I tried to tell him in the system instructions not to write anything for a long time and then respond, but it doesn’t work. Did someone try to turn this guy into a strawberry guy (with chains of reasoning)?
2 Likes
The system typically requires one or two prompts before generating a response. Try responding when it seems to be thinking like “get it? , finish?” , and you might get an answer faster. Good luck !
1 Like
Welcome to the forums!
CoT prompting involves a little more than just saying “don’t write anything for a long time”. Language models don’t work that way, and Gemini is a language model.
At a minimum, you need to include instructions about “consider this question step by step and tell me the steps” or examples that setup how a human would consider the problem.
Remember - these are pattern machines, so you need to set up a pattern for them to work with.
The o1 “model” probably does these sorts of things but expands upon it - either by having some agentic-like processing that guides it along step by step or that may take multiple paths and evaluate how each path is working.
What is interesting is that while Gemini seems to have some of this training (it seems to do about as well as math tasks as o1 from some reports), it isn’t a full agentic or reasoning system.
5 Likes
@afirstenberg has excellent advice. To add to it, you can prompt the model to put the reasoning in <thinking> tags and the result in <result> tags.
Another one is to ask the model to work backwards from result to reasoning how that result would be achieved and if there’s a flaw then correct the flaw in thinking and work towards the result based on this corrected reasoning.
5 Likes
It does not work because the Google Team as a whole is LAZY…! I told them million of time to fix this bug, they did NOTHING!
1 Like