Hmm, the user is asking for a detailed explanation about whether multimodal large models can truly reason and the difficulties involved, specifically requesting it's not too technical. They seem genuinely curious about AI capabilities but likely lack a machine learning background.
First, I need to unpack what "reasoning" means in this context. From the search results 12, it's clear that true reasoning involves step-by-step thinking (like chain-of-thought) rather than just pattern matching. The user might be wondering if these models actually "think" like humans.

The challenges section is crucial here. Several pain points emerge from the literature:

1. Scale issues 1 showing smaller models struggle with reasoning even when fine-tuned
2. Modal alignment problems 3 where combining vision/language meaningfully is hard
3. The "pseudo-reasoning" trap 5 where models statistically mimic reasoning but lack real understanding

What's interesting is how the EMMA benchmark 5 exposes these flaws - models fail at physics questions requiring visual-spatial reasoning that humans solve effortlessly with quick sketches. This highlights the core issue: current models process modalities separately rather than holistically integrating them.

The user might also care about practical implications. The "perception-decoupled" approach 10 mentioned could be a game-changer - it suggests we might achieve better reasoning by having specialized components handle vision vs text separately. This architecture could make the system both more powerful and easier to improve over time.

I should emphasize the human-AI comparison too. The papers constantly note how humans use visual simulation (like sketching force diagrams) while models rely on textual steps 5. This difference in cognitive approach explains why current models still lag behind human-like reasoning.

For the response structure, I'll start simple with analogies, then break down the "does it reason?" question with concrete examples before detailing the three major technical hurdles. The conclusion should offer balanced hope (progress being made) while managing expectations about current limitations.