Unveiling the Basin-Like Loss Landscape in Large Language Models

Huanran Chen, Yinpeng Dong, Zhi Qiang Wei, Yao Huang, Yichi Zhang, Hang Su, Jun‐Jie Zhu

Abstract

We discover the emergence of basins in the loss landscape of large language models (LLMs). As model scale increases, LLMs become progressively more resilient to random perturbations in the parameter space, giving rise to expansive stability regions where models exhibit nearly identical performance, but outside of which their capabilities collapse. We observe that pre-training creates a basic capability basin, and subsequent alignment fine-tuning forms specific capability basins (e.g., safety, math, coding). Thus, we argue that benign fine-tuning confined to the basin should preserve prior capabilities. Besides, we also analyze the loss landscape for worst-case directions, which is consistently sharp and detrimental. We find that adversarial fine-tuning moves along the nearly worst-case directions, thus rapidly degrading model capabilities. Finally, we provide a theoretical analysis demonstrating that the basin size bounds the performance degradation of any fine-tuning, including the adversarial ones, while also guaranteeing the model robustness w.r.t. input perturbations, suggesting the benefit of enlarging basins.

Introduction

Despite the remarkable capabilities of Large Language Models (LLMs), they exhibit a persistent "alignment brittleness." This fragility manifests as catastrophic forgetting—where fine-tuning on a new task causes the loss of previously learned skills—and extreme vulnerability to adversarial fine-tuning that can bypass safety guardrails in just a few steps. This research investigates these phenomena through the lens of the loss landscape, mapping how model performance (measured by generative benchmark scores) fluctuates as parameters are perturbed.

The Geometry of LLM Capabilities: Most-Case Basins

By analyzing perturbations in random directions (the "most-case" landscape), researchers discovered that LLMs do not exist as isolated points of high performance. Instead, they reside within expansive basins. Within these regions, parameters can be shifted significantly without a meaningful drop in benchmark scores.

Figure 1: Loss landscapes for Llama, Qwen, and Mistral across different tasks (Safety, Math, Basic, Coding) show broad, stable basins in random directions.

As shown in Figure 1, these basins are consistent across various model architectures and tasks. Interestingly, the study found that larger models possess significantly wider basins, explaining their inherent resilience to random noise compared to smaller counterparts. Pre-training appears to establish a broad "basic capability" basin, while alignment processes like safety training carve out narrower, specialized basins within or adjacent to this primary region.

The Fragility of Alignment: Worst-Case Cliffs

While the landscape is generally stable in random directions, it is not uniformly robust. By using optimization techniques to find the steepest directions of performance decay, the researchers identified "worst-case" directions. In these specific directions, the loss landscape is incredibly sharp, resembling a steep cliff rather than a broad basin.

Figure 2: The worst-case loss landscape for Llama-3.1-8B is extremely sharp, showing how specific parameter shifts can lead to immediate performance collapse.

Adversarial fine-tuning specifically exploits these sharp directions. Even a tiny movement along these "cliffs" can lead to a total collapse of model alignment or safety guardrails, explaining why models that seem robust to general use can be so easily compromised by targeted attacks.

Enhancing Robustness via Basin Expansion

To address this brittleness, the authors introduced the Gaussian-augmented Optimizer (GO). Unlike standard optimizers that minimize loss at a single point, GO optimizes the expected loss over a neighborhood of parameters. This approach explicitly encourages the formation of wider stability basins during training.

Theoretical analysis using randomized smoothing confirms that the size of the most-case basin ($\sigma$) provides a formal lower bound on performance degradation for any subsequent fine-tuning. Experiments at the 127M parameter scale demonstrated that models trained with GO exhibited significantly less catastrophic forgetting when later fine-tuned on new tasks, suggesting that "widening the basin" is a viable path toward more robust and permanent alignment.

Conclusion

The discovery of basin-like structures in LLM loss landscapes provides a geometric explanation for both the resilience and the fragility of modern AI. While scaling naturally expands these basins, targeted optimization like the GO optimizer can further enhance model stability. Future work may focus on developing scaling laws for these basins and exploring how "over-training" might serve to solidify these regions of capability stability.

Summary updated

Abstract

We discover the emergence of \textit{basins} in the loss landscape of large language models. As model scale increases, LLMs become progressively more resilient to random perturbations in the parameter space, giving rise to expansive stability regions where models exhibit nearly identical performance, but outside of which their capabilities collapse. We observe that pre-training creates a \textit{basic capability} basin, and subsequent alignment fine-tuning forms \textit{specific capability} basins (e.g., safety, math, coding). Thus, we argue that benign fine-tuning confined to the basin should preserve prior capabilities. Besides, we also analyze the loss landscape for worst-case directions, which is consistently sharp and detrimental. We find that adversarial fine-tuning moves along the nearly worst-case directions, thus rapidly degrading model capabilities. Finally, we provide a theoretical analysis demonstrating that the basin size bounds the performance degradation of any fine-tuning, including the adversarial ones, while also guaranteeing the model robustness w.r.t. input perturbations, suggesting the benefit of enlarging basins.
BibTeX
@inproceedings{chen2026unveiling,
  author = {Huanran Chen and Yinpeng Dong and Zhi Qiang Wei and Yao Huang and Yichi Zhang and Hang Su and Jun‐Jie Zhu},
  title = {{Unveiling the Basin-Like Loss Landscape in Large Language Models}},
  booktitle = {International Conference on Learning Representations},
  year = {2026},
  url = {https://openreview.net/forum?id=l4q2Zk2yfk},
}
Figure