MiroThinker verification loop lifts BrowseComp, slashes steps
MiroThinker’s new paper argues that inference-time verification matters more than letting an agent wander longer, with a local verifier on a hard BrowseComp slice lifting Pass@1 from 32.1 to 58.5 while cutting interaction steps from 1,185 to 211. The release also includes open-weight MiroThinker-1.7 and 1.7-mini, while the strongest H1 results remain closed; the open models are competitive on several research and domain benchmarks, but they still trail the best proprietary systems in a few areas.
The most interesting takeaway is that verification seems to improve both quality and efficiency, which is unusual for agent papers and makes the result feel more structural than benchmark-chasing.
- –The +26.4 point gain on hard BrowseComp is the headline, but the 6x step reduction is the deeper signal: a lot of long-horizon agent compute was apparently spent on bad trajectories.
- –The open 1.7-mini is the practical release to watch, but the paper is careful to separate it from the closed H1 system that sets the top numbers.
- –This feels promising but not fully model-agnostic yet: a verifier is only as good as the signal in the base policy, so confidently wrong models may still be hard to rescue.
- –The agentic scaffold matters here too; without tool use, context management, and the verifier loop, you are not reproducing the reported behavior, just running a Qwen3 MoE base.
DISCOVERED
23d ago
2026-03-19
PUBLISHED
23d ago
2026-03-19
RELEVANCE
AUTHOR
Much-Movie-695