Researchers are increasingly focused on understanding why deep learning training is so fragile despite impressive results. Zhipeng Zhang (China Mobile Research Institute & China Mobile GBA Innovation Institute), Zhenjie Yao (Institute of Microelectronics, Chinese Academy of Sciences), and Kai Li, together with Lei Yang, demonstrated that training instability is not random but follows predictable low-dimensional mechanical principles. Their work revealed that factors such as optimization, data, parameters, and signal stability interact during training, and performance is not always a reliable indicator of a stable process. By carefully auditing training trajectories, the team identified consistent patterns such as the protective effects of controlled stochasticity and early warning signs of latent state collapse, providing new and measurable ways to assess and improve the robustness of deep learning systems.
A training policy that adjusts the proximity to instability (entropy or gradient coherence). This study develops a scientific framework to study training stability as a property. In the era of foundational models, this reconfiguration has implications for responsible scaling and safety. Late-stage failures are not just engineering mishaps, but constraints on scientifically investigating, replicating, and managing scaling regimes. Existing scaling laws describe how a model’s functionality increases with size, data, and compute. However, our findings highlight important blind spots. Scaling capabilities does not imply dynamic reliability, that is, whether those capabilities can be stably achieved under unavoidable perturbations.
Researchers have introduced perturbation-based auditing as a methodological approach to study training stability. Rather than relying on anecdotal failure analysis, perturbation auditing systematically investigates the dynamic response of a learning system and provides principled and reproducible insights into the formation of instabilities. They propose StabilityBench not as a benchmark, but as a scientific tool that enables controlled perturbation auditing across learning paradigms such as reinforcement learning and large-scale language models. Through such audits, we uncover cross-domain regularities in how instability progresses, often before performance degradation becomes visible.
They further propose that the meta-state representation serves as a low-dimensional structural summary of training dynamics. Meta-states integrate multiple telemetry channels, such as performance metrics, gradient statistics, and optimizer states, into a joint representation that captures how learning dynamics evolve as a form of instability. Importantly, the metastate is not an average, but a representation of how multiple channels covary as training approaches a structural transition. This aggregation enables conditional closed-loop interactions as a monitoring prototype. Meta-states can support selective, non-intrusive interactions with training dynamics in unstable states while remaining stationary in stable states.
This interaction is used to investigate the responsiveness of learning dynamics, rather than to claim predictive power or deploy control strategies. By providing structured, low-dimensional observability, this research lays the foundation for learning systems that are not only competent but also scientifically interpretable, diagnosable, and auditable. Why use a joint meta-state rather than a single metric?Individual indicators (e.g. performance trends, slope statistics, or short-term instability indices) may provide only partial predictions of the underlying dynamics and may not show consistent anomalies, even in the form of instability.
Across audits, collapsible runs are characterized by multichannel drift that is coordinated over time, motivating a joint latent representation rather than thresholding a single metric alone. This limitation is demonstrated in Section 2.3. Experiments revealed that high final performance is often decoupled from training stability, and this finding was demonstrated both in reinforcement learning and in training large-scale language models. The researchers measured a systematic dissociation between these two factors and demonstrated that models that achieve state-of-the-art results can be highly vulnerable to small disturbances during training.
The results show that reinforcement learning algorithms exhibit significant differences in training stability under optimization perturbations. In HalfCheetah-v3, a single learning rate spike at step 2000 caused an irreversible training collapse in PPO, whereas SAC and TD3 maintained stable learning trajectories despite comparable returns before the perturbation. Our data show that this instability manifests itself as an algorithm-dependent failure mode, rather than gradual performance degradation or noise accumulation, with a consistent pattern observed under perturbations of action noise and reward scale. This breakthrough provides a way to characterize stability as a dynamic property, beyond its dependence on final performance outcomes. Further analysis of large-scale language models confirms similar stability, performance divergences, documenting loss spikes and interruptions during GPT-3 training due to learning rate scheduling issues, irrecoverable divergences in PaLM training from gradient numerical anomalies, and sharp loss surges in LLaMA training.
Measurements confirm that these events highlight the vulnerability of the model to reach peak performance during training, and that this phenomenon maps systematically to specific dimensions of the instability classification. Tests demonstrate that controlled dimension-specific perturbations, rather than a posteriori failure analysis, establish stability as a property independent of final performance. Scientists have documented that training instabilities consistently appear as sudden, non-smooth events and are often caused by single local perturbations, indicating that this is caused by structural transitions in training dynamics rather than by long-term stochastic accumulation. The team focused on gradient-directed coherence, denoted as xgrad, to analyze training instability from a dynamical systems perspective, identify instability manifolds shared across learning paradigms, and characterize the geometry of gradient updates.
