[Critical] Gemini 2.5 Pro Response Error, Possible Memory Error?

[ilya]

Hello Google Team,

There is a problem with Gemini 2.5 Pro, especially on Gemini App and Gemini API

Vertex AI API doesn’t affected with this

Prompt Example to trigger, without any System Instructions:

Please Evaluate This Integral

Expected answer is 4π arctan(√φ) (Famous Cleo’s Integral)

Previously before general availability, this model can answer and identify the steps correctly

But now, it response spit out Random Context, possibly Another Customer Data, or Response

This also maybe the explanation for why sometime its response Answer -> Thinking -> Answer

Is there any probability there are memory corruption?

[ilya]

Sampling for this answer,

From observation, it does thinking twice, before it responding with this response

[Lalit_Kumar]

Hello and welcome to the forum!

To help me understand your question better, could you please share the exact prompt you’re using? Also, could you clarify if you’re providing image to Gemini or a text based mathematical equation?

[ilya]

Hello there, thanks for quick response for this case,

I use image as Input in gemini app

here is my prompt, for Clarity

And here the thinking

---

I also ask the same question through API, and its incorrectly read as “algebra” instead of “integral”

[Richard_Davey]

Isn’t the issue the fact that it failed to see the image? It says as much in the thinking: “I’ve hit a slight snag, as I cannot actually see the image.” - so from that point on, it’s just making it up entirely. I don’t believe this is memory pollution; I think it’s just trying to be overly helpful instead of coming back and saying there was an error parsing the image data - which is what it should have done, but didn’t.

For context, I just tried your image and it worked fine:

[ilya]

Thank you for the response,

You are right, this probably because the model cant see the image, I already repeated the prompt again and again, it finally can read the Integral problem

And this another respond, and finally can read the question (partial)

---

I apologize for identifying wrong subject, I primarily using this question as an benchmark first, but it seems the flaw is not the model itself, but its inability to read the image and trying to be helpful, its actually good but disastrous toward user that think this is hallucinating too much.

You are right to point that in the thinking its point its inability to read image..

But the second try it just really confusing.

[Richard_Davey]

No, you’ve definitely found what I would consider to be a “bug” in Gemini - if it can’t read the image, it should just bail out right there and not waste your API tokens!

I think you’ve found an error for sure, it just wasn’t memory-related. It’s strange because if an image fails to upload, you get an error from the API then. And if it fails to read the image, you get an error from that too (like if the image is too large or corrupted) - but you’ve found something else entirely, where the image uploaded just fine but got lost between the upload and the model thinking about it. Which is quite a fascinating bug!

[ilya]

Yes, its inability to read the proper context, also this happen in the new gemini, I dont know, but this is intermittent

I also have some API response blank, which cost me input token, the rate its happen seems high

But its on Gemini API, on Vertex its like 100% “accurate”

Using Gemini 2.5 Pro using API, the aystem instructions is like blended and not obeyed.

What I mean by “blended”, I told the llm to Stop the persona if this keyword exist,

Its really stop for older iteration on Gemini API, but today its just getting worse

[Lalit_Kumar]

Hello!

I tried to reproduce the issue using the information you provided and in my case the model was able to understand the problem.

It looks like the model might be hallucinating in your case, but we can usually fix this by giving it a bit more context. When you’re passing an image to the model, try adding some additional information about that image in your prompt. This extra detail should help the model understand better and give better results.

I hope this helps you.