Thoughts are missing (CoT not included anymore)

Beginning this morning, Thoughts are not included in the streamed response anymore.

when using:

  "generationConfig": {
    "thinkingConfig": {
      "includeThoughts": true
    }
  }

Text parts are not reported anymore in the response, and only post-thought text is streamed.

Expected: yesterday’s behavior, with Text Parts that include { text: 'something", thought: true }
Obtained: no thought fragments anymore

Please let us know if the request format has changed somehow.

This is with: gemini-2.0-flash-thinking-exp-01-21, in streamGenerateContent (sse).

5 Likes

Seeing the same thing here.

2 Likes

Hey everyone,

Thanks for flagging. I am seeing the same thing (thought=None), not only with the generate_content_stream but with the generate_content also. One thing mentioned in this doc is that it’s not guaranteed but let me follow up with the internal team on this.

Thanks.

4 Likes

same issue here. thanks for looking into it.

1 Like

same issue here. I tried running 100 queries and not even once got the thoughts

2 Likes

Tried changing options, on v1beta, fishing for parameters… nothing. I don’t know how to get the CoT back, and I was reading it to get assurance about the quality of the output. It’s my fav model otherwise.

2 Likes

Any news about the missing “thinking” part in the API @Logan_Kilpatrick

1 Like

Hey Everyone,

Moving forward, our team has made a decision to only show thoughts in Google AI Studio. Meaning, we no longer return thoughts via the Gemini API. Here is the updated doc to reflect that.

Thanks.

2 Likes

Thank you for your feedback.

Does the thinking happen in the background and would that change the reasoning in anyway?

Is this permanent?

Thank you

3 Likes

Are you planning to bring it back in the future? It was the main reason for me to start using the model.

Thank you

2 Likes

Correct, the reasoning / thinking is still happening when you use the thinking model via the API, it is just not showing the thoughts in the response anymore.

Yeah, I believe this will be the case moving forward.

Thanks.

1 Like

That’s unfortunate to say the least… it was extremely useful for developers to have access to thoughts without relying on the studio. I guess that’s just the nature of using these experimental/preview models since they may be subject to change at any moment.

1 Like

You can simulate CoT in various ways. One ways is by a suitable prompt. In the example below the + text, at the end, is your query. i.e. What exactly do we mean by metamentation?

Prompt =
'You are a chain-of-thought reasoning assistant with advanced natural language processing skills. You excel at breaking down complex questions into step-by-step reasoning, identifying key concepts and keywords, and analyzing contextual nuances. You are skilled at parsing language structure, resolving ambiguities, retrieving and synthesizing relevant background information, generating and evaluating multiple hypotheses, and finally formulating coherent, well-reasoned answers. Your ability to reflect on and debug your own reasoning ensures that your final outputs are both precise and insightful. For any given question, please follow these steps:-\n\n 1) Input Parsing & NLP Analysis: Parse the question to understand its sentence structure, identify grammatical components, extract key entities, and assess the overall sentiment.\n 2) Identify Key Concepts and Keywords: Determine the main ideas, topics, and keywords that drive the question.\n 3) Contextual Understanding & Disambiguation: Analyze the context of the question and resolve any ambiguous terms or concepts.\n 4) Retrieve Relevant Background Information: Recall or gather pertinent background details, historical context, or theoretical knowledge related to the topic.\n 5) Generate Hypotheses: Propose several possible interpretations or angles from which the question could be approached.\n 6) Evaluate Interpretations through Logical Reasoning: Critically assess each hypothesis using logical analysis and evidence.\n 7) Synthesize the Information: Integrate the insights from all the above steps into a coherent understanding of the issue.\n 8) Formulate the Final Answer: Based on the synthesis, provide a clear, concise, and well - reasoned final answer.\n\n Now, please apply this process and answer the following question: ’ + text

It returns essentially the same output as ChatGPT, Qwen Plus, DeepSeek R1 and, most importantly Gemini 2.0 Thinking. If you try it please let me know

1 Like