Gemini 3.1 Pro ignores instructions + thoughts on the "thought process"🤔

Google AI Studio
, , , ,
1

This is a fairly common issue. Gemini 3.1 Pro, even with a small context (<32k), completely ignores prompts like “write an answer of X lines/X pages/X words/X characters/X tokens.” I don’t know if this is done to save money or not, but if I ask the model to give a long answer, I, as in the case of 3.0 pro and 2.5 pro, want to get a LONG and detailed answer.

I’m also extremely confused by the model’s inability to determine what I’m asking her to reflect on. I was recently conducting a small psychology study and asked a model to reflect on my past messages and provide a short answer about how Person A and Person B would react in the given situations. The request was for a straightforward answer, so I could copy and paste it as fact, not just a thought. The result: The model’s “Thought Process” only considered my request, not my past messages, which I asked for. This resulted in the model, instead of a short answer like “Person A will want to go to the movies, but Person B will refuse,” producing a huge, meaningless tranche of thoughts about why and how she arrived at this conclusion. In my opinion, this should have been in the “Thought Process,” but not in the answer itself, which was supposed to be brief.

I’d also like to point out that the model doesn’t respect the request for detailed reasoning. Despite the “High” mode enabled, it thinks very briefly, again dumping its thoughts into the main answer, which in some situations is unacceptable for me. 2.5 Pro was more flexible in this regard, as the number of tokens in its reasoning was quite configurable, which made it possible to force the model to truly consider its answer before giving it.

Therefore, I believe that the 3.1 Pro model is completely incapable of following instructions, and its “Thought Process” is also flawed.

Is this issue still relevant to anyone? If so, please write in the comments.

(I don’t speak English, but I hope I’ve made my thoughts clear.)

1 Like
2

At the same time, the model also has a tendency to get hung up on the same old thing, even if you ask her to forget about some detail, after 2-3 messages she returns to the topic again, even if there was no request. Even if you include this in the system instructions, the model won’t care; it will just … its clever “since you’re a student and you love cats, I can tell that your research…”.
This demonstrates both the impossibility of using the Gemini 3.1 pro as a serious instrument (when it doesn’t involve programming, of course) and is further evidence of the inability to follow instructions.

3

This doesn’t seem like a serious problem until the model starts mixing everything up into one mess even on a small context; this is a clear problem of “priority of attention”, which again, does not allow using Gemini 3.1 pro for purposes greater than searching for necessary links.

4

I’ve also encountered the issue of the model lacking a thought process. As strange and counterintuitive as it sounds, the only way to force the model to actually think—rather than just spitting out disconnected text—is to enable the “grounding with Google Search” option.

Failing to follow instructions is a separate issue altogether, and unfortunately, there is no magic button to fix it. I started noticing this around February, after running into several situations that made me seriously question whether it makes sense to continue using the Gemini 3.1 model and Google AI Studio in particular.

Below are a few examples of what I mean. These aren’t the only instances, but they clearly illustrate the problem.

Example 1:

  • Task: Reduce the word count of a literary text to a specific limit while preserving key scenes and the core narrative.

  • Result: The word limit was ignored, key scenes were omitted, the emphasis was misplaced, and the essence of the story was completely lost.

Example 2:

  • Context: A roleplay chat, roughly 40k tokens long.

  • Issue: Character A, instead of tearing a drawing out of their sketchbook, carefully cuts it out with scissors with the intention of gifting it to Character B. Character B (played by Gemini 3.1) completely ignores both this specific action and the emotional tone of the scene. Instead, the model writes that Character B received a drawing that Character A had angrily shredded into tiny pieces with scissors.

Example 3:

  • Context: A short chat, roughly 10k tokens long.

  • Task: Using the model to simplify calculating dietary macronutrients for a strict, health-mandated diet. The prompt included comprehensive and explicit instructions on exactly what was required.

  • Result: The instructions were entirely ignored. Instead of calculating macros, the model provided unsolicited recommendations to follow intermittent fasting alongside a critically low-calorie meal plan. Both practices are potentially dangerous. Furthermore, there were no requests for weight loss—only a request to balance macronutrients.

1 Like