Another thing worth keeping in mind is that LLMs are feedforward networks, meaning they do everything in a fixed amount of time, no looping or second turns allowed. So even if the network has enough information to translate Bengali to English, that still doesn’t necessarily give it the power to answer question in Bengali directly, as by the time it has figured out what the Bengali means, it’s already out of layers to solve the problem.
Proper prompting can sometimes overcome those limitations, e.g. translate the problem into English first, feed the English text back into it to solve the problem and translate that output back to Bengali. The LLM itself can’t do all of that in one step, but if the human sends the proper prompts do to it step by step it can sometimes solve complex problems that it couldn’t do in one turn.