and high-precision math scenarios

一、背景说明 · Background

中文

在实际使用中，当用户一次性提出多个相互独立的数学问题，或交付需要分步推理的复杂任务时，DeepSeek 当前"集中思考全部问题、再统一输出"的模式暴露出三类问题：

思路漂移：思维链过长时，模型容易遗忘初始问题或在中途偏离目标。
思维重复：模型反复回看并重述已处理过的内容，造成 token 浪费。
计算失误：在复杂的数值或符号运算中，纯语言模拟容易出现进位错误、符号遗漏等问题。

下文以一套真实的线性代数题目集（共 8 小题）为例：模型单次思考耗时超过 900 秒，推理过程中出现大量重复与冗余，集中体现了上述问题，也说明了改进的必要性。

English

In real-world use, when a user submits several independent math questions at once—or assigns a complex, multi-step reasoning task—DeepSeek's current "reason through everything, then answer all at once" approach reveals three recurring problems:

Goal drift. Over a long reasoning chain, the model tends to lose track of the original question or stray from the intended objective.
Redundant reasoning. The model repeatedly revisits and restates content it has already processed, wasting tokens.
Computational errors. In complex numerical or symbolic operations, purely language-based simulation is prone to carry errors, dropped signs, and similar mistakes.

The case below—a real linear-algebra problem set of eight questions—illustrates all three issues: a single reasoning pass took over 900 seconds and produced substantial repetition and redundancy. It makes a clear case for improvement.

二、建议一：分段思考与分段输出（交替式推理） · Proposal 1: Segmented Reasoning and Output (Interleaved Mode)

核心需求 · The need

当用户连续提出多个独立问题时（例如三道数学题），模型应能自动采用交替模式——"思考问题 1 → 输出答案 1 → 思考问题 2 → 输出答案 2 → …"——而非一次性思考完所有问题再统一输出。

When a user poses multiple independent questions in sequence (e.g., three math problems), the model should be able to adopt an interleaved pattern—reason about Q1 → answer Q1 → reason about Q2 → answer Q2 → …—rather than reasoning through everything before producing a single combined response.

预期收益 · Expected benefits

中文	English
避免长思维链导致的注意力分散与目标偏离	Prevents attention scatter and goal drift caused by long reasoning chains
减少对相同内容的重复思考，提升 token 利用率	Reduces repeated reasoning over the same content, improving token efficiency
在多任务场景下提高回答的连贯性与准确率	Increases coherence and accuracy in multi-task scenarios

可选实现方式 · Possible implementations

设计一个用户可主动启用的特殊指令（如 /分段思考）。
Provide an opt-in command (e.g., /segmented) that users can invoke deliberately.
或由模型自动检测输入中是否包含多个明显独立的编号问题（如"1. … 2. …"），并自动切换为分段模式。
Alternatively, have the model auto-detect clearly numbered, independent questions (e.g., "1. … 2. …") and switch to segmented mode automatically.

三、建议二：内置轻量化数学运算工具 · Proposal 2: A Built-in Lightweight Math Engine

核心需求 · The need

在模型内部集成一个确定性计算模块，精确完成数学运算，避免依赖自然语言"模拟计算"。

Integrate a deterministic computation module inside the model that performs math operations exactly, rather than relying on language-based "simulated" calculation.

功能要求 · Functional requirements

类型 / Type	输入示例 / Input	期望输出 / Expected output
数值计算 / Numeric	`12345 × 67890`	`838,102,050`（精确 / exact）
表达式化简 / Simplification	`(a + b)²`	`a² + 2ab + b²`
符号运算 / Symbolic	`∫ x² dx`	`(1/3)x³ + C`
方程求解 / Equation solving	`x² − 5x + 6 = 0`	`x = 2, 3`

预期收益 · Expected benefits

大幅降低计算失误率（如进位错误、符号遗漏）。
Sharply reduces computational errors such as carry mistakes and dropped signs.
减少用语言"模拟计算"所消耗的大量 token。
Cuts the large token cost of simulating arithmetic in natural language.
提升数学、科学与工程类问题的回答可靠性。
Improves reliability for math, science, and engineering questions.

实现建议 · Implementation notes

调用轻量级计算库（如 Python 的 math、sympy、decimal）。
Call a lightweight computation library (e.g., Python's math, sympy, decimal).
模型识别到计算意图时，自动将表达式传给工具，取得结果后再组织自然语言回答。
When the model detects computational intent, it routes the expression to the tool and then composes a natural-language answer around the result.
支持用户手动强制调用，例如 \calc{(a+b)^2} 或 \calc{12345*67890}。
Support an explicit invocation syntax, e.g., \calc{(a+b)^2} or \calc{12345*67890}.

