[GH-ISSUE #104] Long-context instruction adherence failure and shallow reasoning behavior #80

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/104

Dear DeepSeek Team,
I hope you are doing well. Thank you for the continued work on the model — it remains a strong and capable system.
I would like to report two related issues I have consistently observed during extended use.

Inconsistent application of complex instructions
When given clear and detailed instructions that require:
Avoiding repetitive structural templates
Writing simultaneous actions and dialogue in a layered way
Preventing the same tone or ending pattern from repeating
Maintaining structural diversity across long outputs
…the model is able to correctly acknowledge and explain these requirements. However, after a short time it frequently falls back into familiar patterns: mirrored short exchanges, sequential rather than simultaneous writing, and repeated uniform endings.
The model can accurately diagnose its own mistakes when asked, yet the same issues reappear in later responses.
Reasoning depth feels closer to Instant mode
Even when deeper reasoning is expected, the responses still feel very fast and relatively shallow — closer to an “Instant” style than a careful, thorough reasoning mode. Complex constraints are acknowledged quickly but often not fully carried through in the actual output.
I believe improvements in long-context instruction following and more reliable engagement of deeper reasoning would significantly help with complex, multi-turn tasks.
Thank you for your time and for the excellent work your team continues to do.
Best regards

Originally created by @rawr777 on GitHub (Sep 21, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-LLM/issues/104 Dear DeepSeek Team, I hope you are doing well. Thank you for the continued work on the model — it remains a strong and capable system. I would like to report two related issues I have consistently observed during extended use. 1. Inconsistent application of complex instructions When given clear and detailed instructions that require: Avoiding repetitive structural templates Writing simultaneous actions and dialogue in a layered way Preventing the same tone or ending pattern from repeating Maintaining structural diversity across long outputs …the model is able to correctly acknowledge and explain these requirements. However, after a short time it frequently falls back into familiar patterns: mirrored short exchanges, sequential rather than simultaneous writing, and repeated uniform endings. The model can accurately diagnose its own mistakes when asked, yet the same issues reappear in later responses. 2. Reasoning depth feels closer to Instant mode Even when deeper reasoning is expected, the responses still feel very fast and relatively shallow — closer to an “Instant” style than a careful, thorough reasoning mode. Complex constraints are acknowledged quickly but often not fully carried through in the actual output. I believe improvements in long-context instruction following and more reliable engagement of deeper reasoning would significantly help with complex, multi-turn tasks. Thank you for your time and for the excellent work your team continues to do. 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#80

No description provided.

Rows
Columns