Comment by jeremyjh
7 hours ago
Why does CoT significantly improve their performance? Most of what they are applying they learned in post-training by solving similar problems themselves. This isn't about regurgitating pre-trained knowledged.
Advancements in Math and coding are because RLVR at massive scale is so cheap.
I have not seen the research on CoT and how it impacts spatial reasoning and outcomes, so I can’t comment on that part. I do see continuously that in almost every example of non-trivial image generation and 3d modeling, there are quirks that point to the fact that the model does not understand relationships within the space based on physics.
CoT seems to work really well for text based generation but once you’re past that and into physics models and detailed relationship mapping, it may work better but it’s not enough to make me believe it’s “reasoning” in a way that humans do.