Unveiling the Basin-Like Loss Landscape in Large Language Models
Fuente:
arXiv
Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866911516592177152 |
|---|---|
| author | Chen, Huanran Dong, Yinpeng Wei, Zeming Huang, Yao Zhang, Yichi Su, Hang Zhu, Jun |
| author_facet | Chen, Huanran Dong, Yinpeng Wei, Zeming Huang, Yao Zhang, Yichi Su, Hang Zhu, Jun |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2505_17646 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Unveiling the Basin-Like Loss Landscape in Large Language Models Chen, Huanran Dong, Yinpeng Wei, Zeming Huang, Yao Zhang, Yichi Su, Hang Zhu, Jun Machine Learning 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. |
| title | Unveiling the Basin-Like Loss Landscape in Large Language Models |
| topic | Machine Learning |
| url | https://arxiv.org/abs/2505.17646 |