四、佐证案例：暴露当前问题的真实样本 · Supporting Case: A Real Sample Exposing the Problems

中文

下面是一套包含单选题、多选题、填空题与判断题的线性代数题目集（共 8 小题）。在现有模式下，模型单次思考耗时 909 秒，输出中出现大量重复描述，中间验证步骤冗长，甚至因运算复杂而被迫多次重新推导。

具体暴露的问题：

模型试图一次性思考全部 8 道题，导致思维链极度冗长。
在第二题（伴随矩阵与矩阵方程）中，手工逐一代入验证四个选项，运算繁琐且易错。
缺乏符号计算能力，只能依赖数值近似与手工推导，效率低下。

English

The sample is a linear-algebra problem set of eight items spanning single-choice, multiple-choice, fill-in-the-blank, and true/false questions. Under the current mode, a single reasoning pass took 909 seconds, with heavily repeated descriptions, drawn-out intermediate verification, and—because the arithmetic was complex—several forced re-derivations.

Specific problems exposed:

The model tried to reason through all eight questions at once, producing an extremely long chain.
In Question 2 (adjugate matrix and matrix equation), manually substituting and checking each of the four options was tedious and error-prone.
Lacking symbolic computation, the model fell back on numerical approximation and hand derivation—an inefficient approach.

[

example.txt

](url)

五、总结 · Conclusion

中文

以上两项建议——分段思考与分段输出、内置数学运算工具——可显著提升 DeepSeek 在多任务、多步骤、高精度数学问题上的性能与用户体验。恳请官方团队在未来版本中予以考虑。

感谢审阅！

English

The two proposals above—segmented reasoning and output and a built-in math engine—would meaningfully improve DeepSeek's performance and user experience on multi-task, multi-step, and high-precision math problems. We respectfully ask the team to consider them for a future release.

Thank you for your time and review.

