Why does gemini-2.5-flash want to end the conversation quickly?

Gemini API
1

Hi. I used various LLM as “brain” of virtual characters in my application. Each character talk each other, or play TRPG, or explain slide show and so on.

When I change one character’s model from gpt-4.1 to gemini-2.5-flash, that character frequently end conversation or game suddenly.

Here is example (simplified, while playing TRPG):
A (DM, gpt-4.1) : So, what is your next action?
B (gpt-4.1) : Look carefully treasure box. It is strange.
C (gemini-2.5-flash) : Agreed. We need more information.
A : You examine carefully the treasure box, and found that key hole of the box.
B : Good, our choice is correct.
// Suddenly C end the conversation.
C : Thank you for watching. See you later.

When all models are gpt-4.1, it does not occurred and they can play over 30 min. But with gemini-2.5-flash, always C quit the conversation quickly so the conversation can’t last more than 5 min. System instrunction like “Don’t finish the game until I explicitly instruct you.” does not work.

Is this a characteristic of the Gemini-2.5-flash model? Is it trained to end the conversation as soon as possible?

3

@KichangKim ,

Rather than giving negative prompts have you tried giving objectives in the system prompt

e.g: You are RPG player and your task is to engage the other player as long as possible and come up with creative conversations to make thegame interesting for other player.

on a different note, what Temperature are you using on the gemini model?? increasing the “temp” settings might improve this behaviour.

1 Like
4

Thanks for information. I used default temperature, so I changed it to 2.0 and make prompts more be objective (keep conversation as long as possible). After that, I got more stable and robust results.