New Breakthrough in On-Device LLM Training: Forward-Only Optimization Enables Personalized Expert Models

MobileSpec uses forward-only training to create 1-5MB expert models on phones, raising agent success rates to 89-100%.
MobileSpec is an on-device LLM personalization system that avoids backpropagation's memory costs and the scarcity of labeled data. It updates models via forward propagation only and uses RLVR to obtain rewards from automatically checked task outcomes. Its core algorithm, FuRA-S, trains only block spectral coordinates (column-block singular values) of pre-trained weights, producing perturbations that resemble RLVR's weight changes, rotate singular vectors, and work with INT4 quantized models without weight modification. Expert models are 1-5MB with 0.015-0.05% of parameters updated; mobile agent success rises from 6-23% to 89-100%, and an 8B model can be trained overnight on an iPhone 16 Pro, performing comparably to full-parameter ES and better than ES on LoRA.
Running large language models (LLMs) on personal devices keeps user data local, but limited memory restricts the model's general capabilities. Most user-specific tasks do not require the breadth of frontier models, so a practical solution is to adapt a shared on-device model into a lightweight expert model. However, backpropagation is memory-expensive, and users rarely provide labeled data. To address this, researchers propose MobileSpec—a system that trains expert models through forward propagation only, learning from automatically checked task outcomes, similar to reinforcement learning with verifiable rewards (RLVR).

