MobileSpec
An LLM personalization training system for mobile devices that trains lightweight expert models solely through forward propagation and verifiable reward signals, without requiring backpropagation or manual annotation.
Timeline (last 90 days)
The MobileSpec system trains expert models solely through forward passes and learns from the results of automatically checked tasks, similar to reinforcement learning with verifiable rewards (RLVR).
MobileSpec's core algorithm FuRA-S trains only the block spectral coordinates of the pretrained weights (the singular values of column blocks), rotating the singular vectors by independently rescaling these blocks.
On the phone-use agent task, MobileSpec raises the success rate from 6-23% to 89-100%.
MobileSpec can train a personalized 8B model overnight on an iPhone 16 Pro.
By avoiding backpropagation and labeled data, MobileSpec lowers the training barrier for on-device LLM personalization, enabling ordinary users to obtain expert models tailored to their own tasks on their phones.
All Facts (5)
The MobileSpec system trains expert models solely through forward passes and learns from the results of automatically checked tasks, similar to reinforcement learning with verifiable rewards (RLVR).
50%UnverifiedMobileSpec's core algorithm FuRA-S trains only the block spectral coordinates of the pretrained weights (the singular values of column blocks), rotating the singular vectors by independently rescaling these blocks.
50%UnverifiedOn the phone-use agent task, MobileSpec raises the success rate from 6-23% to 89-100%.
50%UnverifiedMobileSpec can train a personalized 8B model overnight on an iPhone 16 Pro.
50%UnverifiedBy avoiding backpropagation and labeled data, MobileSpec lowers the training barrier for on-device LLM personalization, enabling ordinary users to obtain expert models tailored to their own tasks on their phones.
50%