[GH-ISSUE #103] Model correctly diagnoses instruction failures but continues repeating the same structural patterns #79

New issue

Open

opened 2026-10-26 21:02:26 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 21:02:26 +00:00

Owner

Originally created by @rawr777 on GitHub (Sep 21, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-LLM/issues/103

Dear DeepSeek Team,
I hope this message finds you well. First, I would like to express my sincere appreciation for the strong reasoning and language capabilities of your model. It is clearly a powerful system.
I would like to respectfully report a recurring issue I have encountered during extended use.
When given detailed and explicit instructions that require:
Avoiding repetitive structural templates
Writing simultaneous actions and dialogue in a layered manner
Preventing the same tone or ending pattern from repeating across multiple segments
Maintaining structural diversity over long outputs
…the model is able to correctly acknowledge and explain these requirements, but frequently fails to consistently apply them in the actual output. After a short time, it tends to fall back into familiar patterns: mirrored short exchanges, sequential rather than simultaneous writing, and repeated quiet or uniform structural endings.
The model can accurately diagnose its own mistakes when asked, yet the same issues reappear in subsequent generations.
I believe improving long-context instruction adherence, especially regarding structural diversity and consistent application of complex formatting constraints, would significantly enhance the user experience for complex, multi-turn tasks.
Thank you very much for your time and for the excellent work your team continues to do. I look forward to future improvements.
Best regards

Originally created by @rawr777 on GitHub (Sep 21, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-LLM/issues/103 Dear DeepSeek Team, I hope this message finds you well. First, I would like to express my sincere appreciation for the strong reasoning and language capabilities of your model. It is clearly a powerful system. I would like to respectfully report a recurring issue I have encountered during extended use. When given detailed and explicit instructions that require: Avoiding repetitive structural templates Writing simultaneous actions and dialogue in a layered manner Preventing the same tone or ending pattern from repeating across multiple segments Maintaining structural diversity over long outputs …the model is able to correctly acknowledge and explain these requirements, but frequently fails to consistently apply them in the actual output. After a short time, it tends to fall back into familiar patterns: mirrored short exchanges, sequential rather than simultaneous writing, and repeated quiet or uniform structural endings. The model can accurately diagnose its own mistakes when asked, yet the same issues reappear in subsequent generations. I believe improving long-context instruction adherence, especially regarding structural diversity and consistent application of complex formatting constraints, would significantly enhance the user experience for complex, multi-turn tasks. Thank you very much for your time and for the excellent work your team continues to do. I look forward to future improvements. Best regards

No labels

bug

pull-request

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

deepseek-ai/DeepSeek-LLM#79

No description provided.

Rows
Columns