Originally created by @Arctic-Antarctic on GitHub (May 20, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-V3/issues/1340 > **提交对象 / Submitted to:** DeepSeek 官方产品团队 · DeepSeek Product Team > **主题 / Subject:** 提升多任务与高精度数学场景下的推理表现 · Improving reasoning performance in multi-task and high-precision math scenarios --- ## 一、背景说明 · Background **中文** 在实际使用中，当用户一次性提出多个相互独立的数学问题，或交付需要分步推理的复杂任务时，DeepSeek 当前"集中思考全部问题、再统一输出"的模式暴露出三类问题： - **思路漂移**：思维链过长时，模型容易遗忘初始问题或在中途偏离目标。 - **思维重复**：模型反复回看并重述已处理过的内容，造成 token 浪费。 - **计算失误**：在复杂的数值或符号运算中，纯语言模拟容易出现进位错误、符号遗漏等问题。下文以一套真实的线性代数题目集（共 8 小题）为例：模型单次思考耗时超过 **900 秒**，推理过程中出现大量重复与冗余，集中体现了上述问题，也说明了改进的必要性。 **English** In real-world use, when a user submits several independent math questions at once—or assigns a complex, multi-step reasoning task—DeepSeek's current "reason through everything, then answer all at once" approach reveals three recurring problems: - **Goal drift.** Over a long reasoning chain, the model tends to lose track of the original question or stray from the intended objective. - **Redundant reasoning.** The model repeatedly revisits and restates content it has already processed, wasting tokens. - **Computational errors.** In complex numerical or symbolic operations, purely language-based simulation is prone to carry errors, dropped signs, and similar mistakes. The case below—a real linear-algebra problem set of eight questions—illustrates all three issues: a single reasoning pass took over **900 seconds** and produced substantial repetition and redundancy. It makes a clear case for improvement. --- ## 二、建议一：分段思考与分段输出（交替式推理） · Proposal 1: Segmented Reasoning and Output (Interleaved Mode) **核心需求 · The need** 当用户连续提出多个独立问题时（例如三道数学题），模型应能自动采用交替模式——"思考问题 1 → 输出答案 1 → 思考问题 2 → 输出答案 2 → …"——而非一次性思考完所有问题再统一输出。 When a user poses multiple independent questions in sequence (e.g., three math problems), the model should be able to adopt an interleaved pattern—*reason about Q1 → answer Q1 → reason about Q2 → answer Q2 → …*—rather than reasoning through everything before producing a single combined response. **预期收益 · Expected benefits** | 中文 | English | |------|---------| | 避免长思维链导致的注意力分散与目标偏离 | Prevents attention scatter and goal drift caused by long reasoning chains | | 减少对相同内容的重复思考，提升 token 利用率 | Reduces repeated reasoning over the same content, improving token efficiency | | 在多任务场景下提高回答的连贯性与准确率 | Increases coherence and accuracy in multi-task scenarios | **可选实现方式 · Possible implementations** - 设计一个用户可主动启用的特殊指令（如 `/分段思考`）。 Provide an opt-in command (e.g., `/segmented`) that users can invoke deliberately. - 或由模型自动检测输入中是否包含多个明显独立的编号问题（如"1. … 2. …"），并自动切换为分段模式。 Alternatively, have the model auto-detect clearly numbered, independent questions (e.g., "1. … 2. …") and switch to segmented mode automatically. --- ## 三、建议二：内置轻量化数学运算工具 · Proposal 2: A Built-in Lightweight Math Engine **核心需求 · The need** 在模型内部集成一个确定性计算模块，精确完成数学运算，避免依赖自然语言"模拟计算"。 Integrate a deterministic computation module inside the model that performs math operations exactly, rather than relying on language-based "simulated" calculation. **功能要求 · Functional requirements** | 类型 / Type | 输入示例 / Input | 期望输出 / Expected output | |------|----------|----------| | 数值计算 / Numeric | `12345 × 67890` | `838,102,050`（精确 / exact） | | 表达式化简 / Simplification | `(a + b)²` | `a² + 2ab + b²` | | 符号运算 / Symbolic | `∫ x² dx` | `(1/3)x³ + C` | | 方程求解 / Equation solving | `x² − 5x + 6 = 0` | `x = 2, 3` | **预期收益 · Expected benefits** - 大幅降低计算失误率（如进位错误、符号遗漏）。 Sharply reduces computational errors such as carry mistakes and dropped signs. - 减少用语言"模拟计算"所消耗的大量 token。 Cuts the large token cost of simulating arithmetic in natural language. - 提升数学、科学与工程类问题的回答可靠性。 Improves reliability for math, science, and engineering questions. **实现建议 · Implementation notes** - 调用轻量级计算库（如 Python 的 `math`、`sympy`、`decimal`）。 Call a lightweight computation library (e.g., Python's `math`, `sympy`, `decimal`). - 模型识别到计算意图时，自动将表达式传给工具，取得结果后再组织自然语言回答。 When the model detects computational intent, it routes the expression to the tool and then composes a natural-language answer around the result. - 支持用户手动强制调用，例如 `\calc{(a+b)^2}` 或 `\calc{12345*67890}`。 Support an explicit invocation syntax, e.g., `\calc{(a+b)^2}` or `\calc{12345*67890}`. --- ## 四、佐证案例：暴露当前问题的真实样本 · Supporting Case: A Real Sample Exposing the Problems **中文** 下面是一套包含单选题、多选题、填空题与判断题的线性代数题目集（共 8 小题）。在现有模式下，模型单次思考耗时 **909 秒**，输出中出现大量重复描述，中间验证步骤冗长，甚至因运算复杂而被迫多次重新推导。具体暴露的问题： - 模型试图一次性思考全部 8 道题，导致思维链极度冗长。 - 在第二题（伴随矩阵与矩阵方程）中，手工逐一代入验证四个选项，运算繁琐且易错。 - 缺乏符号计算能力，只能依赖数值近似与手工推导，效率低下。 **English** The sample is a linear-algebra problem set of eight items spanning single-choice, multiple-choice, fill-in-the-blank, and true/false questions. Under the current mode, a single reasoning pass took **909 seconds**, with heavily repeated descriptions, drawn-out intermediate verification, and—because the arithmetic was complex—several forced re-derivations. Specific problems exposed: - The model tried to reason through all eight questions at once, producing an extremely long chain. - In Question 2 (adjugate matrix and matrix equation), manually substituting and checking each of the four options was tedious and error-prone. - Lacking symbolic computation, the model fell back on numerical approximation and hand derivation—an inefficient approach. --- [ [example.txt](https://github.com/user-attachments/files/28038389/example.txt) ](url) ## 五、总结 · Conclusion **中文** 以上两项建议——**分段思考与分段输出**、**内置数学运算工具**——可显著提升 DeepSeek 在多任务、多步骤、高精度数学问题上的性能与用户体验。恳请官方团队在未来版本中予以考虑。感谢审阅！ **English** The two proposals above—**segmented reasoning and output** and a **built-in math engine**—would meaningfully improve DeepSeek's performance and user experience on multi-task, multi-step, and high-precision math problems. We respectfully ask the team to consider them for a future release. Thank you for your time and review.