Core Challenges: Dual Constraints of Memory and Labeled Data
Training LLMs on memory-constrained devices like phones faces two major obstacles. First is the memory overhead of backpropagation: traditional training requires storing activations and gradients, far exceeding the capacity of mobile devices. Second is the scarcity of labeled data: users typically do not provide supervision signals for specific tasks. MobileSpec's approach is to use only forward propagation and leverage the RLVR framework to automatically obtain reward signals from task outcomes, thereby avoiding backpropagation and manual labeling.
However, forward-only updates based on random perturbations are only effective when these perturbations resemble the weight changes of RLVR. RLVR's weight changes largely preserve the singular values of pre-trained weights and rotate their singular vectors. MobileSpec's core algorithm, FuRA-S, is designed to produce such perturbations.
Backpropagation is the standard algorithm for training neural networks. It propagates gradients back from the loss function layer by layer via the chain rule to update all weights. Its memory overhead mainly comes from two aspects: first, during forward propagation, the activations of each layer must be saved to compute gradients in the backward phase; second, for each trainable parameter, corresponding gradients and optimizer states (such as the first and second moments in Adam) must be stored. For LLMs with billions of parameters, this memory is typically several times the model's own weights, far beyond the capability of mobile devices. In contrast, forward-only optimization methods only need to perform inference and observe outputs, without saving intermediate activations, reducing memory usage by more than an order of magnitude.
RLVR (Reinforcement Learning with Verifiable Rewards) is a reinforcement learning paradigm that uses programmatic rules to automatically determine the correctness of answers. Unlike RLHF, which relies on human preference labeling, RLVR's rewards come from verifiable outcomes, such as whether a math answer matches, whether code passes unit tests, or whether an agent task is completed. This enables training signals to be obtained at scale with zero manual labeling. In MobileSpec, RLVR provides the sparse reward of "task outcome," based on which the model judges whether a forward perturbation improves behavior, thus completing parameter updates without backpropagation.
FuRA-S Algorithm: Efficient Perturbation on Block Spectral Coordinates
The key innovation of FuRA-S is training only the block spectral coordinates of pre-trained weights—i.e., the singular values of column blocks. By independently rescaling these blocks, their relative contribution to each weight matrix can be changed, thereby rotating the singular vectors. Theoretical analysis and experiments both show that this update can be rewritten as a transformation of the layer input, so FuRA-S can also adapt to INT4 quantized models without modifying their weights.
Specifically, the expert models trained by FuRA-S are only 1-5 MB in size, with the proportion of updated parameters at just 0.015-0.05%. On mobile agent tasks, MobileSpec improves the success rate from 6-23% to 89-100%. Additionally, it can train a personalized 8B model overnight on an iPhone 16 Pro. In RLVR benchmarks, FuRA-S is competitive with full-parameter evolution strategies (ES) and outperforms ES on LoRA adapters as well as quantized ES on INT4 models.
Singular Value Decomposition (SVD) and Spectral Coordinates: Any matrix $W$ can be decomposed as $W = U\Sigma V^\top$, where the diagonal elements of $\Sigma$ are the singular values, and the column vectors of $U$ and $V$ are the singular vectors. Singular values characterize the "energy" or importance of the matrix in different directions. The so-called "block spectral coordinates" refer to dividing the weight matrix into several column blocks, computing singular values for each column block separately, and using only these singular values as trainable variables. Independently scaling the singular values of a column block is equivalent to adjusting the block's relative contribution to the overall output; mathematically, this scaling also changes the direction of the singular vectors, i.e., produces the effect of "rotating singular vectors." The weight changes induced by RLVR during training often preserve the general structure of the original singular value spectrum and only slightly rotate the singular vectors. Therefore, directly training block singular values precisely simulates this change while compressing the trainable parameters to a very small proportion.
Evolution Strategies (ES) are a class of gradient-free black-box optimization methods: they sample random perturbations in parameter space, evaluate the reward brought by each perturbation, and then update parameters by reward-weighted averaging. ES requires only forward evaluation, making it naturally suited to memory-constrained scenarios, but full-parameter ES has extremely high search dimensionality and high variance. FuRA-S can be viewed as performing ES on low-dimensional spectral coordinates, making it more efficient under the same forward budget. This is also why it performs comparably to full-parameter ES while significantly outperforming ES on LoRA.
Experimental Results: From Mobile Agents to Personalized 8B Models
MobileSpec demonstrates its effectiveness across multiple scenarios. In mobile agent tasks, the substantial improvement in success rate shows that forward-only optimization can effectively adapt the model. In training a personalized 8B model, an iPhone 16 Pro completed training overnight, proving the method's feasibility on real devices. Comparisons with baseline methods further validate FuRA-S's advantages in memory efficiency and performance.
Significance and Outlook
MobileSpec provides a practical path for on-device LLM personalization. By avoiding backpropagation and labeled data, it lowers the training barrier, enabling ordinary users to obtain expert models tailored to their own tasks on their phones. In the future, this approach is expected to extend to more task types and devices, further advancing privacy-preserving personalized AI.
One-line Summary
MobileSpec trains 1-5MB expert models on phones using only forward propagation, raising agent task success rates to 89-100%.
Paragraph Summary
MobileSpec is a training system for on-device LLM personalization that aims to circumvent the memory overhead of backpropagation and the scarcity of user-labeled data. It updates models only through forward propagation and leverages reinforcement learning with verifiable rewards (RLVR) to obtain reward signals from automatically checked task outcomes. Its core algorithm, FuRA-S, trains only the block spectral coordinates (column-block singular values) of pre-trained weights. The resulting perturbations are structurally close to RLVR's weight changes, can alter singular vector directions, and can adapt to INT4 quantized models without modifying weights. Experiments show that expert models are only 1-5MB with an updated parameter proportion of 0.015-0.05%, mobile agent task success rates improve from 6-23% to 89-100%, and a personalized 8B model can be trained overnight on an iPhone 16 Pro, with performance comparable to full-parameter ES and better than ES on LoRA.
Related articles

RePlan-VLA: Progress-Aware Replanning Breakthrough for Long-Horizon Robotic Manipulation
The RePlan-VLA framework converts VLA models' short-horizon capabilities into reliable long-horizon execution through progress-aware adaptive replanning, improving success rates by 21.9% and 10.4% on RoboCasa365 and CALVIN, and validating real-world deployment on a Franka Research 3 robot.

Geometric Bridge Flow Matching: A New Breakthrough in Blind Rigid Motion Correction
Geometric bridge flow matching enables blind rigid motion correction in head CBCT, achieving 37.1 dB PSNR and 0.27 mm RPE on simulated data in ~9.1 minutes, outperforming existing methods.

Google Releases Gemini 4 Hydrogen: 2.5x Faster Without Sacrificing Intelligence
Google releases the lightweight Gemini 4 Hydrogen: 2.5x faster task completion, AA Intelligence Index of 54 beating its flagship sibling, 7.3s single-step on Agent Arena faster than Claude and GPT rivals, four thinking levels with pay-as-you-need pricing, now available across AI Studio and more.