"Contemplating mode orchestrates multiple agents that reason in parallel, enabling superior performance with comparable latency, achieving a high water mark of 58.4 on Humanity's Last Exam."
"Muse Spark shows smooth predictable gains after additional reinforcement learning steps, improving model reliability without compromising reasoning diversity, utilizing thinking time penalties to balance correctness and token optimization."
"In testing on the AIME 2025 benchmark, a phase transition was observed where the model compressed equally accurate reasoning into significantly fewer tokens, enhancing efficiency."
"The release of Muse Spark coincides with an update to Meta's Advanced AI Scaling Framework, which now addresses a broader range of potential AI risks."
Meta's Muse Spark features Contemplating mode, which utilizes up to 16 agents working in parallel for enhanced performance. This mode reportedly achieves a high score of 58.4 on Humanity's Last Exam. The model benefits from reinforcement learning, showing predictable gains in reliability without sacrificing reasoning diversity. Testing revealed a phase transition where the model compressed accurate reasoning into fewer tokens, leading to improved efficiency. Muse Spark is now available in the Meta AI app and on the meta.ai website, alongside updates to Meta's AI risk management framework.
Read at Ars Technica
Unable to calculate read time
Collection
[
|
...
]