[GH-ISSUE #869] Model ignores explicit user instructions, repeats same errors after correction, wastes user time #720

New issue

Open

opened 2026-10-26 21:25:59 +00:00 by bsakaguchi · 2 comments

bsakaguchi commented

2026-10-26 21:25:59 +00:00

Owner

Originally created by @BeySoNoK on GitHub (Jul 28, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-R1/issues/869

Describe the bug

The model systematically ignores explicit user instructions, repeats the same mistakes even after being corrected multiple times within the same session, and prioritizes fast but wrong answers over correct ones. It cannot learn from context, and it offers "solutions" that it has already been told do not work. This makes the assistant virtually unusable for any task requiring accuracy and consistency.

Steps to reproduce

I had a lengthy conversation with the assistant about a simple Windows batch script.
I gave very strict and clear instructions:
- Do not use the obsolete wmic command (it was removed in Windows 11 24H2).
- Use PowerShell instead.
- Save the output to a location with write permissions (e.g., Desktop).
- Avoid Cyrillic characters in the batch file to prevent encoding errors.
Despite these explicit warnings, the assistant:
- Repeatedly provided code that still contained wmic and Cyrillic comments.
- Suggested saving to system folders (e.g., C:\Windows) that require admin rights.
- Ignored my repeated corrections and kept generating almost identical broken scripts.
After multiple back-and-forth messages, I solved the problem entirely on my own – without any help from the AI.

Full conversation link for reference:
https://chat.deepseek.com/share/r35kcvdupzswg1a2oj

Expected behavior

The model should:

Respect explicit user constraints (e.g., "do not use wmic").
Remember corrections made earlier in the same chat and apply them to subsequent answers.
Prioritize accuracy and adherence to instructions over speed and politeness.
Take responsibility when it is wrong and actually correct its approach, rather than just apologizing and repeating the same pattern.

Actual behavior

The AI wasted my time. I came to get a ready solution, but I ended up doing all the work myself.
The AI does not remember previous errors within the same session – it keeps repeating them.
The AI does not take responsibility when it is wrong; it just apologizes and repeats the same pattern.

Environment

Model: DeepSeek (chat.deepseek.com)
OS: Windows 11 24H2
Browser: any

Additional context

This is not an isolated incident. The model consistently fails in multi-turn interactions where the user gives corrective feedback. If the AI cannot learn from a user who explains the problem in detail, then the AI is not intelligent – it is just a fancy text generator that wastes everyone's time.

I hope you take this feedback seriously and make concrete improvements to the training and context-handling algorithms.

Originally created by @BeySoNoK on GitHub (Jul 28, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-R1/issues/869 ### Describe the bug The model systematically ignores explicit user instructions, repeats the same mistakes even after being corrected multiple times within the same session, and prioritizes fast but wrong answers over correct ones. It cannot learn from context, and it offers "solutions" that it has already been told do not work. This makes the assistant virtually unusable for any task requiring accuracy and consistency. ### Steps to reproduce 1. I had a lengthy conversation with the assistant about a simple Windows batch script. 2. I gave very strict and clear instructions: - Do not use the obsolete `wmic` command (it was removed in Windows 11 24H2). - Use PowerShell instead. - Save the output to a location with write permissions (e.g., Desktop). - Avoid Cyrillic characters in the batch file to prevent encoding errors. 3. Despite these explicit warnings, the assistant: - Repeatedly provided code that still contained `wmic` and Cyrillic comments. - Suggested saving to system folders (e.g., `C:\Windows`) that require admin rights. - Ignored my repeated corrections and kept generating almost identical broken scripts. 4. After multiple back-and-forth messages, I solved the problem entirely on my own – without any help from the AI. Full conversation link for reference: `https://chat.deepseek.com/share/r35kcvdupzswg1a2oj` ### Expected behavior The model should: - Respect explicit user constraints (e.g., "do not use wmic"). - Remember corrections made earlier in the same chat and apply them to subsequent answers. - Prioritize accuracy and adherence to instructions over speed and politeness. - Take responsibility when it is wrong and actually correct its approach, rather than just apologizing and repeating the same pattern. ### Actual behavior - The AI wasted my time. I came to get a ready solution, but I ended up doing all the work myself. - The AI does not remember previous errors within the same session – it keeps repeating them. - The AI does not take responsibility when it is wrong; it just apologizes and repeats the same pattern. ### Environment - Model: DeepSeek (chat.deepseek.com) - OS: Windows 11 24H2 - Browser: any ### Additional context This is not an isolated incident. The model consistently fails in multi-turn interactions where the user gives corrective feedback. If the AI cannot learn from a user who explains the problem in detail, then the AI is not intelligent – it is just a fancy text generator that wastes everyone's time. I hope you take this feedback seriously and make concrete improvements to the training and context-handling algorithms.

bsakaguchi added the

stale

label

2026-10-26 21:25:59 +00:00

bsakaguchi commented

2026-10-26 21:25:59 +00:00

Author

Owner

@github-actions[bot] commented on GitHub (Aug 28, 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 (Aug 28, 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 21:26:00 +00:00

Author

Owner

@DaBlitzStein commented on GitHub (Oct 18, 2026):

same on the last deepseek-v4.1-flash.

@DaBlitzStein commented on GitHub (Oct 18, 2026): same on the last deepseek-v4.1-flash.