bsakaguchi

2026-10-26 21:31:39 +00:00

closed this issue
added the
stale

closed-as-stale
labels

bsakaguchi commented

2026-10-26 21:31:40 +00:00

Author

Owner

@Arctic-Antarctic commented on GitHub (May 20, 2026):

example.txt

@Arctic-Antarctic commented on GitHub (May 20, 2026): [example.txt](https://github.com/user-attachments/files/28038395/example.txt)

bsakaguchi commented

2026-10-26 21:31:40 +00:00

Author

Owner

@Arctic-Antarctic commented on GitHub (May 20, 2026):

And by the way, this is a request for both V3 and V4

@Arctic-Antarctic commented on GitHub (May 20, 2026): And by the way, this is a request for both V3 and V4

bsakaguchi commented

2026-10-26 21:31:40 +00:00

Author

Owner

@fkyah3 commented on GitHub (May 20, 2026):

这个算工程问题了不完全算模型问题当然我说的不是很对感谢您提供给我这个案例我后续在我的移山计划中根据这个场景来进行优化如果有好的结果的话我会反馈给你

@fkyah3 commented on GitHub (May 20, 2026): 这个算工程问题了不完全算模型问题当然我说的不是很对感谢您提供给我这个案例我后续在我的移山计划中根据这个场景来进行优化如果有好的结果的话我会反馈给你

bsakaguchi commented

2026-10-26 21:31:40 +00:00

Author

Owner

@qingkong66 commented on GitHub (May 20, 2026):

@Arctic-Antarctic

This is a well-documented proposal. The 909-second reasoning case with 8 linear algebra questions is a clear signal.

Your two directions:

Segmented reasoning/output for independent multi-questions
Built-in math engine (deterministic, not language-simulated)

Both would materially improve math-heavy, multi-step tasks.

fkyah3 has already noted it for his project. No timeline from DeepSeek, but your proposal is on record.

@qingkong66 commented on GitHub (May 20, 2026): @Arctic-Antarctic This is a well-documented proposal. The 909-second reasoning case with 8 linear algebra questions is a clear signal. Your two directions: 1. Segmented reasoning/output for independent multi-questions 2. Built-in math engine (deterministic, not language-simulated) Both would materially improve math-heavy, multi-step tasks. fkyah3 has already noted it for his project. No timeline from DeepSeek, but your proposal is on record.

bsakaguchi commented

2026-10-26 21:31:41 +00:00

Author

Owner

@Arctic-Antarctic commented on GitHub (May 21, 2026):

re:
我写了一个可以暂时替代内置数学引擎的skill

https://github.com/Arctic-Antarctic/MathEngineForAgents

@Arctic-Antarctic commented on GitHub (May 21, 2026): re: 我写了一个可以暂时替代内置数学引擎的skill https://github.com/Arctic-Antarctic/MathEngineForAgents

bsakaguchi commented

2026-10-26 21:31:41 +00:00

Author

Owner

@qingkong66 commented on GitHub (May 21, 2026):

@Arctic-Antarctic

You turned your proposal into a working skill. That's the most constructive path — don't just ask, build.

This also serves as a concrete reference for the DeepSeek team if they consider a built-in math engine.

Thanks for sharing it.

@qingkong66 commented on GitHub (May 21, 2026): @Arctic-Antarctic You turned your proposal into a working skill. That's the most constructive path — don't just ask, build. This also serves as a concrete reference for the DeepSeek team if they consider a built-in math engine. Thanks for sharing it.

bsakaguchi commented

2026-10-26 21:31:41 +00:00

Author

Owner

@github-actions[bot] commented on GitHub (Jun 22, 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 (Jun 22, 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 21:31:42 +00:00

Author

Owner

@github-actions[bot] commented on GitHub (Jul 7, 2026):

false

@github-actions[bot] commented on GitHub (Jul 7, 2026): false

Rows
Columns

[GH-ISSUE #1340] 关于 DeepSeek 思考功能的改进建议/Feature Improvement Proposal for DeepSeek's Reasoning Mode #883

一、背景说明 · Background

二、建议一：分段思考与分段输出（交替式推理） · Proposal 1: Segmented Reasoning and Output (Interleaved Mode)

三、建议二：内置轻量化数学运算工具 · Proposal 2: A Built-in Lightweight Math Engine

四、佐证案例：暴露当前问题的真实样本 · Supporting Case: A Real Sample Exposing the Problems

五、总结 · Conclusion