HRM confirmed by ARC-AGI team, but also dismissed as non-generalizable
the magic wasn’t in the hierarchical structure, it was in the outer loop. And the outer loop benefited mostly at training time
i.e. it mostly just memorized answers
arcprize.org/blog/hrm-ana...
HRM confirmed by ARC-AGI team, but also dismissed as non-generalizable
View original thread
@timkellogg.me
HRM: Hierarchical Reasoning Model
ngl this sounds like bullshit but i don’t think it is
- 27M (million parameters)
- 1000 training examples
- beats o3-mini on ARC-AGI
arxiv.org/abs/2506.21734
ngl this sounds like bullshit but i don’t think it is
- 27M (million parameters)
- 1000 training examples
- beats o3-mini on ARC-AGI
arxiv.org/abs/2506.21734
26
2
to be clear — it was a legit result. it completed ARC as advertised, no cheating
the problem (as interpreted by the ARC team) is that it won’t work beyond ARC-style problems
the problem (as interpreted by the ARC team) is that it won’t work beyond ARC-style problems
3
1