HRM (hierarchical reasoning model) does make sense (breaking down thinking to hierarchical reasoning - a loop of high level planning + low level computation). Does it work? The benchmark shown are impressive. Having said that, I doubt if it will be a replacement of existing CoT. CoT is also not a replacement of non-thinking architecture either. Many problems are solved better without thinking models. Many problems will be better solved with CoT. And a large number of problems will be more suitable for HRM. We need some more research on which problem space aligns with the non-thinking vs CoT vs HRM.
Old AI architecture is 1000x faster and accurate with 0 training examples: solvers
HRM (hierarchical reasoning model) does make sense (breaking down thinking to hierarchical reasoning - a loop of high level planning + low level computation). Does it work? The benchmark shown are impressive. Having said that, I doubt if it will be a replacement of existing CoT. CoT is also not a replacement of non-thinking architecture either. Many problems are solved better without thinking models. Many problems will be better solved with CoT. And a large number of problems will be more suitable for HRM. We need some more research on which problem space aligns with the non-thinking vs CoT vs HRM.