Hi guys - great job with the Flash 2.0 - thinking of moving to this model for all our primary uses after it goes GA.
Quick question re. Thinking Mode – are there reference prompts available (a cookbook perhaps) to guide the thinking process. Right now, the model sometimes follows the thinking instructions, other times it just re-iterates the input in the thinking process and does the thinking as a part of the output. I get the feeling the model is unable to understand what part of the thinking should be done inside thinking mode and what part is the output expected.
Some reference examples would help sort this out – esp. around how to guide the thinking mode of the model.
1 Like
Hi @Abhishek_Iyer . Welcome to the forum.
We have a cookbook available. You can find the link below.
Gemini 2.0 Flash Thinking
1 Like
Thanks for the pointer!
However, it looks like the prompts here are quite straightforward, and none of the examples are around “guiding” the thinking through a process.
Any examples where the prompt explicitly acknowledges and guides the thinking process?
For eg, after playing around the model for a bit I noticed the model started referring to the thinking process as “Mental Sandbox”. Is this the way to unambiguously reference the thinking process in prompts?
Thanks for your feedback. We currently only have the example notebook mentioned above. We appreciate your suggestion and will consider it as a feature request.
I’ve found that Thinking doesn’t want much prompt engineering. It’s happy with personality and conversational tone but chatHistory and nlpAnalysis (entity, keyword, summarisation and emotion) tend to constrain it such that, as you have observed, it simply returns a rehashed prompt.
Asking Thinking to explain its own thinking processes is revealing.
e.g. Write a template that could be used to emulate the Thinking process used by the Gemini LLM
Yeah, this is my observation as well. Get it to write a prompt to guide itself (ie. iterate with it to write the prompt and use that prompt) seems to provide the best results.
Still would be cool to have something from the team that provides the “prompt style” used during training (similar to what Anthropic does)
From my understanding, Gemini is the sandbox itself…