stability analysis for continual training

This work addresses the vulnerability of neural network training to rare yet severe unstable updates, which can cause irreversible divergence or subtle performance degradation—issues that existing optimizers fail to detect or mitigate at runtime. The authors model the optimization process as a controlled stochastic process and introduce the first runtime stability framework that operates without modifying the underlying optimizer. By leveraging secondary signals such as validation probes, the framework automatically detects instability and triggers lightweight interventions grounded in control theory. Designed for memory-constrained settings, the approach offers low computational overhead, broad compatibility with standard optimizers, and theoretical guarantees of bounded degradation and recovery, effectively preventing training collapse and performance deterioration.

This work addresses the challenge of training failures in large language models, which often persist for thousands of optimization steps before manifesting as obvious loss divergence, leading to substantial computational waste. The authors propose a mechanism-aware, proactive monitoring approach that deploys internal detectors at the earliest points where failure signatures become measurable. Specifically, they introduce diagnostic signals grounded in the functional principles of critical modules—such as spectral entropy derived from the bilinear decomposition of QK matrices under low-precision Flash Attention and behavioral metrics of MoE router expert selection. By leveraging these module-specific indicators, the method enables early and accurate identification of diverse failure modes thousands of steps before loss divergence occurs, significantly outperforming conventional detection strategies based solely on loss values or gradient norms, particularly in scenarios involving low-precision attention, excessively high learning rates, or compound faults.

This study addresses the critical issue of model instability in software engineering optimization, which leads to substantial variability across repeated experiments and undermines both credibility and practical utility. Rather than treating instability as mere random noise, this work conceptualizes it as a quantifiable and manageable property that should be integrated into standard evaluation frameworks. By systematically modulating label usage, model complexity, and partition scoring strategies—combined with multi-objective optimization, causal intervention, data locality analysis, and model calibration—the proposed approach significantly enhances result consistency. Empirical evaluation demonstrates that the optimized configuration reduces the standard deviation of error by 22% on average and outperforms default settings in 119 out of 127 datasets, achieving a 4.8-fold improvement in result consistency.

This work addresses the challenge that existing metrics struggle to capture behavioral drift in large language model (LLM) endpoints caused by changes in weights, tokenizers, quantization, or inference stacks—threatening the consistency of AI applications. To this end, we propose Stability Monitor, the first system that models LLM endpoint behavior through black-box behavioral fingerprinting. By periodically sampling outputs from a fixed prompt set, it constructs output distribution fingerprints and leverages energy distance combined with permutation testing to enable internal-agnostic stability monitoring and change detection. Experimental evaluation demonstrates that our approach effectively identifies differences across model families, versions, quantization schemes, and inference stacks, revealing significant stability disparities in real-world deployments across multiple providers.

In industrial quality inspection, anomaly detection suffers from poor robustness due to high noise levels and sparse defective samples. To address this, we propose Iterative Refinement of Pseudo-labels (IRP), a self-supervised method that alternately evaluates sample credibility and removes misleading instances under feature-space consistency constraints—effectively purifying the training set dynamically without human annotations and generating high-fidelity self-supervised signals. IRP introduces the novel paradigm of “iterative data refinement,” significantly enhancing model robustness against label noise and cross-domain generalization capability. Evaluated on KSDD2 and MVTec AD benchmarks, IRP consistently outperforms existing unsupervised and self-supervised methods. Notably, under high-noise conditions, it achieves substantial improvements in detection accuracy and reduces false positive rates by over 25%.

This work addresses the feedback loops that arise after model deployment due to performativity—wherein the model’s predictions influence the data distribution—particularly under strong interventions where the convergence behavior of retraining remains poorly understood. The paper introduces the “stable signal principle,” positing that the prediction target contains an intrinsic component independent of the model (e.g., inherent item quality), and leverages this insight to analyze the dynamics of regularized repeated risk minimization. Theoretically, it establishes that as long as a non-zero stable signal exists, retraining converges geometrically to its direction, even when model-induced effects dominate. This reveals a novel role for regularization in mitigating performative feedback and extends the framework to nonlinear, heterogeneous, and time-varying settings—including language models—thereby explaining the observed stability of training on generated data.

This work addresses the performance degradation of autoregressive language models during long-sequence generation—manifested as repetition, diminished instruction-following ability, and unstable entropy—for which real-time diagnostic tools are lacking. The authors formalize this phenomenon as “cognitive fatigue” and model it through three dimensions: attention decay, representational drift, and entropy calibration bias. They propose a lightweight, model-agnostic Fatigue Index (FI) that satisfies axiomatic properties of monotonicity, boundedness, and interpretability, enabling online monitoring. Evaluated across nine models ranging from 1B to 13B parameters, FI effectively predicts task performance degradation (AUROC = 0.95) and textual repetition (Spearman ρ = 0.94). The analysis further reveals a non-monotonic relationship between model scale and fatigue dynamics: instruction-tuned models below 3B parameters exhibit accelerated fatigue, whereas those above 7B show significant mitigation.

This work addresses the limitations of conventional continual learning evaluation, which relies on zero-shot forgetting metrics and fails to comprehensively capture a model’s ability to balance retention of prior knowledge with adaptation to new tasks. To overcome this, the authors propose a few-shot continual learning evaluation framework that introduces a novel metric—“per-sample plasticity”—and integrates meta-learning to prospectively model future tasks, thereby enhancing the model’s ability to learn how to learn. Through fine-grained analysis of task sequences in continual image classification, the study systematically reveals the behavioral characteristics of mainstream continual learning algorithms under few-shot settings and demonstrates that the proposed prospective mechanism significantly improves cross-task adaptability.

This work addresses the challenge of efficiently assessing stability in differential-algebraic equation (DAE) systems subject to stochastic dynamic environments, particularly in high-dimensional or real-time settings where repeated simulation is computationally prohibitive. The authors propose a learning framework for testing that uniquely integrates physical constraints with distributional hypothesis testing. By leveraging a neural dynamical surrogate model and uncertainty-aware calibration, the method constructs a physics-informed, regularized latent representation. This enables stability monitoring at deployment to be cast as a distributional hypothesis test in latent space—eliminating the need for repeated DAE solves. The approach provides an efficient, scalable, and statistically reliable means of detecting instability risks under distributional shifts while maintaining strict control over Type I error rates.

本文通过构建CodeInsight数据集，采用多种模型包括RSSM和基于LLM的预测器，研究编程学习中迭代解决问题的过程。