General Question : Why does LLM output the same graph on every trigger ? Why is there no variety in output

Course Q&A Agentic AI
, , ,
1

I am solving the ungraded assignment Chart Generation from LLM using reflection pattern. ChatGPT 4O is prompted to output a graph comparing coffee sales for 2 years (data is pre-loaded). However, on every trigger i notice that LLM gives the same graph as output without anything specific being called out. Even the colours on the bar chart are same. Further, a second LLM is called to read and feedback the graph and that also remains same.

Since LLMs are generative in nature, shouldn’t the output differ. Why is there no variety. Is there some kind of seed/additional in the backend which is leading to repeatability in output. Or is it some kind of hardcoding.

image
image986×633 42.3 KB

2

Hi Gunjan_Sharma,

As you can see, the LLMs output very basic code using matplotlib. As colors are not set by the LLMs, matplotlib uses default settings ( Colors in the default property cycle — Matplotlib 3.10.8 documentation ).

My interpretation is that the output does not differ because the issue at hand is rather straightforward and can be solved using basic code. There’s not much room for or interest in variety with this type of issue.

3

I agree. If you ask the tool to simply plot a given set of data, hopefully it won’t do anything random and give you a misleading plot.

4

Thanks , makes sense. I tried a new instance, and at least the text of the feedback generated by the refection LLM is changing.