Li Yang

Lei Yu, Peng Wang, Jia Xu et al.

Bash scripts are the cornerstone of system administration and DevOps automation, where code quality directly impacts system stability and security. In automated Bash script generation using Large Language Models (LLMs), two interconnected failures emerge: unauditable "black box" reasoning and critical robustness vulnerabilities in generated code. To address both, we propose BashCoder-R1, a novel framework for robust and explainable Bash script generation. Our pipeline combines: (1) Continual Pre-training (CPT) to specialize the model on Bash paradigms; (2) Long Chain-of-Thought Supervised Fine-Tuning (L-CoT SFT) on expert-validated reasoning-and-code samples to emulate proactive risk-aware thinking; and (3) Robustness-Aware Group Relative Policy Optimization (R-GRPO), a reinforcement learning phase optimizing a weighted reward for syntax correctness, robustness (via shellcheck), and format correctness. We evaluate on BashBench, a new benchmark of 952 real-world tasks (773 single-line, 179 multi-line). BashCoder-R1 achieves SyntaxPass (100.00%/94.97%), RobustWarnRate (4.01%/16.47%), RobustPass (95.99%/79.33%), FuncRate (93.01%/93.85%), and FullRate (90.04%/73.18%) for single-line/multi-line tasks, outperforming the strongest baseline DeepSeek-V3.2 (Reasoning) by 37.82% and 20.18% in FullRate. Human evaluation on Functionality, Robustness, and Clarity further confirms BashCoder-R1 achieves the highest quality ratings.

Alec Dewulf, Dhruv Pai, Li Yang et al.

We show that for tall matrix parameters, like projection matrices in the MLP layers, the Muon update can have row norms that are arbitrarily non-uniform. This can lead to a self-reinforcing feedback loop whereby neurons receive persistently small updates and eventually do not contribute meaningfully to network outputs. This problem is effectively mitigated by an additional row normalization step, but current methods do this in a way that moves the Muon update geometry away from the polar factor of the momentum matrix, which we find is undesirable. We propose Aurora, an optimizer that enforces row-uniformity of matrix parameter updates while respecting Muon's polar factor geometry. Aurora outperforms Muon in our pre-training experiments and, when combined with existing methods, achieves state-of-the-art performance among spectral optimizers on the optimizer track of the modded-nanoGPT speedrun. Additionally, we find that Aurora's empirical gains over Muon scale with the MLP expansion factor, suggesting that Aurora may allow for effective training of very wide MLP layers.