deepseek-ai

/

DeepSeek-R1

like 11k

Follow

DeepSeek 44.1k

Environment:

• Model: DeepSeek-R1
• Broken Revision: [both 8a58a132790c9935686eb97f042afa8013451c9f and a157fa3d494497a54586a333a23df6c2143e7697]
• Working Revision: f7361cd9ff99396dbf6bd644ad846015e59ed4fc

Issue Description:
The model has stopped outputting the <think> tag in responses, which was previously available in revision f7361cd. This breaks expected chain-of-thought behavior.

Reproduction Steps:

1. Send prompt:

   Hi
   

2. Response:

Expected Output (from revision f7361cd):

"<think>\n\n</think>\n\nHello! How can I assist you today? 😊"

Actual Output (latest revision):

"Okay, the user just said \"Hi\". That's a friendly greeting. I should respond in a warm and welcoming manner. Maybe say something like, \"Hello! How can I assist you today?\" to invite them to ask for help. Keep it open-ended so they feel comfortable to share what they need. Make sure to maintain a helpful and approachable tone.\n\nI need to check if there are any specific guidelines about responding to greetings. The instructions mention being helpful, using clear language, and avoiding markdown. So keep the response simple and professional. No need for any formatting here. Just a straightforward reply.\n</think>\n\nHello! How can I assist you today?"