MiroThinker-H1 verifies more, loops less

This feels less like a “give agents more steps” scaling story and more like a “teach them when to distrust themselves” story.

• –The Local Verifier is the interesting bit: it forces the model to seek disconfirming evidence before committing, which appears to cut wasteful loops instead of just adding more search.
• –The strongest numbers are tied to the closed H1 system, so the architecture looks promising but not fully reproducible on the flagship model.
• –The dramatic step drop may partly reflect fixing a looping baseline, so the efficiency win is real but probably not a universal law of verification.
• –The Tree of Thoughts comparison is only partial: ToT explores branches internally, while MiroThinker leans on actual tool feedback in the environment, which matters a lot for agentic tasks.
• –The compute curve also smells like diminishing returns: scaling from 16x to 64x buys only a small extra lift, so more budget helps, but not linearly.