Key Info
Meta described two capabilities that make Muse Wonder suited for long-running, multi-step agentic work, and reported test-time compute scaling results showing Muse Wonder Full handles unfamiliar tasks better than the Muse Spark 1.3 trend predicts.
Highlights
- Active Memory carries context across conversations and surfaces it when relevant.
- Active Focus keeps models working toward the goal on challenging tasks, trying alternative approaches, working through blockers, and following through to completion.
- Meta compared test-time compute scaling on tasks seen during post-training versus other tasks to quantify handling of unfamiliar work.
- Both models improve similarly out-of-distribution, but at the same in-distribution score, Muse Wonder Full performs better on unfamiliar tasks than the Muse Spark 1.3 trend predicts.
- Meta attributes this upward shift to a stronger base model that carries post-training gains more effectively to unanticipated tasks.
