Supervised Fine-Tuning Needs to Unlock the Potential of Token Priority
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2026
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866908822200647680 |
|---|---|
| author | Shen, Zhanming Qin, Zeyu Hu, Jiaqi Ye, Wentao Chen, Hao Hu, Xiaomeng Xu, Haokai Chen, Gang Fung, Yi R. Wang, Haobo |
| author_facet | Shen, Zhanming Qin, Zeyu Hu, Jiaqi Ye, Wentao Chen, Hao Hu, Xiaomeng Xu, Haokai Chen, Gang Fung, Yi R. Wang, Haobo |
| contents | The transition from fitting empirical data to achieving true human utility is fundamentally constrained by a granularity mismatch, where fine-grained autoregressive generation is often supervised by coarse or uniform signals. This position paper advocates Token Priority as the essential bridge, formalizing Supervised Fine-Tuning (SFT) not as simple optimization but as a precise distribution reshaping process that aligns raw data with the ideal alignment manifold. We analyze recent breakthroughs through this unified lens, categorizing them into two distinct regimes: Positive Priority for noise filtration and Signed Priority for toxic modes unlearning. We revisit existing progress and limitations, identify key challenges, and suggest directions for future research. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2602_01227 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Supervised Fine-Tuning Needs to Unlock the Potential of Token Priority Shen, Zhanming Qin, Zeyu Hu, Jiaqi Ye, Wentao Chen, Hao Hu, Xiaomeng Xu, Haokai Chen, Gang Fung, Yi R. Wang, Haobo Computation and Language Artificial Intelligence Machine Learning The transition from fitting empirical data to achieving true human utility is fundamentally constrained by a granularity mismatch, where fine-grained autoregressive generation is often supervised by coarse or uniform signals. This position paper advocates Token Priority as the essential bridge, formalizing Supervised Fine-Tuning (SFT) not as simple optimization but as a precise distribution reshaping process that aligns raw data with the ideal alignment manifold. We analyze recent breakthroughs through this unified lens, categorizing them into two distinct regimes: Positive Priority for noise filtration and Signed Priority for toxic modes unlearning. We revisit existing progress and limitations, identify key challenges, and suggest directions for future research. |
| title | Supervised Fine-Tuning Needs to Unlock the Potential of Token Priority |
| topic | Computation and Language Artificial Intelligence Machine Learning |
| url | https://arxiv.org/abs/2602.01227 |