Feb 2026· arXiv.org· Vol abs/2602.01734· 1 citation· 29 references
Computer Science
TL;DR
This work proposes MSign, a new optimizer that periodically applies matrix sign operations to restore stable rank in a 5M-parameter NanoGPT model, and proves theoretically that these two conditions jointly cause exponential gradient norm growth with network depth.
Abstract
Training instability remains a critical challenge in large language model (LLM) pretraining, often manifesting as sudden gradient explosions that waste significant computational resources. We study training failures in a 5M-parameter NanoGPT model scaled via $\mu$P, identifying two key phenomena preceding collapse: (1) rapid decline in weight matrix stable rank (ratio of squared Frobenius norm to squared spectral norm), and (2) increasing alignment between adjacent layer Jacobians. We prove theoretically that these two conditions jointly cause exponential gradient norm growth with network depth. To break this instability mechanism, we propose MSign, a new optimizer that periodically applies matrix sign operations to restore stable rank. Experiments on models from 5M to 3B parameters demonstrate that MSign effectively prevents training failures with a computational overhead of less than 7.0%.
The results suggest that wide TopK SAEs should be evaluated not only by reconstruction, sparsity, and feature count, but also by feature reliability under semantic variation and boundary geometry for stable interpretability.
Zhen-Ting Huang, Bo-Han Jiang, J. Liu et al.· 0 citations
The gradient flow of the best rank-$r$ approximation is derived and it is pointed out that the offset of DLRT comes from a curvature-coupling term which is large and thus non-negligible under aggressive compression, and a stable dynamic low-rank training method is proposed, named SDLRT.
Existing LLM unlearning methods struggle with robustness: unbounded forget losses degrade model coherence, fixed-weight balancing cannot adapt as retain difficulty shifts mid-training, and methods that work on one benchmark falter under scaling or repeated application. We propose BLADE, a constrained bilevel framework...
Md Toufikuzzaman, Ahmad Mousavi, Dongwon Lee· 0 citations
This paper investigates SoRA algorithm's performance in large-scale models, and derives an improved learning rate configuration and proposes SoRA+, an enhanced variant that introduces a hyperparameter to adjust learning rates proportionally and mitigates the suboptimality of the original SoRA method.
Results establish ELR as a common coordinate linking LR scheduling, norm control, and loss dynamics, and Controlled interventions further show that weight decay and Hyperball shape loss dynamics primarily through the ELR schedules they induce.
Zi-Han Liu, Rui-Heng Zheng, Shaobo Zhang et al.· 4 citations
Low-Rank Adaptive Residual Connections (LARC) give a frozen model a compact numerical state that can learn from feedback. The map $h+BAh$ adds a low-rank correction to a hidden representation. A slow state $\rho$ learns starting factors across tasks; a private fast state $\Phi$ copies them, changes with feedback, and r...
Jun-Yi Zou, Avrova Donz· 1 citation
Related blog posts
MIT News · Artificial Intelligence· news.mit.eduNov 17, 2026
By studying the computational foundations of language learning, Associate Professor Jacob Andreas builds intelligent systems that can communicate effectively with humans.
Microsoft Research Blog· microsoft.comNov 16, 2026
AI coding agents are increasingly help developers fix bugs and build features largely on their own, but every interaction with their models consumes tokens. In each session, an agent sends instructions, tool descriptions, and session history to its model. Most efforts to reduce this cost focus on individual tasks. Yet developers return to the same […] The post TokenEconomizer: Learning from Experience to Make AI Coding Agents More Efficient appeared first on Microsoft Research.
MIT News · Artificial Intelligence· news.mit.eduNov 13, 2026