mlsanityData to 7 Oct 2026

Stability as a Liability:Systematic Breakdown of Linguistic Structure in LLMs

arXiv 2601.18588 ↗First posted 26 Jan 2026cs.AI cs.CL cs.LGcited by 0

Not on the map. Why?

Abstract (arXiv)

Training stability is typically regarded as a prerequisite for reliable optimization in large language models. In this work, we analyze how stabilizing training dynamics affects the induced generation distribution. We show that under standard maximum likelihood training, stable parameter trajectories lead stationary solutions to approximately minimize the forward KL divergence to the empirical distribution, while implicitly reducing generative entropy. As a consequence, the learned model can concentrate probability mass on a limited subset of empirical modes, exhibiting systematic degeneration despite smooth loss convergence. We empirically validate this effect using a controlled feedback-based training framework that stabilizes internal generation statistics, observing consistent low-entropy outputs and repetitive behavior across architectures and random seeds. It indicates that optimization stability and generative expressivity are not inherently aligned, and that stability alone is an insufficient indicator of generative quality.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Builds on 21 of its 39 references that are in the map. First posted on arXiv 26 Jan 2026.

Builds on

21 of 39 references.

References · 21

  1. 2019Normalizing Flows for Probabilistic Modeling and Inferencecited by 1,315
  2. 2025The Entropy Mechanism of Reinforcement Learning for Reasoning Language Modelscited by 463
  3. 2020Understanding the Difficulty of Training Transformerscited by 210
  4. 2025The Unreasonable Effectiveness of Entropy Minimization in LLM Reasoningcited by 160
  5. 2016Improving the Robustness of Deep Neural Networks via Stability Trainingcited by 156
  6. 2020Understanding Gradient Clipping in Private SGD: A Geometric Perspectivecited by 140
  7. 2025Where LLM Agents Fail and How They can Learn From Failurescited by 118
  8. 2023Revisiting Gradient Clipping: Stochastic bias and tight convergence guaranteescited by 83
  9. 2025R1-Reward: Training Multimodal Reward Model Through Stable Reinforcement Learningcited by 65
  10. 2024Entropy Law: The Story Behind Data Compression and LLM Performancecited by 42
  11. 2015Understanding Adversarial Training: Increasing Local Stability of Neural Nets through Robust Optimizationcited by 32
  12. 2024In-the-loop Hyper-Parameter Optimization for LLM-Based Automated Design of Heuristicscited by 28
9 more
  1. 2024Edge of Stochastic Stability: Revisiting the Edge of Stability for SGDcited by 24
  2. 2024Under the Surface: Tracking the Artifactuality of LLM-Generated Datacited by 23
  3. 2025A Survey on LLM Mid-Trainingcited by 18
  4. 2023RoBoSS: A Robust, Bounded, Sparse, and Smooth Loss Function for Supervised Learningcited by 15
  5. 2024Evaluating Uncertainty-based Failure Detection for Closed-Loop LLM Plannerscited by 15
  6. 2024Methods of improving LLM training stabilitycited by 13
  7. 2021Characterizing and addressing the issue of oversmoothing in neural autoregressive sequence modelingcited by 10
  8. 2025Can LLMs Leverage Observational Data? Towards Data-Driven Causal Discovery with LLMscited by 4
  9. 2025Control LLM: Controlled Evolution for Intelligence Retention in LLMcited by 3

Built on by

Not cited yet.