In neuroscience, we learned about foveated vision — our eyes focus sharply at the center and blur the edges to conserve energy.

I was curious how LLMs handle that kind of input, so I wrote a quick script to apply a foveated blur to 30 CAPTCHA images of cars, with the car in different regions each time.

At first, I asked GPT-4o:

“Do you see a car in this?”

But that felt too leading. So I switched to:

“What do you see in this image?”

No matter how intense the blur, GPT-4o consistently identified the car — even the blur made the image indescribable.

But in a fresh thread? It struggled. At one point, it even called a staircase a dessert. While in older conversation thread, it correctly identified it was a staircase with a car.

Turns out my original hypothesis — the blur would impair recognition — wasn’t quite right. Instead, the limiting factor was something more subtle: memory persistence.

I suspected the model might be relying on a previously mentioned keyword (“car”), so I asked if it was using prior context. It said no. But after pointing out the inconsistency in performance across threads, it admitted that earlier context had helped. Not maliciously — it just didn’t seem fully aware of how much prior memory shaped its output.

Claude 3.7 was different. It gave more cautious and consistent responses across threads, even when I primed it with the word “car.” Its answers weren’t influenced the same way.

It reminded me of a LessWrong post from exactly two years ago that questioned why LLMs don’t have access to long-term memory beyond their immediate context windows. Now, they seem to do exactly that — but at a cost. By heavily utilizing this long-term memory, even when instructed not to, model controllability is more difficult. 

I think it's because, for these models, instructions like “ignore previous context” are just more tokens. The instructions affect probabilities but aren't hard rules that override the prior activations that have already shaped their knowledge.

Curious if others have noticed something similar, especially with multimodal models. Are you seeing this kind of implicit context carryover too?